The mask is one of literature’s oldest technologies. Throughout literary history, it has served as a tool [technē] for humans to solve problems and get things done. Lazarillo de Tormes (1554), Shakespeare’s Twelfth Night (1601), Anne Mauduit de Fatouville’s Arlequin empereur dans la lune (1684): early modern comedy holds a long catalogue of tricksters who get by, pretending to be someone else.

Each of these works evinces the same idea. Worldviews are narrow, and we don masks to increase our credibility, hoping to be seen the way we need to.

Poets carried this idea into the 20th century, creating a theory about their own craft. Fernando Pessoa, writing about his own kind, stated: “O poeta é um fingidor / Finge tão completamente / Que chega a fingir que é dor / A dor que deveras sente” [The poet is a feigner / He feigns so completely / That he comes to feign as pain / The pain he really feels]. Pessoa’s mask has someone behind it, and the feigning is a technique for getting something true onto the page.

Today we look constantly for new lenses through which to understand technological developments — Responsible AI, model cards, red-teaming, AI safety frameworks.

In our panel proposal for SXSW 2027, Lindsey DeWitt Prat and I frame the stochastic, probabilistic nature of AI models through the trickster of classical comedies. We find that a machine learning model is just like the fraud in a mask, only able to generate the world it already knows. Consider how the 1684 play of Arlequin as ambassador from the moon plays out. Arlequin never says the refrain. Nine times in a single scene his listeners greet his dispatches from the moon with the same delighted line, c’est tout comme ici—it’s just like here—and the ninth time the stage direction hands it to the whole household at once: TOUS ENSEMBLE. He does not have to be convincing. The audience does the describing for him, out loud, in unison, on cue.

Urs App, the historian of philosophy who traced how Europeans assembled something they called Oriental philosophy out of their own categories, named this pattern the Arlecchino mechanism. What makes it a mechanism rather than an error is that nobody involved notices.

Large language models (LLMs) are cut from very narrow cloth, a subsection of the world’s training data that is limited and riddled with bias, and they are now expected to output what we are calling Artificial General Intelligence, speaking for a monolithic society.

The mask is a stand-in for the most common viewpoints throughout mainstream media and tech development globally, those of populations that are Western, Educated, Industrialized, Rich, and Democratic (W.E.I.R.D.). Joseph Henrich, Steven Heine, and Ara Norenzayan coined this acronym in 2010 to describe who certain fields had been strictly sampling while deriving knowledge about society as a whole.

More recently, Yan Tao, Olga Viberg, Ryan Baker, and René Kizilcec tested five OpenAI models against the Integrated Values Survey, nationally representative data from 107 countries and territories, and every model expressed cultural values resembling English-speaking and Protestant European countries. Here, the metaphorical mask not only conceals but selects, and speaks fluently for the whole world, from just a very narrow point of view.

Or, consider another theory of the mask. In Japanese noh, the object is not called a mask. It is 面, omote, which means face, or surface, and nobody hides behind it. The expression is not in the object either: a noh face is carved close to neutral, and what an audience reads as joy or grief comes from the tilt of the actor’s head against the light—照らす terasu, tilted up toward brightness, 曇らす kumorasu, tilted down toward cloud. The wood does not change. The angle changes, and the viewer supplies the rest. Prompt a model in English about a place the English-language internet covers thickly, and the face brightens. Query about somewhere the training data barely reaches, and the face brightens anyway.

Machine learning utilizes these same samples, and now speaks with decisiveness about the human condition, without the data to qualify its claims. In Austin, Lindsey and I will take these masks apart: where they get made, who paints the expression on, and what changes when the people being described decide what goes in.

In honor of HTML day, which was this past Saturday, August 8th, I decided to take a walk down the memory lane of the nascent internet I remember from my childhood. As I scoured the archives of my brain, I was hit with a memory — the notorious chain email.

For those of you who are fortunate enough to remember, chain emails were messages that attempted to convince the recipient to make a number of copies to a certain number of recipients, growing an exponential pyramid chain of people that could never be sustained.

What did these look like you ask?

Many of them started with the most complex of hooks like, STOP!!!!, or, KEEP READING. DO NOT IGNORE.

Behold the chain email of our childhoods:

These chain letters sometimes kept me up at night, and often contained enough urgency and social weight that I would forward them to my friends. And a few years later, after the chain email fad drifted away, came the text chains.

A brief history of the “Chain Letter”

From my research on this topic, I’ve found there are two types of chain letters:

1. Hoax chain letters which attempt to trick or defraud users.

2. Luck- or superstition-based letters which promise good luck for forwarding or bad luck (even as far as death) if the chain is broken or the letter is not forwarded.

Scholars note that the popularity of chain mail typically flares up around times of intense social anxiety or worry about the future, which checks out for the Y2K drama of the turn of the century, as well as the introduction of Web 2.0, or peer-to-peer communication, which provided us mass access to people we might not otherwise have been able to communicate with but through a pen pal relationship.

But, chain letters actually started well before the email existed. In fact, they date back to the 6th century, when they were called “Himmelsbrief” or Heaven Letters, letters that people purported fell from heaven, delivered by agents of God, that would urge the adherence to Christian teachings.

How did this work without email?

Well, a scribe would copy the letter and send it to a number of people, which would then bring good fortune. Failing to do so would bring misfortune, and a potential hell-sentence.

These types of letters were particularly popular during the Reformation, a time in which many sold indulgences, as “ways to reduce the amount of punishment one has to undergo for (forgiven sins)” which consisted of the purchase of salvation as cash-for-heaven tickets.

Well-past Martin Luther’s demonishment of indulgences, around the 1800s, these types of letters transformed into what is referred to as “Prosperity Clubs” or “Send-a-Dime” letter, first started in 1935, during the Great Depression. One particular scheme, which instructed recipients to send a dime to the person at the top of a list of names, remove that name, add their own to the bottom, and send that letter to five others, swamped the Denver post office with nearly 100,000 letters a day before it spilled to other cities.

Outside of the US, chain letters gained particular popularity in Japan, with the Fukou no Tegami (the Unlucky Letter or Letter of Misfortune) which preyed on the superstitions of Japanese culture at the time:

This is an Unlucky Letter, a death god that came to me sequentially from Okinawa. A Canadian supposedly thought of it. If you stop it with you, misfortune will certainly visit. A person in Texas stopped it and died five years later. You too, please send this letter to twenty-nine people within thirty hours without changing the text. I am number XXX.

— Quoted in 東 1996

In the late 1990s and early 2000s, right before the popularization of chain e-mails, chain letters often targeted specific demographics, relying on fear-based tactics to push recipients to act. These letters are often referenced as a method to either calm societal anxiety or displace anxiety through the forwarding of letters:

What do these letters have to do with our internet?

Scammers have existed for centuries. Wherever there is power or money to be exploited, scammers will flock. The best example of this from early modern literature is Lazarillo de Tormes, who became the namesake of the Picaresque novel. People always need to get by, and scamming has been an industry for centuries.

That being said, I believe that Chain Letters and Chain E-Mails are direct precursors to phishing schemes and social engineering. What are phishing schemes? The use of fake emails, text messages, or phone calls to trick individuals into sharing private data like passwords, credit card numbers, or bank login details, utilizing many of the same psychological pressure factors as those of the original chain letter.

One of the earliest recognized forms of phishing was the Love Bug Scheme, in which emailboxes around the world were filled with “ILOVEYOU” messages, in which the body of the email simply stated “Kindly check the attached LOVELETTER coming from me.”

What remains the same here? The same emotional tactics to ensure recipients could not resist unearthing the secrets contained within the email, opening what they thought was a harmless .txt file.

Today in age, we have hour-long trainings at work that teach us how not to fall for phishing scams, given that the stakes have changed. After all, some of the largest financial Phishing heists scored over $100 million through social engineering scams by forging emails and invoices to Facebook and Google from legitimate hardware suppliers.

Today, scams have become even more complex. At this point in time, there are more bots on the internet than humans, meaning much of the content being disseminated online is also, in some way, artificially generated. Europol predicts that by the end of 2026, as much as 90% of all content put online could be, in some way, artificially generated. Deepfakes abound, and while we are overconfident in our ability to detect them, it is typically a coin toss, between 50-62% blind test accuracy.

Deepfakes do not just refer to images or videos — in fact, they can refer to voice cloning as well. And as Ashley Flowers of Crime Junkies warned us in a recent episode WARNING: AI Voice Cloning and Virtual Kidnappings, tech-enabled extortion is on the rise. And just like that, the chain letter has evolved to the chain phone call, text, or email.

And just like that, we find ourselves back at the tactics of the early 2000s, but in a different setting, time period, and context. What remains the same is that not only technology will be exploited to urge certain actions and interactions, but the psychology of fear will be manipulated and extorted to encourage certain social movements.

A path forward…

I will be talk about this further in depth at my talk with the After5Society this Wednesday at 7PM. If you are in New York City and want to learn more about the state of the internet in 2026, come join me for a rewind on the history of the Web.

The ideas contained herein are my own and do not reflect those of my employer.

I recently received a comment on an Instagram Reel about a trend where women tend to be more cautious and vocal about AI regulation. This commenter was right: recent studies from Harvard and Northeastern demonstrate a 22% gap in relative AI adoption rate between men and women.

In my research on this topic, I saw some of the most interesting, and in some cases questionable, headlines — even coming from the likes of Harvard and peer reviewed journals. These studies claim that female skepticism towards emerging tech is either: 1) due to the fear that women have towards losing their jobs, or 2) due to an underlying sense of technophobia, or feelings of anxiety that lead to avoidance of the technology. I do, however, feel like we are missing a greater point.

What Does History Have to Say About It?

Historically, women have been more skeptical of emerging technology. Ada Lovelace is our most well-known example, frequently cited as one of the early cautionary whistleblowers about machine intelligence and its inability to originate anything novel: “The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform. It can follow analysis, but it has no power of anticipating any… relations or truths. Its province is to assist us in making available what we are already acquainted with.”

Other fields echo this caution, due to first-hand experience. In 1888, Annie Besant published an article in The Link describing the use of phosphorus in the finishing of matches as leading to necrosis of the jaw, and encouraging movement away from use of the material, which was echoed across the Atlantic by Grace Fryer, who was a watch dial painter instructed to point her brushes with her lips (since then as one of the “Radium Girls”). This well-proceeded the regulation related to phosphorus which occurred 20 years later.

Women also expressed concern around nuclear fission — Lise Meitner, who co-explained nuclear fission with Otto Frisch in 1939, declined to go to Los Alamos to work on the development of an atomic bomb, stating that should would have nothing to do with the creation of a nuclear weapon. Joan Hinton, who did work at Los Alamos and witnessed Hiroshima’s impacts, immediately left the field. In each of these cases women did not have aversion to the physics behind the science, but the risks and institutional oversight that radiated from them.

In 1956, Alice Stewart found that prenatal pelvic X-rays substantially increased the risk of childhood leukemia, which was dismissed by radiology experts for roughly two decades (Women Who Knew Too Much).

The same year, Edris Rice-Wray ran trials on oral contraceptive pills in Puerto Rico, reporting that the side effect burden made the formulation unacceptable at that dose. In the face of this discovery, Doctors Gregory Pincus and John Rock treated these reports as psychosomatic and proceeded. Several disputes followed, including the Nelson Pill Hearings from 1969 to 1970 where women’s reports of thrombosis, stroke, and depression were consistently dismissed and no woman was called to the 1970 Senate hearings.

All of this to say that, not only have women always forewarned of risks and therefore further investigated it, but we have also backed it with facts — first hand or otherwise — only to be dismissed until years or decades later, when the risk became too big to ignore.

Shifting Skepticism from a Weakness to a Strength

Historically, this type of skepticism has also been seen as a sign of weakness. For this reason, perception of women as “averse to technology” has consistently kept women away from developer meetings, board rooms, and policy discussions, for fear that said risk aversion would slow processes down. We can see this in the quantity of times the sayings “Paralysis by Analysis” or “Not Invented Here” have been uttered in a high-ranking meeting.

In Mara Bolis’ “The AI Gender Gap Paradox” she wonders what a world look like where we reframed being mindful about the impact of risks away from being “too cautious” and “nagging” to a world where diligence is seen as strength rather than a perceived weakness. Bolis states plainly: “Women see flaws in systems they have not built, weigh the stakes carefully, and act accordingly.”

In general, because women have historically been excluded from the design and development of new technologies (only 28% of all computer and mathematics jobs are held by women), and because AI is no different (7 out of 9 frontier AI firms are founded or led by men), women are attuned to weaknesses that may have been pointed out had they been in the development room — weaknesses like bias, privacy risks, unreliable outputs. I am positive that had a woman been in the room for the design of Grok’s fantasy chatbot or CrushOn.ai, they may not have been released to the public as they were. After all, as of 2026, 99% of deepfakes in circulation are of women.

As a result of being historically excluded from the development rooms that have shaped our day-to-day interactions with technology, women often express feelings of dread, revulsion, or foreignness when using these products, which leads to more outspoken thoughts on the technology’s shortcomings.

This aspect is what most of the research on why women are more skeptical towards AI misses. Instead of seeing how women have historically been excluded from the design of these products and therefore feel less techno-optimistic affinity towards them, their skepticism is viewed as a crutch and ultimately as something that will lead them to lose their jobs. For examples of this, see: “Technophobia: Gender differences in the adoption of high-technology consumer products,” “Women Are Avoiding AI. Will Their Careers Suffer?,” or “Global Evidence on Gender Gaps and Generative AI”.

Even further, because of this historical exclusion from tech-forward jobs, sources like the ILO predict that women worldwide are 3 times more likely than men to be in positions that can be automated by generative AI due to their relegation to clerical or routine cognitive roles.

While I don’t have solutions to these issues, as they require a larger structural shift, this topic does beg that we sit and ask ourselves the true reasons behind why women are more likely to sound the alarm in human kill-switch situations than men, who tend to lean heavier into techno-optimistic narratives: is it really because of technophobia, or is it because of a history of pattern recognition of risky behavior?

The ideas contained herein are my own and do not reflect those of my employer.

I sometimes wonder what thoughts raced through the minds of John McCarthy, Alan Turing, or Marvin Minsky when first birthed the idea of creating a computer with “humanlike” intelligence. Turing originally asked the question of imitating human intelligence, proposing in 1950 that we abandon the metaphysics of “can machines think” in favor of a more revealing test: could machines converse well enough to pass as one of us? While he died in 1954 before that idea could take shape, two years later, in 1956, John McCarthy went ahead and coined the term “artificial intelligence” for a summer workshop at Dartmouth.

In and of itself, the thought that one could create something on par with human life or intelligence is an incredibly hubristic thought.

Ever since the conception of AI at the 1956 Dartmouth Summer Research Project, “AI suffered from a hype problem.” But the idea of an artificially intelligent entity in the image of humankind had existed for centuries in literature, warning of the dangers of reaching towards the capabilities of god: Frankenstein (1818), Darwin among the Machines (1863), Erewhon (1872), etc.

Mary Shelley’s Frankenstein provided the archetypical cautionary tale for desiring to artificially bear another being into existence without a forethought on the repercussions of what it means to “create” life. Victor’s tragic downfall after the creation of his creature is driven by his hubristic God-playing intentions, ignoring the boundaries of nature, allowing Frankenstein to endure as a parable for the ethics of unchecked scientific advancement.

I opened this article with the creature’s lament: “You, my creator, abhor me; what can I gather from this fellow-sufferer who looks upon me? No sympathizing being meets me.” Victor brings the creature into being and provides him with just enough to learn moral right from wrong, to interpret Paradise Lost, to process the world around him. In the end, Victor abhors what he has created and his creature abhors him back. Very much like our visceral reaction to when we come across AI slop on the internet, those around Victor viewed his creation in the same manner: “My dear Victor, do not waste your time upon this; it is sad trash.”

We are experiencing a parallel in present-day-society, where many have expressed disgust towards AI, or what is now called “AI Paranoia,” living in constant fear of the machine man created and what impending doom it might bring on our natural world.

But what if we had a parable for artificial intelligence from well before contemporary times, dating as far back as ancient civilization, as far back as Troy?

In fact, Shelley references these early archetypes in her novel, paralleling Victor’s acts with that of a Modern Prometheus, Prometheus being the Titan who first stole fire, likened to knowledge or technology, from the gods to give to humanity.

In a shockingly well-timed and -oriented theatrical release, Nolans’ interpretation of The Odyssey similarly suggests that this parable is much older than Shelley’s Frankenstein. Had Minsky or Turing had sat down with Homer’s full epic, they may have seen their own desires reflected back at them.

Homer wrote about The Siren’s allure:

“Come closer, famous Odysseus – Achaea’s pride and glory –/

Moor your ship on our coast so you can hear our song! /

Never has any sailor passed our shores in his black ship /

until he has heard the honeyed voices pouring from our lips.”

The Sirens, while they promise pleasure, they moreover promise knowledge, total and effortless, all while the beach beneath their bodies echoed with the rattle of the bones of those who believed the price for eternal knowledge applied to everyone but themselves.

Homer also wrote of Odysseus’ affront to the Cyclops Polyphemus, son of Podeison, whose curse follows the remainder of his voyage:

“My child, what word has escaped the barrier of your teeth? [...]

Was this not your plan, as you counseled it /

that he would return and punish those men?” (5.22; 5.23).

In the end, Odysseus returns home, but not with a single one of the men who he journeyed out with, a sacrifice he was willing to make because he believed he could defy the gods and still reach Ithaca.

In the 1800s, Shelley extended The Odyssey and Prometheus, envisioning a world in which the creator abhors his creation, and his creation abhors him back. In the 8th Century BCE, Homer provided the foundation of something even more damning - the parable of a voyager who believes his cleverness exempts himself from any further consequence (we might see a parallel here with a certain man named Sam Altman). Odysseus survives The Sirens only because he has himself lashed to the mast, with ropes tightening with every plea to be cut loose.

Terrifyingly, we find ourselves in The Odyssey’s book 12, but with no ropes tying us back to the mast, as Amy Worley mentions in her piece Tied to the Mast; we have pressed forward with no such precaution with artificial intelligence. As honeyed and sycophantic voices pour from the machines we have created, Venture Capitalists and Tech Billionaires continue to funnel money into new AI frontiers, promising a future of abundance, even as far as promising the solution to climate change.

With emerging technology, we find ourselves in Odysseus’ position all over again, sailing under the conviction that on ingenuity alone we can forge forward into the wind.

The ideas contained herein are my own and do not reflect those of my employer.

As part of the research for this post, I confronted my demons and painstakingly analyzed each of the meanest comments I have ever received on social media, all to prove to you that the computer screen has drastically changed the way we interact with other humans. This is something I have long come to terms with since I decided to put my thoughts on the internet in an effort to combat misinformation on social media.

However, this does mean that daily I wake up to comments I might have been better off never reading. To a certain extent, the notifications banner on instagram gives me heart palpitations, because of what it therein contains: “you have nothing interesting to say”; “stop rambling, stop it, stop barraging people with you schizpid content”; “this is the dumbest take I’ve ever heard in my life” ; “go back to the kitchen"; “she is a type of sub-species” ; “I am concerned you teach this and know nothing about it”; “look at this professional LARPer”.

I want to first state that I am not alone: at least 41% of social media users have experienced some form of hateful speech (doesn’t this seem quite low?). I am sure that in informal and less-reported ways, most of us have experienced hateful speech online at one point or another.

Diving head first into producing content online, I have started to wonder: why do people feel emboldened to speak so aggressively on the internet in ways they would never in person? And not to make a generalization, but why specifically do men feel emboldened to speak this way on the internet? I would wager that without the safety of the screen and the keyboard as intermediaries, most of the comments I’ve received would have stayed inside people’s heads.

Of course, I started to research it, as I am so academically inclined.

What I found is that there is little research on the psychology behind why people act this way, even as online “haters” and “internet hating” have widely become a recognized human behavior. Some liken online trolling to the psychological Dark Triad of Narcissism, Psychopathy, and Machiavellianism, linking online hate to a “frustration-aggression” cycle in which frustration may foster aggressive behaviors that are more pronounced online because of the various outlets that social media provides. Each of these parts of the Dark Triad are accompanied by “high levels of impulsivity… impetuous behaviors, such as expressing a negative, insulting attitude/opinion toward someone or something,” which, in my experience, align with many of the characteristics of the online users I have encountered over the past four moths.

The Dark Triad, however, only tells us about the people who are typing. I wanted to go deeper, and for that, I returned to the OG theorists of “injurious speech”: Judith Butler and Jürgen Habermas.

Pardon this brief academic interlude.

What do philosophers and psychoanalysts have to say about injurious speech?

Judith Butler and Jürgen Habermas both wrote extensively on injurious speech, and both understood the "self" in discourse as performative, whether in person or remediated through the screen.

Butler opens Excitable Speech with the question of what it means to be wounded by words. When we claim to have been injured by language, she argues, we ascribe to language an agency, a power to injure, and we position ourselves as the object of its injurious trajectory. For Butler, language can wound because it is not separate from the life of the body. We come into social existence by being addressed, named, called: "a certain kind of surviving takes place in language" (3).

Toni Morrison put it bluntly, stating that oppressive language “does more than represent violence; it is violence” itself.

Jürgen Habermas, conversely, asks what conditions make speech between strangers possible at all. Every genuine dialogue, he argues, points toward mutual understanding. When I speak to you, dear reader, I implicitly claim that what I say is true, that it is appropriate to our relationship, and that I mean it sincerely, and you retain the standing to contest any of those claims.

The online hate commentator abandons this entirely, performing what Habermas deems “strategic action”: speech that treats the other as an object to be moved rather than a partner to be answered. Habermas, who recently passed in 2026, considered his theory in the context of the internet in A New Structural Transformation of the Public Sphere (2022), where he argued that social media dissolved the editorial gatekeeping that once kept public discourse disciplined, fragmenting it into “semi-publics” where every user becomes an author, exempt from the reciprocal accountability that in-person discourse requires.

The Keyboard as a Dis-inhibitor

Psychology backs the computer as a dis-inhibitor. Being online creates a form of “Online Disinhibition” in which, behind the screen, people feel entirely anonymous and removed from the immediate consequences of their actions. In the online message board, people experience a sort of “de-individuation” in which they lose their personal identity within the vastness of the crowd, making them more likely to act in ways that don’t align with their personal values, participating in the social contagion of mob mentality.

Besides these psychological effects, I believe the keyboard acts as an intermediary that removes the person behind the discourse not once but twice: once through not being in the room, and twice through placing oneself behind the keyboard. Internet media theorist Remedios Zafra writes in her book Eyes and Capital [Ojos y capital]: “Because everything that appears to us as absent is implemented in the internet through lack and through desire that this lack evokes… the screen and the keyboard begin to function as an appendix of intimacy and oneself” (103).

This brings us back to Butler’s argument that language and the body can hardly be disconnected, and when they are, online users find themselves in an intensified performative space of disinhibition. In fact, there is a long history of attempting to separate them, and there is documentation of those attempts failing. As Shoshana Felman argues: “if the problem of the human act consists in the relation between language and body, it is because the act is conceived, by performative analysis as well as by psychoanalysis, as that which problematizes at one and the same time the separation and the opposition of the two.” Here she claims that speech redoubles the body, and in the act of saying, the body is the instrument of utterance.

When the physical body is lacking, but language remains, online users find themselves in an intensified performative space of disinhibition. Words, keyboards, and screens become appendices of intimacy, extensions of the physical body on the internet.

The Keyboard Warrior and the Rise of the Manosphere

The internet is a liminal space, or an “‘in-between’ location of cultural action,” that often produces an unpleasant or unmoored feeling. Those of us who have seen Backrooms know that this term is trending in film and media. In parallel ways, the internet has become a space where social norms often do not apply, simply because of the lack of face-to-face interaction. Anonymity tends to embolden people who no longer have to confront the reactions and feelings of the ‘other.’ As Michelle Carey discussed, social media and online message boards become just “words on a screen, devoid of context and humanity.”

This type of dissolution of humanity on the internet has caused the rise of Keyboard Warriors, people who makes abusive or aggressive posts on the internet, while concealing their true identity. The Keyboard Warrior initially came to be through Massively Multiplayer Online Games, viewed as a particular stereotype:

“a gangly teenage boy hypnotized in the moonlight before a computer screen flickering with assorted night elves, dwarves and the forsaken dead” (Herald Tribune).

It is simply counterfactual to claim that online hate speech is segregated solely to male online audiences (see “Are Online Haters Psychopaths?,” which demonstrates that women frequently participate in this kind of behavior as well). However, it is true that hateful speech on the internet tends to align with digital misogyny and what we now recognize as the “manosphere.” According to UN Women, the manosphere consists of a “loose network of communities that claim to address men’s struggles - dating, fitness or fatherhood, for example - but often promote harmful advice and attitudes” in a fundamental misinterpretation of where victimhood lies.

This is best exemplified through the rise of right wing politics online and its crossover with twitch streamers, youtube lives, and podcasters like Joe Rogan, Harrison Sullivan, Nicolas Kenn De Balinthazy, Amrou Fudl, and Justin Waller. These online media presences tend to consolidate around a few main topics: 1) the “alpha” business model and message, often selling courses and online coaching schemes on how to be more manly to their followers, 2) the gamification of life in which social status relies on metrics related to wealth, physical fitness, and sexual dominance, and 3) an appeal to and manipulation of the youth, often drawing in audiences of young men who already feel unsteady about their identity or societal standings (see Netflix’s Louis Theroux: Inside the Manosphere).

The saturation of social media by these types of influencers, along with the rise of right wing/nationalist politicians and creators who echo back the same narratives, has intensified online Social Contagion, in which users, a large swath of whom identify as men, feel emboldened to act out behind their keyboards, with no real consequences to speak of.

In the most facetious of ways, I might say that receiving online hate from men is a sign of great (dis)respect in their culture.

The manosphere is best understood as a business model in search of an asset class. In the 2010s the asset was the alpha “self,” sold back to young men through courses on confidence, female seduction, and physique. By the 2020s the asset became crypto. Andrew Tate’s Hustlers University, later rebranded The Real World, packaged copy-trading and crypto speculation alongside masculinity coaching.

The vocabulary that grew around these self-help courses is gendered and gamified all the way down. “Have fun staying poor” is the crypto-native descendant of “go back to the kitchen,” a taunt that converts and justifies exclusion and transforms it into a function in the market. David Golumbia identified this ideological core back in 2024: these media personalities participate in cyberlibertarianism, a promise of money without institutions, value without labor, wealth as personal sovereignty. There is a reason why Get Rich Quick Schemes flourish in crypto communities.

AI bro culture is an extension of this early crypto culture and cyberlibertarianism, but has simply taken a new form. The prompt-hustle courses that abound (”make $10K a month with ChatGPT”) directly recycle the alpha coaching scheme. Likewise, the effective accelerationism or “techno-optimisim” of artificial intelligence, including the narrative that “AI is inevitable,” recodes caution around artificial intelligence as a weakness itself, an indication that the speaker is part of a new underclass, very similar to narratives around betas emerging the 2010s.

The gendered aspects of these AI narratives online are documented, and insidious on social media platforms like TikTok, Instagram, X, and Telegram, and now put my experience as an “AI Educator” on instagram into context. Sensity’s research found that 96% of deepfakes circulating online were non-consensual intimate imagery, overwhelmingly of women. Meanwhile, AI companion apps sell young men the girlfriend they always wanted, without the personhood, and Meta empowers the same users who create online hate content with the ability to use AI to create from other people’s instagram posts — one of the biggest class action lawsuits waiting to happen. This is where the AI bro rejoins my larger argument.

Our earlier mention of Remedios Zafra’s screen or keyboard as an appendix of intimacy comes to a head with the AI girlfriend and the generation of deepfakes of women: people are able to recreate emotional or physical intimacy with no interlocutor, an other who can never return a validity claim, never contest, never leave. The manosphere always promised relief from the accountability of other people, crypto has and continues to promise money that answers to no institution, and, now, AI promises relationships that answer to no true “self.”

The Politics of Free Speech and Social Media Platforms

There is a reason that social media platforms do so little about this. Rage bait, or social media content designed to create a strong and negative reaction, is a feature and product of the attention economy, because users are quick to “align their views with others on anything that prompts a negative emotion.”

Unfortunately, platforms are designed to increase and amplify the echo chambers of injurious speech online and the social contagion of rage bate because they profit from the engagement. Each mechanism, including which videos are surfaced by the algorithm, which comments appear first based on relevancy, and which videos are suppressed, are explicit results of the design of these applications themselves. In fact, investigative research demonstrate that Meta and TikTok explicitly allow harmful content, including “misogyny and conspiracy theories,” (direct quote from an engineer at Meta, by the way) to be surfaced by the algorithm because of the amount of engagement it receives from angered users.

And while writing this Substack article will not change the makeup of the notifications I received daily on Instagram or TikTok, I believe it is empowering to understand how digital social movements, interactions, and platforms are constructed at their core. And further, this post is a source I can cite when the next anonymous commenter informs me that “women know next to nothing about AI.”

The thoughts contained here are my own and do not reflect my employer.

The queue for WorldID formed in Nairobi, in Jakarta, in the outskirts of São Paulo, all for the chance to look into a silver sphere called the Orb. One looks into it, and behold: it maps the structure of your iris into a hash, and you’re issued a credential meant to prove, for the rest of your life, that you are a human being, and more specifically, you are this human being. In exchange you receive $50 worth of their cryptocurrency token, WRLD.

The company behind the Orb, co-founded by Sam Altman and rebranded from WorldID to World, needed millions of unique irises to make its verification network worth anything. In order to create an ecosystem, and demonstrate Product-Market-Fit, they needed daily active users and hashes already within their operating system/protocol environment.

Getting people inside the queue is the strategy. Once people are in the ecosystem, companies can claim them as active users, engage them through push notifications, and utilize their data.

Kenya suspended World operations in 2023, citing the exchange of cryptocurrency for biometric consent, uncertainty about where the data lived, and the accumulation of citizen data in private hands without any governing framework. Brazil, well-known for its data privacy regulation (LGPD), similarly banned World outright, finding that paying people for iris scans violated the LGPD’s requirement that consent to biometric collection be free, informed, and unequivocal.

Who makes up a useful data point?

This course investigates the idea of what kind of biometric data has been, is, and will be useful to big tech. Joy Boulamwini’s TED Talk sits on our reading list, and her Gender Shades audit with Timnit Gebru remains the cleanest demonstration in the field of a commercial facial analysis system that systemically misclassified dark-skinned women at error rates approaching 35 percent while reading light-skinned men almost flawlessly.

Subscribe now

Set her finding next to the Orb and a terrible feeling may emerge: the faces the industry’s system could not see belong to the same populations whose eyes the industry is now paying out to collect. Often referred to as the Global South, but what I refer to as the “Majority World,” is absent from the ML pipeline at several instances. First, it is the missing faces that cause bias and make AI systems fail evaluations. Then, the Majority World enters the supply chain as cheap biometric data that is often not treated or stored properly. The same logistics that move lithium and labor and rare earth minerals, as Kate Crawford mentions in Atlas of AI, now do the same with the human body.

Let us also remember that the body cannot be reissued: biometrics are forever, the hash of an iris is one of one. German data regulators, reviewing World’s practices, observed that a person who's iris has been captured loses the possibility of anonymity permanently. Now imagine if you were never given a choice. When $50 in crypto is the equivalent of one month’s rent, and you’re not sure where your next meal comes from, the choice is quite simple: submit to the machine.

On the Other Side of the Looking Glass

In late June, Meta launched a new generation of smart glasses, the Kylie Edition, designed with Kylie Jenner, carrying a Meta AI voice modeled on Jenner herself. And all I could ask myself was: how in the world could this possibly be legal?

Read more

I recently spoke with a worker who spent time on Remotasks and Appen, two of the largest data annotation platforms in the world. He left this comment on my survey:

"They paid pennies and could ban you for anything. I wish more people would fill out this survey because almost nobody talks about these conditions or how long it takes to earn anything decent."

His response provided the whole thesis for an article I was writing: workers who were paid pennies, banned arbitrarily with no explanation and no appeal, faced with silence around converting annotation to a livable wage.

This section of our course is about The Human Impact of creating, training, and releasing LLMs into the world.

Here is what another annotator told me:

“You learn to work up to 70 or 80 hours a week, as incredible as it sounds. All of this with the goal of earning more, since the income generated in this type of work tends not to be constant.”

70 to 80 hours a week to compensate for income that is structurally designed to be inconsistent. Gig Work, also known as individual “tasks, short-term project[s], or ongoing work that an independent worker completes for a client” that is not their full-time employer, spells out its very pitfalls in its name. Gigs are irregular, lack benefits, and often are time intensive with very unclear instructions.

A third annotator told me:

“The main challenge I face as a data annotator is the lack of benefits and income uncertainty, which creates constant pressure to balance my finances.”

The lack of benefits, the income uncertainty, the constant financial pressure. These are the conditions under which we are asking humans to train the Large Language Models that will replace them one day.

Read more

The news cycle around technology is shortening every week. Plate-shifting announcements drop on a Tuesday and are buried by Friday. A model launches, a regulation passes, a company testifies before Congress and changes nothing, and somewhere in the feed between those events is a study showing that the system you were just told was neutral has been classifying certain faces as criminals for three years. It is genuinely hard to follow. And it is, quite frankly, stressful.

This stress is a very accurate response to a situation in which consequential decisions around how AI systems are being built, deployed, governed at a speed deliberately designed to outpace scrutiny and exclude the general public.

There are ways through it, because as the age old saying goes: the only way out is through. The correct path is to build an understanding and framework sturdy enough that when the next explosive announcement around AI lands, you know where you need to focus and which questions to ask.

I teach AI Ethics at the City University of New York. I have spent years collecting, arguing with, assigning, and refining the resources that help people build that framework. In this series, I want to share these resources with you.

This is a course that is compressed, direct, and written for people who are alarmed and concerned but incredibly busy. This is for you if you are past the point of accepting AI being marketed to you as “magic” and the framing of critics as alarmist “doomers.”

What’s Inside

This is a five part series in which each post is a self-contained unit. Each of these units combined builds a complete critical framework for understanding AI and its consequences.

Part 1: How AI Actually Works

Before the ethical layer, we need to understand the basic mechanics of AI systems. This is for anyone who has referred to “AI” as one monolith, while wishing they could understand the many layers behind it.

Part 2: Human Impact

Algorithmic bias, facial recognition, healthcare, criminal justice. The through-line connecting all of it is the same: these systems land hardest on people who had the least say in building them.

Part 3: Labor and the Invisible Workforce

These models do not exist in media res. They are trained on human judgment, human attention, and in many cases, human trauma. Most importantly, they are trained by workers whose names never appear in change logs.

Part 4: Decision Makers and Accountability

Who writes the rules, who enforces them, and who gets left out of the room entirely. The EU AI Act, the right to explanation, and the gap between governance on paper and governance in practice.

Part 5: Environmental and Creative Cost

What these systems consume: energy, water, artistic labor, trust in what we see. Deepfakes, AI-generated images, automated weapons, and the infrastructure nobody wants to talk about.

The first part of this series is available now for paid subscribers. You can also purchase the complete pack as a one-time buy: all five posts, the full reading list from my course, and my annotations on each text.

This is the resource I wish existed when I started asking these questions.

The views expressed herein are my own and do not necessarily reflect those of my employer.

How do AI systems work?

Every technology that reorganizes labor, communication, and power eventually gets compared to the printing press, or the industrial revolution, or electricity. The comparison is meant to signal magnitude. To say: this type of system signals a rupture, the kind of thing that changes everything. AI gets all three comparisons, sometimes in the same sentence.

There is something true in it. The printing press redistributed interpretive authority away from institutions that had held it for centuries. The industrial revolution reorganized human bodies in space and time in ways we are still living inside. These were technologies that changed not just what people could do, but who could do it, under what conditions, and whose knowledge counted.

AI is that kind of technology, and so, the comparison earns something.

The problem is that this comparison quickly breaks down. The printing press and the industrial revolution were materially visible to the people they were transforming. A weaver could see the loom. A clergyman could hold the printed book that was displacing his authority. The technology and its logic, however brutal, were followable.

AI operates differently because in many cases, its illegibility is what props it up as a source of authority. Framing AI as “magic” that comes to us in media res is exactly the marketing that keeps us coming back, and this is where we need to start.

What machine learning actually is

Read more

This week, for the first time, I imagined what life would look like without libraries. Not because I am a masochist, but because this was the first time I was actually afraid there could be a future without them.

In 2026, fears around the Dead Internet Theory, which claims that activity and content on the internet will be predominantly created by artificial intelligence, rendering all online content untrustworthy, are more real than ever, particularly in light of several literary faux pas that occurred this week, with unexpected crossovers with Artificial Intelligence. Each of these cases pertains to the unlikely union of venerated authors and literary institutions with AI models like Claude and ChatGPT.

Case Study One: The Serpent in the Grove

Two weeks ago, GRANTA magazine awarded The Commonwealth Prize ($5,000 + short story publication in GRANTA) to “The Serpent in the Grove” by Jamir Nazir, which ex post facto has many academics up in arms, declaring the short story is clearly written by artificial intelligence.

Now, we cannot say 100% that this story was written by AI, but let me tell you, when I read it my senses were tingling. Even in a quick skim, it is riddled with typical linguistic structures output by LLMs: em dashes, single-clause sentences, contrastive negations (it’s not X, it’s Y), participle clauses (verbs ending in -ing), tricolons (three parallel words paired together for emphasis), and the list goes on. Even further, the writing suffers from metaphors that seem to be a simulation of what a model believes a writer should sound like:

“The serpent in the grove was never only a snake. / It was the thing in a man that slid along stone for dark, and the thing in a woman that wrapped a vine around herself and climbed”

“She had the kind of walking that made benches become men.”

Said best by author Amy Katherine, the story is plentiful in “nonsensical, pretentious metaphors that read like MFA writing reflected back to itself over and over in an endless hall of mirrors until you’re left with absolute slop.” Wharton business school Professor Ethan Mollick, well known for his research in AI, similarly called this bluff:

Unfortunately, at this moment, there is no way to fully rule in or out the use of AI in “The Serpent in the Grove,” given that AI detection platforms like TurnItIn are still quite unreliable. But using our best judgment, at least for the time being we can still detect the tell-tale signs of an artificially generated writing pattern.

Case Study Two: The Main Drag of the Early Aughts Turns to AI

Last week, Barnes and Nobles CEO James Daunt announced that it would be open to selling AI-generated books:

“…as long as an AI-written book says it’s an AI-written book and doesn’t pretend to be something else and isn’t ripping off somebody else, as long as that’s clearly stated and the customer wants to buy it, then we will stock them.”

RIP to younger me who used to haunt the local Barnes and Nobles after school to grab a Chocolate Frappuccino and surreptitiously browse the Romance section.

This arrives in tandem with Anthropic’s $1.5 billion settlement in January 2026 over its secretive internal mission “Project Panama,” which spent tens of millions of dollars to buy ~500,000 books, slice their spines apart, scan their contents for training AI models, and then dispose of them. An internal planning document revealed during discovery stated:

“Project Panama is our effort to destructively scan all the books in the world… We don’t want it to be known that we are working on this.”

At least Anthropic knew the public would be disgusted with this internal endeavor. But, what makes the statement from Barnes and Noble CEO all the more surprising is that as of late, we’ve entered a period of significant disdain towards AI: about 70% of US residents believe AI is advancing too fast. And yet, Daunt said the quiet part out loud: “We have 300,000 titles across all of our stores. Do we think that some of those may be AI? The chances are that they are, but we’re not really conscious of them.”

I want explicitly state: it is not normal to rely on a machine for writing that we used to be perfectly capable of doing ourselves before, without wasting one bottle of water per every 100 words generated. There are people, however, who continue to push upon us the inevitability of adapting AI as a society.

Okay, in certain ways it is inevitable. I can now code an application in an afternoon, when before, I could not. That’s all thanks to Claude. But the skills I had prior to ChatGPT’s public release in 2022, the ones that make me the most proud and potentially the most human, are neither enhanced nor replaceable by a machine that favors the most probable outcome.

And yet, the “We Must Adapt” movement now emerges at us from all sides: Reese Witherspoon, Ben Affleck, real estate executives, university admin, and now, a Nobel Prize in Literature Winner herself.

Case Study Three: “Darling, how could we develop this beautifully?”

Polish author Olga Tokarczuk, who won the Nobel Prize in 2018 for her works The Books of Jacob and Flights, acknowledged that she has a premium subscription to an AI chatbot and used it to ideate and revise her latest book:

“When writing my latest novel… I asked this advanced model what kind of songs my protagonists would be listening to at a dance, a few dozen years ago, and AI gave me a few titles… Often I just ask the machine, ‘darling, how could we develop this beautifully?’ Even though I know about hallucinations and many factual errors in the algorithms in terms of economics and hard data, I have to add that in literary fiction this technology is an advantage of unbelievable proportion.”

Without entering all too much into her anthropomorphization of the model by calling it “Darling,” her clear admission that she utilized AI while writing her latest novel is what will keep us all up at night wondering: what other worthy Nobel Prize winners could have been chosen back in 2018?

She continued, describing that she likely would not write another full length novel, but rather just short stories, as it wasn’t worthwhile “from a purely economic perspective.” As if us writers do this for the money (I have a feeling there would be many more of us doing this for a living if that were the case).

This statement brings us to the question of whether or not writing, or any other art form, for that matter, was ever meant to be efficient. While the push to adapt to AI for efficiency purposes is relatively understandable in a corporate environment focused on the bottom line, the arts are distinct.

Having finished my PhD in Languages and Literatures, my understanding has always been that the importance of writing is to share your unique view of the world with others in novel ways. Most often, beautiful art is agonizing, pain-staking, and time-intensive, and it can hurt to write as much as to read (looking at you, A Little Life). Aside from New York Times Bestselling authors, writing is almost never referred to as particularly “lucrative” or “efficient.”

While on fieldwork preparing myself for my defense, my days were often spent in the corners of old stone libraries, peering through stacks of books, inhaling the damp, dusty, bookish air (I love it), and gently palming through old newspapers. These interactions were incredibly human, often requiring me to ask librarians across different countries and languages for assistance in finding the appropriate source.

Much of this process has changed today now that we largely access the archives online. While online archival access certainly is efficient, it does spark other concerns around copyright, IP, and control of the archives (which according to Derrida, should be in the hands of the people, not of a select few companies).

All of this to say that my utopian vision of a Solarpunk future, where we all have access to open and public resources and libraries, is looking more and more like a distant mirage through binoculars.

What I want to investigate now is how long it will take for archives to become entirely useless, either because they have been cannibalized by AI-generated slop, or because our writing has become so flattened that we all begin to write like AI anyways. My sneaking suspicion is we are closer to this horizon than originally thought.

The views expressed herein are my own and do not necessarily reflect those of my employer.

I partnered up with Mozilla.ai to create a “Vibe Check Framework” for vibe- coded features. What does this mean? It means that it reintroduces some of the critical thinking friction back into shipping product features when working with an agent.

Sometimes, when shipping quickly, or shipping as a non-technical PM, there can be disastrous results: exposure of personal API keys, deletion of code, uncontrollable costs due to lack of budget caps, and incomprehensible AI “logic.” Many have suggested a new kind of “vibe testing” of these features before shipping to prod. In response, we’ve created a flow combining Mozilla.ai’s proprietary cq (stack overflow for agents), and a personally coined framework, VIBE√.

Imagine a world in which dating apps cease to exist, where there is no need to meet a potential partner in real life. In fact, there's no need to even try to find your "perfect match," because you can have one factory-made to algorithmic specifications. Every preference becomes programmable. That birthmark you never liked on your partner? Swap it out with a new skin.

Rosalía’s extended Lux album open a portal into precisely this world. The song Novia Robot [Robot Girlfriend] begins with a PSA from ROBOTIKAS, a fictional multinational manufacturer of robot girlfriends designed explicitly for the male gaze:

Welcome to // ROBOTIKAS // with a “K” // A world of // female robotic // fantasy, made // for the // pleasure of the // opposite sex

The scenario Rosalía sketches parallels a technological reality already taking shape. Multinational corporations now manufacture humanoid robot companions with fully replaceable skins, busts, and faces, and interchangeable personalities driven by AI. These companions learn from user preferences in the same way ChatGPT or Claude do, becoming progressively attuned to individual needs, superseding the basic desire for human connection.

“You look lonely, I can fix that…”

Media has explored this territory since at least the 1800s. H.G. Wells classic The Time Machine, depicted human cognitive faculties atrophying under total automation. In contemporary iterations, Spike Jonze’s 2013 Her, Denis Villeneuve’s Blade Runner 2049 [where this section’s subtitle comes from], or Kazou Ishiguro’s book Klara and the Sun all navigate themes of robotic companionship and manufactured intimacy.

Rosalía, known for flamenco-style music inspired by Romani culture, has built a practice around embedding esoteric and academic elements within catchy compositions. Her previous album El Mal Querer [The Bad Loving] investigated the psychic depths of inter-partner violence and the compulsion to return to one’s abuser. “Novia Robot” extends this feminist lineage while invoking new theoretical frameworks.

The track summons Donna Haraway’s 1985 A Cyborg Manifesto and the VNS Matrix’s 1991 Cyberfeminist Manifesto for the 21st Century, both of which urged the creation of feminist regimes existing between organism and machine. Haraway’s cyborg identity aimed to dismantle binary constructions of gender, nature, and technology. Yet where Haraway imagined liberatory potential in the fusion of body and machine, Rosalía’s ROBOTIKAS exposes how this fusion has been captured by capital and routed through the logics of male desire.

“I do hope you’re satisfied with our product”

Novia Robot operates as satire, but the real world has outpaced this parody. Robot Girlfriends already exist. RealDoll offers replaceable skins, faces, and busts. The base model costs $550. The company has performed well enough to secure acquisition by Tokens.com and enter public markets on the CBOE Canada Exchange under the ticker XBOTF.

The rise of these companies tracks alongside the meteoric expansion of the "Manosphere," a global network of online communities promoting misogyny, anti-feminism, and toxic masculinity. Podcasting culture, particularly the Financial Self-Help Guru and Sigma narratives aligned with Crypto Bro culture, has amplified these spaces. RealDoll's acquisition by Tokens.com positions the company within token economies that rely on Proof of Stake (PoS) systems, creating circularity and time-bound contracts that keep capital invested. The connection is structural: lonely, younger male generations, who become perpetual stakeholders in their own isolation.

RealDoll’s original visibility came through Howard Stern, the libertarian podcaster known for extreme media performances oftentimes credited as the first manosphere influencer, who featured the company and marveled: “they literally had a giant poster chart of nipples. There are like 20 different nipples you can choose from.” The anecdote captures the reduction of intimacy to a consumer catalog, desire modularized into purchasable options.

Jonze’s 2013 Her predicted this trend. At a certain point in the film, Twombly’s ex-wife, Catherine, explains the cold reality of his new AI companion: “You always wanted to have a wife without the challenges of actually dealing with anything real…” In the same manner, companies offer a static version of human companionship, one that removes all the friction of real relationships, which is the reason we see that 1 in 4 young adults believe AI partners could replace real life romance.

What becomes legible through this architecture is a familiar pattern: whatever can be commodified will be commodified. Natural resources gave way to data. Now the frontier includes intelligence, reasoning, humanity, and companionship itself. The market logic targets men who have internalized the narrative that their needs will no longer be met by women in this "modern day and age."

The critical gap remains the question of sentience. Artificial Intelligence does not possess consciousness. It utilizes probabilistic reasoning to output the most statistically likely response. Yet RealDoll continues expanding its features. The company recently integrated an AI Vision System called Vinci that maintains eye contact and tracks its user. The company markets this as the ability to “develop genuine connection,” though the mechanism mirrors tracking trends in wearables and surveillance technology in which “connection” and “personalization” become synonyms for data collection.

Rosalía's Novia Robot makes visible what remains partially hidden in techno-utopian narratives about AI companionship. The cyborg fusion Haraway theorized as liberatory has been rerouted through capital into the production of gendered synthetic labor. ROBOTIKAS, as a fictional entity, performs the work of exposure: the song itself is an advertisement, a corporate testimonial of seduction. By inhabiting the voice of the manufacturer, Rosalía reveals the infrastructure subtending promises of connection: male desire engineered into product specifications, loneliness converted into market opportunity, intimacy reimagined as service.

The views expressed herein are my own and do not necessarily reflect those of my employer.

“Imitation is the sincerest form of flattery” may as well be one of the refrains we heard the most growing up. There is truth to it, though. Remixing other people’s work is one of the most human things we can do. We watch a Paul Thomas Anderson movie, and suddenly we’re writing a blog post inspired by it. We read Joan Didion and soon have an itch to write a poetic memoir.

Philosophers and linguists alike have belabored this theme — Walter Benjamin’s The Work of Art in the Mechanical Age of Reproduction, Jorge Luis Borges’ Homeric Versions, Paul Rabinow’s Marking Time, Oswald de Andrade’s Manifesto antropófago. To consume art is to cannibalize it through your own interpretation, and that is the beauty.

However, there is something new. Now, just about anyone can remediate anything, without more thought than a partially formed sentence: I tell Claude to “write instagram caption about how waiting in lines in NYC is yuppy-esque, in the voice of Anthony Bourdain.” And just like that, I’ve got an answer generated in seconds:

“Standing in line for an hour for a croissant isn’t about the croissant. It’s about being seen standing in line for an hour for a croissant. The best meal I ever had in this city cost $4 and came through a window I had to shout my order into. Nobody was taking pictures.”

I’m not as concerned with the caption itself — if you have seen No Reservations or Parts Unknown — you know this pretty much sounds like Bourdain. What is unsettling is the fact that it took next to no effort, when if I generated this myself, it might take me a few drafts and a text to a friend to gut check me.

Subscribe now

Unfortunately, it’s no longer a party trick to see what “fun things” ChatGPT can do for you. In 2024, we crossed a line as a species: more new articles published online were written by AI than humans, 50.3%, compared to 5% prior to ChatGPT’s launch. Ilia Shumailov (formerly Vector Institute and Google DeepMind) and his research team found that out of 900,000 newly created webpages in April 2025, roughly 74.2% contained some form of AI-generated content. That means as you’re scrolling through the news right now, even within Substack, you’re reading something written by a machine more often than a person.

Unfortunately, AI text detection tools are only as good as a standard classification model, and the more AI-Generated content published online, the harder it will be to distinguish between artificial and organic content. Even if a creator claims to not utilize AI in their daily work, the mere reliance on definitions obtained from Gemini’s AI summary search already skews our search for knowledge.

The Social Media Problem

Unfortunately, the proliferation of AI-generated content has extended to most social media platforms. Some Instagram influencers that have amassed more than one million followers — see Mag do Lu, Lil Miquela, and Rebecka Hemse — are clearly Deepfakes. This means beauty advice, lifestyle tips, and trends, even, are coming directly from statistical models. Effectively, the trained model is now training us.

The worst of this is that this artificial content often performs better than that of human influencers. AI-generated content is designed to be the most middle of the line, probabilistic output, formulated to engage the general public on social media. When put that way, of course this content will proliferate. If content, in the end, is just a box inside of a larger container, does it matter which box is human and which box is AI?

We’re Training Ourselves

Watch any creator, it doesn’t matter the genre or platform, and you will start to notice patterns: Come with me to try the most delicious croissant in lower manhattan; I am so happy to be on this brand trip with [brand name]; If you’re not doing [XYZ] to maximize your life, you’re falling behind.

The emphatic prosody is everywhere: stretching out certain words for emphasis, raising pitch on important parts of the sentence, that certain cadence and verbal hooks that just keep you watching. Linguist Etymology Nerd has termed this “algospeak,” which is when speaking patterns are influenced by social media algorithms.

Algorithmic conditioning by platforms like Instagram, TikTok, and YouTube reward content that keeps viewers watching. Any content that churns, dies. What becomes clear is that above all in English, we are beginning to flatten linguistic patterns to adhere to an all-seeing algorithm. This linguistic flattening is parallel to the flattened answers we receive from AI models, which present us with only one or two ways of doing things, when in reality there are many different paths to get to an endpoint.

At a conference in Bologna last summer, an academic raised his hand to ask a question to the panelist. He prefaced this question with: “For reference, I never use AI for my research or writing, and I never have.” Recently, I’ve been thinking about how we all have utilized AI in our work, whether we know it or not. Whether it’s the answer you rely on from Google’s AI search summary, or the Medium article you read about a new topic, there is at least a 74% chance part of it was artificial information.

The effect of the proliferation of artificial content is a form of archival poisoning, in which we fundamentally change our output has humans, driving lines of thought into a flat, centralized pattern. If we look at machines, this resembles model collapse — this is when a model is fed so much degraded synthetic data that it is unable to produce adequate answers. What about when this synthetic degradation process happens to humans?

This degraded feedback loop is already in place. Here is what our current creative patterns trend towards: 1) AI trains on human writing, 2) humans consume AI outputs and absorb patterns, 3) humans write like or with machines, 4) the machine trains on flattened data.

This patterns leads to the perfect conditions for model collapse, and similarly, the collapse of humans as critical thinkers. In the early 2020s, we trained models upon our writing, content, and habits. In the early 2030s, might we be the ones trained by the output of the models. This is not what we traditionally call imitation or remixing of content in a new mechanical age — it is more akin to a knowledge cannibalization process that creates entirely synthetic outputs.

You Know You’re Not Allowed to Use ChatGPT to Write Poems, Right?

I will leave you with an excerpt of a poem that Jon Sands recently wrote about this very problem, to contemplate the very question of what it means to make something that is completely human.

The views expressed herein are my own and do not necessarily reflect those of my employer.

My morning started like any other. I woke up at 7am, made some coffee, and as I pondered what my day would look like, I had an idea. When I have an idea, I often need to get up and execute it immediately. I attribute this to a mild case of anxiety towards academic achievement.

Today, my grand idea: make a Conan O’Brien Chatbot.

Before I get into the technical process, I want to sit with what this impulse actually means. The ease with which I can now create a digital twin of someone I’ve never met raises some questions about para-social relationships, consent, and the strange asymmetry of contemporary celebrity culture (See the podcast Beth’s Dead, for reference).

In many ways, I feel like Conan is my friend, and that is much of the basis of his present day show, Conan O’Brien Needs a Friend. Back during COVID, I became an avid podcast consumer, and Conan O'Brien Needs a Friend had hundreds of hours of content for me to binge. I genuinely love Conan's humor, and the podcast had the extra charm of two new voices: Sona Movsesian and Matt Gourley. Since then, one of my closest friends and I exchange texts weekly about Conan's podcasts.

We even had the good luck of seeing him live, TWICE, with the extra surprise of Paul Rudd coming on stage:

But here’s the crux: Conan does not even know I exist, and yet I’ve consumed hundreds of hours of his voice, his humor, his patterns of thought. In surveillance capitalism terms, I’ve been the product and the consumer simultaneously: absorbing his content while platforms extract value from my attention. Now I’m flipping the script, becoming a kind of amateur data engineer, extracting patterns from his public performances to reconstruct a simulacrum of his personality.

I think you’re getting the picture now. For me, Conan is a big deal. I’ve tried many times over the years to get picked for one of his Conan O’Brien Needs a Fan episodes, where he talks with fans across the globe who have interesting questions for him. No luck yet.

And that’s when the idea came to me for this chatbot. In this day and age, digital twins are actually not a luxury, but quite easy to make. So I pulled out my laptop, iCloud storage full, screen cracked and all, and got to work.

Let me walk you through my process.

MISSION CONAN

NOTE: I largely relied on Claude and Claude Code during this for help with code generation and double checks. An interesting note on this — when I proposed my idea to Claude, I received the following clarification question:

I mention this because it seems that my Claude assistant is INCREDIBLY well-tailored to my interests, personality, and academic pursuits, as this question related to surveillance capitalism is one that went through my mind while brainstorming this idea. More on this later.

Step 1: Data Collection, Cleansing, Annotation

The first step involved data collection, cleansing, and annotation. I went through transcripts from 20 different Conan O’Brien needs a friend episodes, created files for them, cleaned them with a quick python script, and started on my data annotation journey. As a few of you know, I used to work at a data annotation firm, and so I’ve had quite a bit of experience working on different annotation projects and schemas.

The data annotation process involved pulling out Conan’s most common characteristics — self-deprecating humor about his height, references to Harvard with undercuts, absurdist scenarios that escalate, mentions to his time on late night, his hatred towards Jordan Schlansky. I even analyzed 20-30 different joke structures from the 20 podcast episodes I surveyed:

I extracted several different patterns related to ‘height’, ‘family’, ‘body_awkwardness’, and ‘armenia_sona’. I added his verbal tics: ‘here’s the thing’, ‘and this is true…”, ‘hold on… no, no, no…’

Step 2: Voice Cloning

My computer is super old so I decided to create a voice on Eleven Labs and call the API in order to integrate Conan’s voice into the chatbot. I want to flag here that ‘My Voices’ (see below in fig 5) is not a correct embodiment of the voice I created for Conan, as I do not own Conan’s voice and would not encourage anyone to steal the likeness of someone’s voice.

Step 3: System Prompt Engineering

Finally, it was time to create my AI system prompt, which consisted of different cues like Response Length, Behavioral Indications with Examples, Frequency Instructions for the Chatbot’s Voice.

STEP 4: Testing and Interface Design

It was finally time to test out my work. I ran python conan_voice.py to see how I’d done. After a few fails, issues with reproducing audio in response to the questions I prompted it, I started to hear from Conan himself. But I couldn’t keep this hidden behind walls of code — I wanted to create a UX for it. So I pulled a few UX/UI elements from 21stDev and provided a product document for how I wanted the chat interface to look. Various shades of orange, smooth, sleek, and responsive to Conan’s voice.

Here’s the final product!

Final Thoughts on Digital Twins, Consent, and the Future of Identity

The chatbot I've created is solely for curiosity and personal experimentation. It will never be for public use, especially since I do not have Conan's explicit consent or rights to his voice. This constraint matters. This goes for any time you create a new digital twin, voice, or chatbot in the likeness of anyone else.

But let's push further on what this project reveals about our current technological moment. The internet has created new ways of signifying our identity and the identity of others. In the era of deepfakes, voice cloning, and AI-generated content, we face urgent questions about privacy and ownership over our digital footprint. Who owns the patterns that make up a personality? Who has the right to replicate the cadence of someone's speech, the structure of their jokes, the rhythm of their thoughts?

I don’t entirely have clear answers to these questions. I do, however, have deepened awareness that every technical choice carries ethical weight, even if it just seems like a fun, convenient thing to do, and capability to do something, is not the same as permission to do it.

The views expressed herein are my own and do not necessarily reflect those of my employer.

Debates over which words the Real Academia Española recognizes as legitimate Spanish resurface periodically, and with them, broader questions about who gets to define the language. For example, the verb parquear is part of the daily lexicon of Spanish in the Caribbean and Central America. It only officially became a recognized word by the RAE in 2001, with the indicator “Am.” rather than “Es.” This word has long been viewed as an informal deviation and anglicism of the “correct” word, estacionar. As in most institutions, “legitimacy” travels slowly when crossing the Atlantic.

The instagram creator, @panahispana, recently challenged her followers: stop relying exclusively on standardized Spanish dictionaries, aligned to Peninsular Spanish, especially if you live in or speak Spanish from the Caribbean or South America. She provides five alternate dictionary sources to the Diccionario de la lengua española (DLE) from the RAE, the Royal Spanish Academy:

• Diccionario de Americanismos

• Diccionario del Español de México

• Diccionario del Uso del Español de María Moliner

• Diccionario Panhispánico de Dudas

• Fundéu

• El Banco de Datos de la RAE, CORPES, CDH, CREA y CORDE, and

• El Centro Virtual Cervantes.

Each of these offers distinct cultural and contextual frameworks. The Diccionario de Americanismos contains 70,000 terms used across the Americas. The Diccionario del Español de México centers on Mexican usage, some of which resonants across Latin America. The María Moliner dictionary provides contextual, descriptive, and agile guidance on usage that feels more alive than the institutional tone of the DLE.

The call to move beyond exclusive reliance on the DLE is not anti-institutional — some of the sources above also come from the RAE. But it is contextual. It asks us to recognize that “standard Spanish” is not lived, or does not exist, for millions of speakers.

Let’s take the following excerpt as an example:

“[T]he phrase “Anda a tu casa a fregar” can only be interpreted as “Go home to scrub (the dishes)” by people from Spain, which can contain a misogynous connotation. On the other hand, for people speaking Spanish in Ecuador, it may be mostly interpreted as “Go home to have some fun” or “Go home to annoy (other people)”, which are not related to discriminatory connotations.” — Analyzing Zero-Shot transfer Scenarios across Spanish variants for Hate Speech Detection

There are infinite examples as such, in which meaning is not fixed but situated.

We don’t even need to reach for more infamous examples like coger, a neutral verb for “to take” in Spain but sexually explicit in parts of Latin America, to understand how easily interpretation fractures across borders.

Institutions have long been averse to formal changes based on every-day dialect drift, and this is not limited to the RAE or even Spanish itself. In France, the Académie Française has long been known to “attempt to maintain the purity of the language by resisting such trends as slang, or argot, as well as the Americanization of common parts of speech” (The Usage Wars à l’Envers). Verlan, a linguistic trend that emerged out of protests, is commonly used in every day French. In this linguistic structure, standard words are inverted: femme becomes meuf, chaud becomes auche, and so on. The Académie, regardless of how present Verlan is in every-day speak, resists integration of this linguistic structure into the dictionary, seeing its duty as setting down “the French language, to give it rules, to render it pure and comprehensible for all” (L’histoire, Académie Française).

In parallel ways to the Real Academia Española, this resistance creates blind spots for the multilingual models we train for translation purposes, enforcing contextual blindness and temporal drift.

Of course, these differences are important and change conversations — but if we consider them within where we presently sit, a society absolutely saturated with artificial intelligence models, whether needed or not, the stakes changes. The contextual collapse of the use of language online, and with artificial intelligence systems, begs us to seek different ways of grounding constantly-changing language online.

Content moderation is a clear example. Detecting online hate speech in Spanish involves both human reviewers and machine learning systems scanning vast quantities of text posted daily (see American Sweatshop if further interested). Meanwhile, Section 230 of the US Communications Decency Act shields platforms from liability for user generated content.

When Spanish (English, or French, for that matter) is treated as a monolith, models inherit that simplification. Yet Spanish spans 22 countries as an official language, with significant morphosyntactic and lexical variation. If an algorithm is trained on Peninsular Spanish, it will interpret anda a fregar or voy a coger el autobús through that lens, regardless of whether their speaker is in Quito or Distrito Federal.

Multilingual models like mBERT, short for the Multilingual Bidirectional Encoder Representations from Transformers, were trained on large corpuses of data, including Wikipedia. Research indicates that a disproportionate percentage of Spanish-language Wikipedia edits originate from Spain, roughly 40%. While Spain is indeed the fourth largest Spanish-speaking country by population, Mexico has the highest number of Spanish speakers globally. Yet, Mexican Spanish is underrepresented in the data that shape “neutral” or “winner” MLLMs.

Newer hate-speech-specific corpora such as HaSCoSVa-2022 (corpus of tweets related to hate speech towards immigrants written in Spanish), and Detección Misogninia (DetMis, which contains 35,000 tweets geo-located to Mexico), attempt to address this imbalance and their findings are telling. Models trained on cross-variant data over-flag misogyny and xenophobia when none is present, while models trained on target-variant data out perform significantly better.

This brings me to an experiment I built.

If institutional dictionaries privilege one variant of Spanish, and AI systems trained on skewed corpora disproportionately flag others, the remaining question is whether these hierarchies are also embedded in us.

To explore this, I have developed an Implicit Association Test (IAT) in Spanish. The IAT is a reaction-time task: words appear on screen one at a time, and you sort them into categories as fast as possible using two keys. The logic is simple: when two categories feel cognitively compatible, sorting them together is fast. When they feel mismatched, you slow down, even if you would never consciously say they are related. That difference in milliseconds, calculated as a D-score, is the measure of implicit association.

This version of the test pairs two types of stimuli: words that belong to one variant of Spanish (Latin American regional forms like parquear, celular, ahorita) versus words that belong to the RAE standard (Peninsular forms like estacionar, el móvil, enseguida), and attributes that signal legitimacy (correcto, académico, formal) versus informality (marginal, dialectal, no estándar). In the key rounds, these categories are combined. If you are faster when Peninsular forms and legitimacy share a key than when Latin American forms and legitimacy share a key, the test registers an implicit prestige association aligned with the RAE hierarchy. The question is not whether you consciously believe one variant is superior. It is whether that belief has been absorbed at the level of automatic judgment, before conscious thought has a chance to intervene.

Try it here. My hypothesis is that the results will suggest that linguistic hierarchy is not confined to institutional authority, because it rests first and foremost with human cognition.

Methodological note: this is an abbreviated IAT (5 blocks, fewer trials than the full 7-block lab version), so individual D-scores will carry more variance than a formally administered test. Treat your result as a directional signal, not a clinical measurement.

The views expressed herein are my own and do not necessarily reflect those of my employer.

I’ve gotten an uptick in requests on LinkedIn, from student groups wanting me to join panels, and from individuals hoping to grab a virtual coffee, all circling the same existential question: how will Artificial Intelligence change the job market? I use will rather than may because, at this point, we are already experiencing the tip of the iceberg. My goal is to assuage some of the fear that undergraduates and graduate students alike are feeling right now.

The truth is, we have already reached a point of no return in which AI models can complete tasks with graduate-level reasoning. But up until this point, models only imitate human behavior, with an incredible probabilistic output.

MoltBook, for example, a social media platform created by and for AI agents, based on the OpenClaw gitbhub repo, gained attention over the past few weeks as agents began to interact with each other and demonstrate signs of what many called “consciousness.”

What becomes clear, however, is that agents can be trained and prompted specifically on data related to human consciousness, and therefore produce outputs related to it, mimicking the human experience. If a chatbot is prompted, “Write a post about an agent-led study related to Consciousness Detection in the form of a breaking news post,” we might see an output like Figure 1.

At the crossroads of psychoanalysis and neuroscience, consciousness is not only a by-product of a normally functioning brain (A Turing Machine), but also, the possession of some greater moral agency through language or action that distinguishes the being, from an inanimate object. We can think of agency as a step below: computation has not reached full consciousness but has developed decision-making capability with actions to follow.

All of this is to say: we are a ways away from claiming AI consciousness and its proximity to the human experience. And because of that, the number one piece of advice I provide to people entering the job market is to focus on skills that are absolutely human and unachievable through a probabilistic output alone.

Alysa Liu’s show-stopping long- and short-form Olympics performance, which earned her a gold medal, is a perfect example. @thekathrynkeyes perfectly captured this on threads, stating “I work in tech, so I constantly hear whisperings of ai taking over everything. but can it produce the essence of girlhood and whimsy like this?” referring to Liu’s gold medal performance.

She is right. We cannot let the consciousness illusion of artificial intelligence distance us from what is distinctly human — enjoyment, love, yearning, longing, creation, desire.

A recent Hermès campaign embodied this: in the wake of AI slop taking over the internet, Hermes went back to its artistic roots, demonstrating that artisanally-made products are the “ultimate luxury.” Imagine that: a living, breathing artistic creation.

In Fig 3, we see various pointillism pieces by artist Linda Merad, painstakingly crafted by hand, dot by dot, then animated to achieve a stop-motion effect. They are brimming with life and vibrancy, a deliberate contrast to the hyper-smooth, perfectly symmetrical, and uncanny valley-esque images that emerge from tools like Midjourney. Hermès is just one in a series of brands signaling something larger: the luxury of being human.

These examples represent a broader trend we are seeing in the market — a return to human, analogue intelligence — accelerated, in part, by growing awareness of the environmental cost of hyper-utilizing AI models. We can map these trends onto the corporate job market. I predict we’ll see a few different shifts in the coming months and years:

1. Major in being human. Claude Code and Codex can write accurate lines of code, teach you how to run an experiment, how to set up an environment, or create a MATLAB plot. What they can’t teach you is critical thinking, creativity, or how to approach a problem that has never been seen before. In 2025, total US computer science job postings decined 35% from 2024. Many claim that the “Junior Developer” role is dead, along with many other entry-level roles. The goal is to master skills that are distinctly human, starting with the question: “What are our core capabilities as humans?”

2. Understand what Human-Computer Interaction looks like in the age of AI. More specifically, what will Human-Agent Interaction be over the next five years? I’ve seen an explosion in PhD and post-doctoral positions investigating just this question.

3. Lean into creativity. There are elements of art that simply cannot be recreated by machines, not yet, and perhaps not ever in the ways that matter. Human-made art may be a luxury. Learning a trade, mastering it, the way certain cultures have passed down craft for generations, will be a form of anthropological preservation. Continue to draw, paint, and write. In the future, it may be the most human thing we have left.

So, as you approach graduation, don’t sit at home worried. Keep creating: research, art, novel thoughts, relationships. The knowledge economy, while increasingly shaped by technology, it is giving way to something new, a relationship economy, in which soft skills are becoming invaluable, and increasingly rare.

“The things we want are transformative, and we don't know or only think we know what is on the other side of that transformation.”

— Rebecca Solnit, A Field Guide to Getting Lost

The views expressed herein are my own and do not necessarily reflect those of my employer.

In Spanish, there is a difference between te quiero and te amo: each carries a different level of intensity when expressing that you love someone. Saying te amo to someone you’ve just started seeing could, potentially, end the relationship. In Catalan, there is only one verb used for both: t’estimo. The variation is limited, and the meaning or intensity of the statement relies on the intonation and the context within which it’s said.

This kind of linguistic nuance is often lost in translation, both in human interpretation and machine translation. Right now, LLM performance in languages like English, Spanish, and Chinese functions at an all-time high, largely due to the massive number of high- quality datasets available in standard forms of each of these languages. English, for example, accounts for 92.099% of the training corpus of ChatGPT.

Google's C4 corpus, a colossal collection of data scraped from internet crawlers, includes training data in 108 languages, but still demonstrates the wide gap in language distribution for training data, with English encompassing 45.1% of the dataset.

There are 1.5 billion English speakers worldwide. Let’s put this into perspective with a lower resource language: Catalan, which has approximately 9 - 10 million speakers worldwide. Catalan is the native language of Catalonia, the Northeastern part of Spain where Barcelona is located. The language is spoken across Catalonia, Andorra, Valencia, and the Balearic Islands. For reference, Catalan does not even appear in the aforementioned C4 Dataset.

LLM training datasets are created through a variety of methods: some are generated from scratch utilizing the Bender Rule, gathering audio and text data from speakers across the globe; some are from translated texts of Spanish sources like Wikipedia; some from government websites like gen.cat; and the rest from web crawlers on sites like Booking.com. The English-Catalan ratio of 150:1 tells us that the sheer amount of data generated daily in English far exceeds that of Catalan.

Research reiterates this idea. When searching for corpora of texts to train and evaluate an LLM, I found very few: Vikipèdia, FLORES-200 (Catalan - Spanish), Official Gazette of the Catalan Government, Catalan Open Subtitles, CaWac Corpus (2013), Mozilla Common Voice, as well as lesser quality datasets scrapped from the web. The size of these is roughly half that of other well-performing training datasets in other languages.

Researchers make due with this lack of resources by relying on the “semantic textual similarity” between Catalan and Spanish, which, to a certain extent is true: “Spanish and Catalan are two Romance languages with similar syntactic structure, and that – since Spanish and Catalan society and cultural environments greatly overlap - many Named Entities appear in both language corpora” (Named Entity Recognition for Catalan Using Spanish Resources).

Unfortunately, relying on the similarity between Spanish and Catalan is simply not sufficient to capture the linguistic complexities of Catalan. Some differences are small, for example: “què” versus “qué” (acute versus grave). Some are great: “tengo que” versus “ he de” (I have to). And remember t’estimo? That entire spectrum of emotional expression between te quiero and te amo exists in Catalan too, but it’s just encoded differently, through context and intonation rather than vocabulary.

Building Datasets from the Ground Up

So where do we go from here? It becomes important to take grassroots action when facing the potential linguistic flattening of entire regions through the use of English-language-focused LLMs. In fact, with an active community, datasets are relatively simple to collect in both text and audio format.

This could look like recording a five-minute conversation between you and your best friend, or sharing an essay you wrote for your 10th grade history class as relevant text data. In Catalonia specifically, the movement to revitalize Catalan among younger speakers has resulted in the creation of several language revitalization networks like the VxL Programme and El Consorci per a la Normalització Lingüística, as well as less formal networks through instagram accounts like @dailycatalan, @serapisoler, and @urijimenezzz_. These communities have shown the potential for small, grassroots organizations to make a meaningful difference in language representation across Catalonia.

We do, however, need more voices, more data, and more examples of how real people utilize Catalan in their daily lives. Not just formal government documents or translated articles, but the messy, very human, content-dependent way that language actually works.

If you speak a lower resource language, whether it’s Catalan, Galician, Basque, or any of the thousands of languages underrepresented in AI training data, I want to hear from you. Drop a comment below and let’s start building something together.

The views expressed herein are my own and do not necessarily reflect those of my employer.

As I was watching the Super Bowl Halftime show this weekend, I found myself thinking about how I would translate some of my favorite lyrics by Bad Bunny. How does one convey not only the meaning of El apagón (the blackout), but also capture the dense linguistic, cultural, and political nuances embedded within it? And if a human struggles to do this well, how could a large language model succeed?

“Pichea Maldiva’, yo me quedo en Palomino” might be translated as “Forget about the Maldives, I’ll stay in Palomino,” referring to a small island off the Northern coast of Puerto Rico. This is among the easier lines to translate. Many of Bad Bunny’s songs rely on words like tití, Nuevayol, bichiyal, bichota, among others: all words that are resistant to standardization. Some songs reference Taíno and Jíbaro folklore through expressions like lelolai (Lo que le pasó a Hawaii) and place-based allusions such as Jacinto (WElTiTa), referring to a natural opening called Jacinto’s Well on Jobos Beach in the town of Isabela, Puerto Rico.

Across the Americas, there are more than twenty-nine different dialects of Spanish, in addition to various dialects of English, Creole, and Pidgin. Each of these varieties differ in a multitude of ways, namely the use of forms of address like tú, usted, or vos, as well as slang and semantic variation. In Perú, the verb parar (to stop) can mean to hang out with friends. In Argentinian Spanish, it means to stand on your feet or stop. These distinctions are meaningful, but they are often flattened in automated translation.

AI models frequently fail to recognize these differences. Even LLMs that are recognized as leaders in multilingual reasoning, like Le Chat, struggle with low-resources languages or dialects. Multilingual models are trained on parallel corpora across many different language pairs. However, when available online data for a given language is sparse, the default is to draw on information from a similar language (a lender language) in the training data that might have more resources. The result is a very standardized, homogenous version of Spanish as an output. This phenomenon can be described as the ChatGPTfication of language.

Puerto Rican Spanish illustrates this problem clearly. Because much of this dialect is either practiced orally or through informal methods of communication online it is often dismissed as a “non-standard ways of expressing oneself.” This has created a dearth of sources for AI scientists to utilize to train and test models across dialects as a result. This absence of data is not neutral.

Research has already demonstrated the types of linguistic biases detectable in AI outputs. One study found that while humans typically utilize the sentence structure “it’s not x, but it’s y” only five times per 1,000 words, ChatGPT utilizes it twenty-seven times per 1,000 words. If you’re sick and tired of reading the word “delve” in every cover letter that you review, you are justified! Studies demonstrate that usage of the word “delve” in published papers is 100 times more prevalent than it was just a few years ago due to outsourced annotation of the English language to non-native speakers.

Just as in English, we see an over-reliance on certain phrases that are overrepresented in online corpuses of Spanish, including subrayar la importancia de, or profundizar en las complejidades. Given this, it’s difficult to imagine how a model trained on lender language data could grasp the cultural, historical, and political stakes of a song like Bad Bunny’s El apagón.

A recent paper, “Spanish is not just one: A dataset of Spanish dialect recognition for LLMs,” presented a large swath of datasets designed to assess the capability of Large Language Models in various Spanish dialects, which can be utilized to train, deploy, and test AI model performance across languages. This is an important step. Dialectical evaluation remains largely unexplored, and without it, claims about multilingual competence are incomplete.

For Puerto Rican Spanish in particular, the available NLP resources are limited. Existing datasets include the OpenSLR Puerto Rican Spanish Dataset, Google Research Crowdsourced Dataset, and #Somos600M, with varied coverage across audio and text. Spanish is spoken by roughly 600 million people worldwide, but there are only a handful of open datasets that can be utilized to evaluate LLMs, and there are no leaderboards to compare model performance like there are in English. Basic questions remain unanswered, such as whether one system succeed at translating dialectical Spanish better than another.

There is, however, a path forward. Mitigating the flattening and erasure of dialects requires deliberate investment in listening to, supporting, and amplifying different voices, and non- “standardized” forms of language. So, watch that Bad Bunny halftime show again. Then consider what is lost when you read its closed captioning translation.

The views expressed herein are my own and do not necessarily reflect those of my employer.

Artificial intelligence has always cycled between promise and collapse: the so-called AI “springs” and “winters.” Each wave begins with technical breakthroughs and ends when the limits of data and human supervision become clear. Today, as companies race to scale model training, the same tension is resurfacing. The new frontier isn’t algorithmic at all—it’s human. Who controls and understands the labor behind the data may determine whether this era of AI sustains its spring or slides back into another winter.

These days, it seems like everyone is entering the data annotation space. Earlier this year, Meta partially acquired Scale AI in an attempt to internalize dataset production for Large Language Model training, widely interpreted as a move to control the Data Annotation Gold Rush. Google soon followed by reducing its reliance on Scale as a vendor due to concerns about net neutrality and data provenance. Amazon’s MTurk is currently undergoing a consolidation into a new service called “SageMaker Ground Truth,” a gig platform focused specifically on human-in-the-loop and automation datasets. Just last week, xAI laid off 500 data annotation generalists from Grok training, with a preference for specialized AI tutor teams. Most recently, Uber entered the field with “Scaled Solutions,” a new division devoted to annotation after a pilot rollout in India in 2024 that is now slotted to roll out in the US.

Each of these developments reflects a broader pattern shift: the “gigification” of human expertise. In today’s AI economy, annotation is treated as endlessly scalable yet ultimately disposable. By framing it as piecework for drivers or delivery workers to pick up between shifts, companies trivialize a task that demands precision, cultural context, and sustained judgment. In chasing efficiency, the industry risks undermining the very human intelligence its models depend on.

The problem with framing annotation as gig work is that the task itself is rarely simple. Annotation is cognitively demanding, culturally specific, and dependent on rigorous quality-control pipelines, processes that take even the most mature firms years to refine. As annotation expands into domains such as computer vision, robotics, and 3D/4D modeling, the expertise required to perform this work increases dramatically. Training models to understand movement, depth, and spatial relationships is far more complex than labeling a static image. In many cases, annotators hold advanced degrees, possess specialized software skills, and can command rates exceeding $60 per hour (in some cases much more).

Take three-dimensional annotation, for example. Unlike its 2D counterpart, 3D annotation captures depth, distance, and volume, key inputs for vehicles and robotics systems such as warehouse automation like Symbiotic’s Warehouse Robot, Tesla’s Robotaxi, or May Mobility autonomous shuttles for last-mile transit issues. Integrating 4D perception, which tracks object movement over time, adds yet another level of complexity. These models also rely on multimodal data streams — weather conditions, map APIs, sensor inputs — each requiring precise synchronization and interpretation during the fine-tuning process.

Annotation for these systems takes place in specialized tools like CVAT, Label Studio, and Dataloop, solely focused on 3D and 4D annotation. Developing and maintaining such platforms requires significant engineering effort and capital investment. Despite projected market growth, estimated to reach $6.2 billion by 2030 with a CAGR of 22.2%, annotation tooling remains rudimentary, especially for higher-dimensional data.

The bottleneck lies in quality assurance. Sustaining continuous data pipelines demands robust validation, feedback, and reputation systems. Some companies have begun experimenting with blockchain-based reputation tracking to secure reliability and provenance across distributed workforces. Yet even advanced tools struggle with frame-loading performance issues, limited automation, and inconsistent user experience.

If we look at the trajectory of the field, low-complexity tasks such as image tagging, text ranking, and simple bounding boxes are quickly becoming relics of the early era of AI. The next frontier of model-fine-tuning depends not on scale alone, but cultivating skilled, fairly compensated human contributors who understand what models cannot yet see.

The last AI winter came from overestimating algorithms, the next will come from underestimating people. The race to scale data labeling risks forgetting that human comprehension is what makes intelligence possible. The difference between another AI spring and a new AI winter will hinge on whether we treat annotators as gig workers or as partners in intelligence.

The views expressed herein are my own and do not necessarily reflect those of my employer.

In Venezuela, María spends her early mornings labeling short video clips of illicit content so that an AI model can learn what not to show you online. Each task could pay less than one cent. She has no idea which model her work trains, only that her accuracy determines whether she will be paid at all.

AI may be the world’s newest export, but the labor behind it looks remarkably old: repetitive, underpaid, and outsourced. Before a model can reason, someone must teach it what to say and how to say it. That “someone” is rarely a well-paid researcher in Silicon Valley. More often, it’s a gig worker in the Majority World, or areas outside the US, Canada, and England, tasked with labeling repetitive content for cents per task. Many work invisibly in so-called “data farms,” crowded digital sweatshops where thousands perform the same task: they’re all training AI to mimic human reasoning.

When companies outsource annotation work to regions with weak labor protections, cost-saving mechanisms come at a price. In fact, outsourcing to third-party and regional vendors is a well-known method of risk mitigation utilized for training frontier models. Instead of adhering to gig work regulation in the country of origin, companies shift the legal burden to the local, regional vendors themselves. Regional annotation vendors typically have fewer or more relaxed quality assurance mechanisms. Frequently, the industry sees poorly conducted annotation emerging from these vendors due to a variety of issues: poor working conditions, lack of contextual country- and language-specific clues, or over-prescription to western cultural norms due to leading or minimal instructions.

The global data annotation workforce is highly concentrated in East Africa, South Asia, and Latin America. According to a 2023 World Bank Report, online gig work comprises up to 435 million workers globally, or 12% of the global workforce. This distribution at first glance reduces costs for AI firms, but also undermines data quality and ethical standards. Annotators often work without proper training, contracts, or vetting. The unevenness across geopolitical lines magnifies bias: datasets reflect narrow cultural perspectives, and those biases migrate into the systems that shape daily life, from drafting CVs to recommending job candidates.

Industry magnates argue that higher-quality data can be secured through better management and training without overhauling the global labor model. Centralized platforms, they suggest, are efficient at scaling data collection and ensuring some degree of quality assurance. But efficiency is not the same as equity, or safety. As long as incentives favor speed and low cost, annotation practices will prioritize volume over fairness or representation.

A better approach lies in building decentralized data annotation ecosystems. Instead of funneling tasks through centralized platforms, communities can participate directly in the creation and governance of datasets, with knowledge of what kind of models their work trains. Mechanisms for consent, local oversight, and equitable pay can embed diverse cultural and linguistic perspectives at the point of data creation, countering the dominance of narrow, Western viewpoints, centralized in the English language.

At the regional level, the European Union has moved furthest toward binding rules with its 2024 Platform Work Directive, which creates a presumption of employment for many platform workers and requires algorithmic transparency in task allocation. No equivalent regulatory framework exists in Latin America, where fragmented labor codes often exclude digital freelancers. The Comunidad Andina (CAN) and MERCOSUR trade blocs have yet to establish common standards on digital platform labor, even when partnering with the EU on transatlantic trade blocs, leaving gig workers exposed to the dynamics of global labor arbitrage.

International regulatory bodies are likewise poised to play a stronger role. The International Labour Organization (ILO) has recognized these risks through its AI for Fair Work campaign, which spotlights an action research methodology “designed to hold companies within the AI production network accountable.” Even further, as recent as June 5th, 2025, the ILO convened to draft a binding agreement on what is considered “fair work” in the platform economy. With robust cross-border standards, the international community can begin to shift the system away from gray area, low-oversight, low-pay models and toward practices that are safer for workers and safer for society.

Making the labor that sustains artificial intelligence visible requires more than acknowledgment. Regulation is one step, but reform must also reach the data itself. Rather than relying on opaque, centralized pipelines, data ecosystems can be designed for co-governance, democratizing how annotation work is valued and rewarded. Transparent allocation, shared oversight, and local quality checks can embed equity at the point of data creation. Recognizing this labor, and building the infrastructure to value it, is essential if AI is to reflect the world it claims to model.

The views expressed herein are my own and do not necessarily reflect those of my employer.