Remember the Will Smith spaghetti video?

No one would have believed it could advance from a scrap patch of random pixels and a mosaic of disarrayed frames to a realistic video with rich texture and smooth actions in a few years – the quality jumped drastically from March 2023, when Generative Artificial Intelligence (AI) for video production first appeared in front of the public, to early 2026.

While ordinary users were amazed by the quality of a short clip generated by a simple prompt, workers in entertainment industries were undoubtedly nervous. Is there a future where job replacement comes after them?

In the US and EU, the impact of AI on show business is barely felt – popular drama series booked, blockbusters shipped; there are few headlines about AI replacing actors or popular dramas.

But not in China – where technology spreads fast, where some of the world’s best AI models are built, where the audience is enormous, and where, most importantly, labour rights are barely enforced.

Due to the economic and political environment, there aren’t many incentives for the government to protect labour from the rapidly changing business environment. A perfect storm is swirling against the entertainment industry in China.

The diffusion of smartphones has drastically reshaped how people, especially the elderly, in China consume video products. Just like the way television drew people from the cinema to their couch, smartphones shifted their attention from the 32″ TV to their Xiaomi. Instead of 90-minute movies or 60-minute soap operas, in recent years, the rise of vertical microdramas has dominated the Chinese entertainment industry.

The rise of microdramas dates back to 2022. Certain unique features set them up for success in the post-pandemic world: short (single episodes range from 60 seconds to 15 minutes), appealing for quick emotional responses, and melodramatic. In 2025, China had more than 100,000 microdrama-related businesses; the market size had risen to 63.4 billion yuan ($9.4 billion) and was projected to be a 100-billion-yuan sector. The income generated by microdrama surpassed that of movies in 2024, and the gap is widening.

Screens unlocked, apps opened, viewers were hooked by the first 5 seconds of the clip, and finished 10 episodes in 15 minutes. They thought about whether to pay for the rest of the series. This is the daily routine of up to 70% of netizens in China.

The microdrama industry experienced rapid expansion, and actors enjoyed explosive growth in fame and income. But that seems no longer to be the case, at least not for humans. The market expands continuously – but the production and the performers are increasingly replaced by AI.

The dynamics shifted drastically in 2026, after ByteDance rolled out the powerful AI video producer Seedance 2.0. Of the 128,000 series of microdramas released in the first quarter of 2026, 95% were fully generated by AI. Industry insiders revealed that the cost of production using AI is only 20 to 30% of traditional filming. It has opened up an enormous amount of “creative freedom” so that production can be more visually impactful and diversified. Playwrights can now write stories about Chinese warriors in combat, an apocalypse or an idyllic countryside river without thinking about the cost of production.

The performers are the unfortunate ones. In an interview with the BBC, a leading microdrama actress was out of work almost instantly when the AI replacement came. She was shocked, as she had been fully booked for the past two years; even after she accepted cutting her salary in half, opportunities were lacking. In the span of months, hundreds of companies went dark; tens of thousands of industry workers lost their jobs.

“The microdrama attracts people by spiking emotions; emotions are driven by melodramatic themes and frames. How can human performance possibly be more dramatic than AI?” The actress said desperately.

But the replacement is more than the opportunity to appear in front of the camera. It’s also another industry where actors make handsome money – advertising.

“Endorsements from AI actors are more expensive than humans!” A sensational news headline read. “The Laid-Off Girl” (被裁掉的女孩), the viral AI microdrama about a young model pursuing her dream in a fancy city, attracted close to 300 million views across all platforms. The leading AI actress, Fang Taozi (方桃子), has a Douyin (Chinese version of TikTok) account with more than 400k followers and interacts with fans regularly. Reportedly, a 60-second advertising clip on Fang Taozi’s account charges up to 25,800 yuan (around $3,800). The price tag is comparable to influencers with millions of followers.

I was curious about the quality of the phenomenal “The Laid-Off Girl” – it must hit something right to be so sensational. I watched the first 60 seconds of it, and I gave up. It did not attract me. One of the first scenes is a group meeting in a restaurant. The background is so inconsistent – every time the camera takes a different angle, the characters’ seating order is scrambled, and even the shape of their seats changes. The inconsistencies were so annoying that I couldn’t even finish the scene.

Then I tried to watch some of the social media posts from Fang Taozi – the same inconsistency appeared. The cat in the background of a wide shot looks completely different from the one in a medium shot. The fully stacked rack emptied when the camera zoomed in.

The advantage of a real set is that at least most of the background is consistent, and I appreciate that. I don’t think any of these AI microdramas would attract me if the quality stays like this.

But I can also understand why these productions attracted hundreds of millions of viewers: it’s the dopamine spike – storytelling so fast it barely gives you an establishing shot. A psychiatrist offered an explanation for this phenomenon:

“The rapid change of scenes stimulates the brain’s reward system, triggering rapid dopamine release and producing an immediate sense of pleasure.”

Most microdramas revolve around fixed formulas: the underdog’s rise, justice served for bad people, etc. They may seem old-school or even obsolete, but they satisfy expectations unmet in viewers’ real lives.

Then there is the psychological mechanism of “suspension of disbelief” – we may know the content is fictional and still be moved to tears by the story.

“In the same way, as long as a story resonates, people will temporarily overlook the unnaturalness of AI-generated images and focus instead on the emotional experience.”

Culture has historically been transmitted through humans.

A scenario in the AI safety field that receives wide attention is “gradual disempowerment.” The proliferation of AI microdrama is arguably the closest thing we have to it. Even though producers can engineer detailed prompts, fine-tune the software to meet their demands, and eventually direct and edit the videos by themselves, a large portion of the visual product is unavoidably determined by the algorithm and mathematical calculations behind the machines.

In the face of AI’s rise, tens of thousands of workers were displaced in months, with no independent union, no bargaining power, and a government that doesn’t act enough to alleviate the situation; no one needs them for anything, so nobody has to answer to them. The AI-dominated microdrama industry portrays an environment where machines replace humans – without a system of democratic oversight and checks, when human resources are disposable. I find it hard to imagine what that anxiety must feel like.

That shows that a democratic system that centres people is important in the age of the AI revolution.

Will this test come to the “West”? We have not seen the equivalent of the rise of vertical video and the addiction to AI-generated content (though the US is the largest microdrama market outside China – roughly $819 million in 2024), and yet we are unsure whether it’s just a matter of time, or whether the audiences are fundamentally different such that the viewers in the US or the UK are more resistant to the change.

Part of it is, of course, that the speed of industry shift is slower here. It’s hard to imagine a business, or an entire industry, being transformed in a matter of months – tens of thousands of workers being laid off without triggering substantial backlash in society and from unions. The rights of creative industries workers are generally more protected, and more checks and balances are in place to ensure good corporate practices.

The attitude towards AI also marks a difference. People in China have a more positive attitude – or you can claim that it’s driven by survival instinct or anxiety to embrace the technology as an irresistible trend – therefore the diffusion of AI-generated content is met with less resistance.

I believe most want to live in societies where paintings, songs, and dramas are products of, and by, humans – that’s the only way to escape a gradually homogeneous world and avoid flattening our diversity and creativity. The evolution of the vertical microdrama industry in China can serve as a great case study to show how a democratic society can increase our friction against a trend of taking away people’s agency and independence.

When we talk about “openness” in tech, we often associate it with an internet connecting the rest of the world, a diversity of thoughts, an attitude of exploration and curiosity. An open-weight model does not automatically embody these values.

Over the past eighteen months, Chinese labs have seemingly become the standard-bearers of open release, while American ones have retreated behind APIs and safety justifications. Some even describe Chinese open weights as a democratising force, American proprietary models as the artefact of a rent-seeking capitalism.

This could not be further from reality.

The start of censorship

In the late 2000s, I travelled back to China often. I learned that relatives in the mainland started using the messenger application QQ and Xiaonei/RenRen, while we in Hong Kong used MSN Messenger and Facebook. I noticed that the internet ecosystems in the two places had been gradually diverging.

There was a time when the Chinese internet was not as insulated. You could still access Facebook, Google, and Twitter while Chinese-incubated competitors like QQ, Renren were on the rise; it wasn’t until Beijing strengthened political and speech control in the online space that the mobility stalled.

It ended up with Facebook and Twitter being blocked in 2009 after the Ürümqi unrests; Google retreated in 2010 and eventually gave up the whole market because of speech control concerns. The parallel replicas of these American tech giants thrived inside the Chinese firewall because they submit themselves to overwhelming political control and obey every order to silence and erase dissent.

The very first priority of the Chinese regime is political stability; in their own terms, it is the continuation of the incumbent Chinese Communist Party. If a service is seen as not wholeheartedly and unreservedly assisting this goal, it faces the risk of extinction.

Fanfou(飯否), the first to be called the “Chinese version of Twitter”, the popular microblog site started in 2007, went dark for roughly sixteen months because it was allegedly not fully complying with the government’s online speech control in 2009 after the Ürümqi unrest. Several of China’s first-generation microblogging services were also shuttered after the incident.

Fanfou started way before Weibo, gained enormous success with the rise of social media, but eventually died down, opening space for alternative services that were more willing to comply with the government’s agenda.

Why open-weights, and the dangers of Chinese models

Fast forward to 2026, Chinese open-weight AI models have been progressing successfully around the world. Alibaba’s Qwen family passed 700 million cumulative downloads by January 2026 and has since crossed a billion, overtaking Llama on Hugging Face; Chinese open-weight models went from roughly 1% of global usage share in late 2024 to something closer to 30% within months. Moonshot released Kimi K3 on 17 July 2026 at around 2.8 trillion parameters, pictured as the largest and most powerful open-source model in the world.

The push for open-source communities in China started in 2020, when OpenAtom Foundation was founded, and the government pushed Gitee as a domestic alternative to GitHub; since then, the Chinese government has been implementing policies and allocating resources to support and endorse an open-source AI ecosystem.

Three explanations are usually offered for the open-weight decisions:

• On compute. Even though reports reveal chip export evasions from Chinese companies, they have far lower compute than American counterparts. That deficit led to choices of lower inference costs – open release externalises serving costs to third-party hosts worldwide.

• On market access. Open-weight models are generally more accessible than proprietary models – if you have a decent computer, you can run models locally. These models also generally offer lower per-token prices. In markets where Chinese models are not the most advanced, they flood the market with accessibility and price.

• As a state policy. Building a powerful and open system helps China shape itself as a “benefit sharer” on the global stage, and sidestep vulnerability chokepoints such as Huawei losing its Android licence in 2019.

Behind the ideological push as a “benefit sharer” and embodiment of openness, the danger of Chinese open-source models has been well researched and documented.

CrowdStrike reported in November 2025 that DeepSeek-R1 produced insecure code around 19% of the time at baseline – and that adding politically sensitive context to a prompt, such as specifying the software was for use in Tibet, raised the likelihood of severe vulnerabilities by up to 50%, on coding tasks with no political content at all. In 35% of implementations, the model used weak password hashing or none. The researchers also documented what they called an intrinsic kill switch: the model reasoning through a complete technical solution and then declining to emit it, refusing outright in 45% of cases involving Falun Gong.

The broader picture is consistent. A 2026 PNAS Nexus study tested nine models against 145 politically sensitive questions: Chinese models refused between 30% and 60% of the time, against 0 - 3% for Western models. When they did answer, they disputed the premise, omitted material information, or fabricated.

The R1dacted study found that the censorship is applied at the model level, present in locally run weights, not merely at the API. Another study found sensitive material appearing in internal reasoning traces and disappearing from final output: the model knows and decides to suppress. China Media Project obtained Qwen3’s internal alignment instructions, which direct the model on China-related queries to “focus on China’s achievements” and “avoid any negative or critical statements,” while instructing it to be “neutral and objective” about the United States, Kenya or Belgium.

A CEIAS study published in July 2026 on the political inclination of Chinese models shows that Kimi substituted CCP messaging in roughly one of every three answers on politically sensitive topics; DeepSeek in one of four. The trigger was the topic keyword – a question about press freedom in Estonia or minority rights in Mongolia elicited the same CCP framing as one about China. Different prompt language performs differently: significant CCP insertion rose from 24% in English to 59% in Mandarin across the ten most sensitive topics.

The most censored version of these models is the one served to Chinese speakers. A technology marketed globally as democratising delivers its most constrained output to the population with the fewest alternatives.

Policy-backed model behaviour

None of this is attributing bad faith to Chinese developers. Many are, I have no doubt, building open tools because they believe in a freer and more borderless internet. But that just does not manifest under the CCP’s tightened political control. The biases models exhibit have policy backing.

The Interim Measures for the Management of Generative AI Services, effective 15 August 2023, require adherence to “core socialist values” and prohibit content inciting subversion of state power, endangering national security, harming the national image, inciting separatism, or undermining national unity and social stability.

To deploy the models publicly, a provider must pass a 2,000-question test set with at least 90% acceptable responses while wrongly refusing no more than 5% of benign prompts, and must maintain a keyword library of at least ten thousand entries. As of November 2025, 611 generative AI services and 306 applications had completed mandatory registration.

A parallel legal architecture governing data is also in place. Article 7 of the National Intelligence Law requires all organisations and citizens to support and assist state intelligence work; Article 28 of the Cybersecurity Law requires network operators to provide technical assistance to public and state security organs; Article 35 of the Data Security Law requires cooperation with security-organ data collection.

Accessing Chinese models through their servers has enormous data risks; even if you operate it locally, the biases permeate each and every answer, and they shape the epistemic environment you and your companies’ users are in.

What’s next?

This article is not an advertisement for US frontier models, or to justify the race dynamics. In fact, I’d love US tech products to be more guardrailed and secured; but the fact that the policies and the product are not up to certain AI safety standards is not a convenient excuse to get behind Chinese models with so many hardwired biases and national security risks.

It could be a cost calculation that some businesses decided to use non-US models; but the framing of China as a global ‘benefit sharer’ should stop. The diffusion of models filled with authoritarian ideologies is net-negative to freedom and democracy around the world.

The wide adoption of Chinese models should be treated as an emerging risk. Political-alignment and security testing should be mandated to mitigate those risks. And in high-stakes settings, such as public services, critical infrastructure, and products used for education, models that fail those tests should be excluded outright. Consumers should also be told when a product builds on models that failed those tests.

It would be foolish to pretend that the only thing worth asking of these systems is how good they are at coding and maths. The cost of building on models that reproduce an authoritarian state’s framing by design is high. Intervention is what incentivises corporations to recalibrate – and once that cost is priced in, these models stop looking cheap.

^{Photograph: Jenny Kane/AP}

1. Historian Timothy Snyder (one of my favourite authors) writes extensively on democracy, authoritarianism and fascism. He describes the politics of inevitability in his book “Road to Unfreedom” as:

“The future is just more of the present that the laws of progress are known, that there are no alternatives, and therefore nothing really to be done.”

2. The politics of inevitability guided Americans through a capitalist vision of history, from nature to market to democracy; the European version was nation, peace, and integration; the Soviet Union version took a different yet disastrous turn, from nature to technology, then social change, and utopia through revolution. The collapse of the Soviet Union gave birth to the Russia that launched one of the most atrocious wars in our times and continues to suppress free will within its borders. The American, European and Soviet Union politics of inevitability have promised progress and prosperity to the people, in exchange for their approval and tacit contract.

3. The problem is the deal is too good to be true. The magnitude of the discourse implies something more than national flourishing – it mandates that the certainty of progress will eventually triumph over the odds, thus, all hurdles and objections that stand in between the destined progression are either futile, lame or even hateful.

4. In this version of determinism, as Snyder argues, we lose sight of history. We lose track of learning from other events that ended up in failure; the ability to reflect, revisit or even resist has been amputated.

5. The politics of inevitability, in the 2020s, has come to another form – the promise in the age of abundance from Artificial Intelligence (AI) progress. The notion of AI, as accelerated and explosive technological progress, as a form of assistance and eventually substitution for human intelligence, has decades of history, tracing all the way back to the 1950s. One of AI’s founding scientists Herbert Simon in 1965 said machines would within twenty years do any work a man can do; mathematician and cryptologist Irving John Good originated the concept of intelligence explosion in 1965; esteemed AI scientist Marvin Minsky told Life magazine in 1970 that human-level general intelligence was three to eight years out.

6. The prediction of AI revolution is as old as the concept. But the idea that human-level intelligence AI is materially feasible with the injection of compute and capital is fairly recent. The introduction of ChatGPT in November 2022 was the public’s first contact with how revolutionary this technology can be.

7. Since then, everyone has been asking where AI leads us, and leading figures have invited us to peek into the future they envision. One of the “end stages” of the AI revolution can be understood through a recent piece written by co-founder and CEO of Google DeepMind Demis Hassabis:

“The magnitude of this technology’s impact will be unprecedented, perhaps 10x of the Industrial Revolution at 10x the speed. It will help us solve some of the biggest problems society faces, from accelerating drug discovery to developing new clean energy sources to creating novel advanced materials. We could even reach a point where resources are no longer the limiting factor for human progress, leading to an amazing new era of abundance.”

8. Demis has expressed clear reservations about the future of AI, the need to steward the technology and more regulatory control. But the foundational tone of inevitability is also hard to neglect – that it permeates the AI executives’ narratives in the past years and continues to be reinforced by technological accelerationists.

9. The public has absorbed the inevitability of AI, but not the executives’ optimism. Pew put the same questions to AI experts and to the American public: 47% of experts were more excited than concerned about AI, against 11% of the public; 76% of experts expected to benefit personally, against 24% of the public. Optimism about abundance is concentrated among the people building it.

10. The politics of inevitability in AI goes: scaling AI, achieving Artificial General/Super Intelligence, and the world of abundance ensues. The growth of AI intelligence, benefits generated, and the age of abundance come hand in hand – ostensibly, if we solve the mysterious task of AGI, we will reach “the end of history.”

10. With the incredible success of ChatGPT’s launch (achieved 100m users in two months after launch) and the scaling growth of AI’s capacities, the idea of AI inevitability has gradually been ingrained in the commercial world, politics, and our daily lives. How do I use AI to help my work and life? When will we achieve superhuman intelligence? How do we govern intelligence that’s superior to us? Changes are seemingly irresistible, and the dreamland of powerful AI and the new era of abundance is destined to come. The questions of whether these futures will arrive, and what has been sacrificed in the process, are just unimportant remarks in this train of inevitable future.

11. Capital gladly participates in the politics of inevitability. The infamous over $300 billion Oracle-Nvidia-OpenAI circular investment loops illustrated the eagerness to whisper the tale – these investments are entirely built on lofty expectations of AI. It boosted AI hype and made headlines everywhere. The explosive growth of AI-related infrastructure industries and the construction of data centres reflect the same optimism.

12. The problem in the politics of inevitability is that it omits history. It neglects the different alternatives the progress could have marched towards, and the failure of other technologies demonstrated similar promising features. When the force of progress, in an analogy of almost a saviour of humanity, crusades so forcefully, it turns out to be both authoritarian and suffocating: legitimate opposition would, in the narrative of inevitability, become ash of history and be justifiably dismissed.

13. Most people rejecting the prophecy are not anti-technology. Many who appeal for a halt or slowing down the progress are core contributors to the technologies. Most resist the projection of a rosy, abundance future not because they hate AI or prefer a less thriving society, but because they see alternative visions of what the current AI progress could become. The AI revolution may result in individual loss of agency, concentration of power, or annihilation of the human race.

14. Some people are more impacted by the early stage of the AI revolution, such as entry-level white-collar workers and graduates. Do those who shaped the whole narrative and industry care to explain why these people are sacrifices of the great journey leading to abundance?

15. The public backlash to the politics of inevitability in AI has been enormous: most polls in the US show people’s negative sentiment, more protests on the construction of data centres, and waves of boycotting the use of AI in younger generations. It has propelled a certain change of tone. Executives had been bragging about the mass job replacement of AI to lure investments; yet, CEO of OpenAI Sam Altman eventually had to backtrack his prophecy1. He wrote this a few weeks ago:

16. Altman was one of the loudest to warn against “AI job apocalypse” – in the past two years he’s been cautioning society that AI will replace most white-collar employment within a decade. I felt the change of tone quite abrupt. AI executives’ predictions that are in favour of them raising capital are always doubtful to me – if a forecast as significant and supposedly scientific (though I know it’s difficult and economists agree that it’s difficult) as job replacement can flip so drastically in a short period of time, it shows how narratives have been morphed into prophecy to manipulate.

17. Yet, even if we dislike the future AI executives depict, the positive impact of the technology should easily be felt. I use AI to retrieve information or as a thought companion. AlphaFold was the most salient example that AI benefits humanity: it transformed structural biology and compressed a step that used to take years.

18. That’s why solely assessing the direct impact of AI on the public now does not explain the anguish that runs much deeper. When the speaker of the University of Central Florida claimed that AI is the next industrial revolution, the attending students reacted with a long BOO; former Google CEO Eric Schmidt received a similar reception when he addressed the importance of AI during his commencement speech at the University of Arizona.

19. The emotional responses derive from the resistance to the politics of inevitability in AI. It’s not only about AI safety or ethical risks, but a rejection of being seen as noise, fed into an irreversible narrative, and under the enormous weight to comply with “the trend” without public discussion or seeking their consent.

20. Undoubtedly, AI has the potential to bring abundance to humanity. But that’s not going to happen without a struggle and clear-minded planning – figure out the diffusion of benefits and wealth, the mechanism to address loss of agency and meaning, a social system that specifically supports those most impacted. It’s difficult to see politics that plans for the future, but that’s the prerequisite of success.

21. This is measurable. In June 2026, Athena Insights asked a nationally representative sample of American adults to choose between two statements. Around 70% said AI is coming into their lives whether they want it or not. 15% said people like them can shape how much AI is part of their lives. Pew separately found 61% want more control over how AI is used in their lives, against 17% content with the control they have.

22. That the narrative matters, and not only the technology, also has empirical support. Social Change Lab ran a randomised trial in the UK to test which AI harms move people to act. Anger predicted willingness to act; anxiety, fear and a sense of powerlessness did not. A politics of inevitability produces the second set of feelings rather than the first. It tells people the outcome is settled, and people who believe an outcome is settled do not organise against it.

23. I want to see the promise of AI abundance succeed – that’s why the discussion of possibilities and democratising the decision-making is so important. People always want to be listened to and respected. Demonstrating that the companies building the technology are indeed treading carefully, and that the voices of opposition are legitimate and worth engagement, is the key to breaking the politics of inevitability.

Thanks to Colin Bortner, who sharpened several of these arguments.

It turns out to be an AI pet companion. It was produced by a company that was eventually acquired by Casio. The creature, which turns out to be a robot, can react to the way you touch with simulated responses from animals. I suspect that it’s learnt from hamster – the reactions don’t look like any I encountered with my cat, at least.

It was really cute. It shared one very important feature as zodiac – it helps you open up a conversation. I spent hours chatting with my friend, and every 15 minutes, another guest would come by and display a sense of shock. I suspected that if I got myself one, my conversation conversion rate would be exponentially increased.

Would I like the hamster-like robot? I actually googled it and thought about buying it. It looks like a great companion – it reacts to you but not too much. Lasts for a long time after one charge. You don’t have to feed them, and probably if you decided it was no fun, you could toss it without hurting anyone.

It’s convenient, isn’t it? But, is that good?

It’s always tempting to replace something difficult with an easy alternative. Hussleless. Or even, cheat to achieve your goal.

I did that in gaming. Not those competitive games where other players’ gaming experiences are destroyed when you cheat – I did that in those, the whole universe is you and the game. To put it in simpler terms, a solo game.

I like playing strategy games. One category of them is idle games – games that you need to make decisions and wait for them to be completed, thus you receive rewards to repeat these actions. I played Melvor Idle a while ago. It was often ranked among the best idle games. Very easy to play, low demand for hardware, well-balanced gameplay, and enough depth that people could invest hundreds of hours and still find fun in it.

I spent tens of hours playing it once I downloaded it. In game, I tried to figure out the best way to utilise resources and build power to beat the next monsters. I quickly got stuck in a bottleneck – probably designed that way to intentionally keep players at a certain level, and force them to invest time to figure out how to improve before advancing.

But I was quite impatient. I wanted to use something similar to a magic remote control in the movie “Click”, where you can fast-forward parts you don’t like in life. I did that without magic power – I found out a way to install an in-game time machine. In these idle games, you must wait before a crop is produced or a weapon is forged. But with these in-game editors, you can immediately get what you want.

To beat the next monster without wasting more time, I used the time machine without hesitation.

Immediately, the all-mighty monster that I struggled to kill for a while became so weak in front of my newly acquired even-mightier weapons. I had such an adrenaline rush and thrill that I felt I had unlocked a fast pass of the game, and I got to enjoy the gist of it without the waiting and repetitive actions. I didn’t have to do the planting, roaming, forging, killing weaker and more boring monsters to get resources, mining, and waiting around the clock.

I thought I would enjoy this god-like feeling. I thought it would be fun.

No it wasn’t. And I should know it wasn’t – it had happened so many times before. All the spoilers that ruin the surprises of the plot twist, the run-through videos that gave me the answer to the puzzle but I lost the fun of cracking it, and the cheat code making me so powerful that I won’t encounter any enemies I could not overpower, yet it took away all my sense of satisfaction, and the sweet feeling of “my hard work paid off.”

A wise person once said that it’s the journey that matters. Both the joy and the pain. It is the blood and sweat that make the reward tasty. It is the hard work that makes the result cherished.

I thought about it when I was deciding whether to spend hundreds on the hamster-like robot I found cute and carefree. I literally had it in Casio’s shopping cart, but I was so hesitant about whether I should buy it. It’s not a big sum of money – though it’s not a small purchase. And I can clearly see that I will have a bit of fun with it. But, do I treasure it as much as a real pet companion, when I don’t even need to take any responsibility, or spend time to understand them? I don’t even need to commit anything – my time, my effort, my comfort in exchange for the companion feeling happy and satisfied.

When you are in the cheat mode in a companionship, what’s the fun in it? What are you truly looking for?

What can you gain in a relationship where you don’t have to compromise or make adjustments, where you don’t have to juggle between priorities, and where you don’t have to give any effort to have the expected “joy” in return?

Nah, I don’t feel like I need it. And when I really thought deeply about whether to buy it, I thought about all the shovelling of the extremely odorous wastes that my cats produced, fairly regularly, that I disdained and disliked. But I realised that, because I loved and adored them, that’s what I do for them, and why I was proud of myself and felt a sense of satisfaction. It’s these little things that define a relationship, that I don’t think AI-robots can produce, as their fundamental business logic is to optimise for what the tech firms believe humans are pursuing: less agency, less independence, less effort, less commitment, and more dopamine rush.

The companion product, as much as we thought they “know” you, at the end of the day, it’s about hollowing out your attention and the ability to live a normal life without a screen. Hope to be understood and accompanied are real needs and could address some of the most deeply-rooted problems in our society. But is it the way out?

These firms and products must have gotten something right to become so popular and in such high demand. But also something very wrong.

The only scenario in which I would buy the Casio AI-robot is if I am 100% disregarding it as a companion – it’s just another toy, as cute as a jellycat. Having that critical distance will keep me sane, as much as if I were to cheat code whenever I decided I’m going to ditch the game. It won’t mess with how I think about what actually matters.

China’s military parade in early September was staged not only to demonstrate strength, but also to showcase its alliances. Seated beside President Xi Jinping were Russian President Vladimir Putin, North Korean leader Kim Jong-un, and senior Iranian officials. The display of solidarity among these governments was widely noted, with some observers describing them as the “Axis of Upheaval.” The phrase evoked memories of the alliances of the Second World War and underscored the extent to which authoritarian governments are increasingly comfortable signalling their alignment on the world stage.

What captured international attention, despite the weaponry and the guest list, was an unscripted exchange. During a live broadcast, as Xi and Putin walked together, microphones picked up a conversation about organ transplantation and immortality. Putin suggested that organs from younger people could prolong vitality, while Xi referred to claims that human lifespans might stretch to 150 years.

Had this been a conversation between random passersby, it might have been dismissed as fantasy talk. But given the positions of the speakers—and longstanding allegations of forced organ harvesting in the PRC—the exchange drew widespread scrutiny. The video circulated quickly online, provoking debate before Chinese authorities requested media platforms to retract it.

Despite Beijing’s request, the clip spread rapidly across international social media platforms, which were beyond their reach. Inside China, officials acted swiftly. Discussions in Chinese social platforms about “150 years old” and related topics once spiked but were later censored.

A user on “X” (formerly as Twitter) shared his experience on WeChat. He sent a message to colleagues joking about Xi’s reference to “immortality(長生不老).” The message appeared on his desktop version of the app but disappeared from his phone. His colleagues were unable to read them. Another user tested the same phrase and confirmed that the message was automatically censored.

It’s a lively example of how censorship in China covers from public to private domains. Such a robust digital authoritarian infrastructure has been through numerous changes and is now enabled by powerful AI systems.

The shift to AI moderation

In the early days of China’s internet, content moderation was carried out primarily by human reviewers employed by government agencies and technology firms. They monitored posts, flagged sensitive material, and deleted content deemed politically problematic.

As the number of internet users grew (reaching more than 1.05 billion by 2023), the sheer volume of material became impossible to manage through manual review alone. To meet this challenge, major platforms such as WeChat, Weibo, and Douyin adopted AI-driven censorship tools.

These systems employ natural language processing, machine learning, sentiment analysis, and image-recognition technologies to scan vast amounts of text, photos, and videos in real-time. They can automatically remove or block sensitive content, shift censorship from reactive deletion to proactive suppression, and reduce the time needed to enforce government directives.

The Authoritarian Data Problem

Despite their capabilities, AI censorship systems face inherent limitations. Scholars describe one of the limitations as the “Authoritarian Data Problem.”

The more tightly a state represses speech, the less relevant data it generates for training its algorithms, resulting in biases and the incapacity to detect some dissenting content.

Citizens, aware of the risks, adapt strategically: they self-censor and rely on coded language. This behaviour reduces the amount of politically sensitive material available for AI training. As a result, the algorithm’s ability to distinguish between acceptable and censurable content is weakened. The challenge grows as users devise new homophones, euphemisms, puns, and cultural references designed to evade detection.

The dynamic mirrors biases seen in other AI applications, such as facial recognition, where unbalanced datasets lead to systematic inaccuracies. In censorship, repression itself produces the imbalance: the more dissent is suppressed and people refrain from expressing themselves, the fewer examples exist for machines to learn from.

To offset this shortfall, authoritarian governments can turn to international data. Posts from platforms like Twitter or Facebook, generated in less restricted environments, provide a rich source of politically sensitive discourse. Incorporating such data can improve censorship algorithms by exposing them to patterns absent in domestic platforms.

Yet this approach has limits. The language and style of political discussion inside China often diverge sharply from what appears on Western platforms. Coded speech used on Weibo or WeChat may not resemble posts on Twitter, making the imported data only partially useful. Academic studies simulating this technique also found that the impact of these patching data does not offset the constraint of the politically biased data either.

Outsourcing the generation of politically sensitive data to foreign workers also encounters difficulties. The coded languages and slang employed by Chinese netizens often require deep cultural familiarity to decode. Without this knowledge, identifying subversive content becomes difficult.

For these reasons, AI has not eliminated the need for human censors, as the AI system generates problems, including a prevalence of false positive cases. Instead, the two systems function together. The possible mechanism is that AI tools conduct initial scans, filtering or flagging potentially sensitive material, while human reviewers make final judgments, particularly on ambiguous cases.

The latest report by InterSecLab analyses leaked materials revealing how China exports its internet surveillance and censorship technologies to other countries. A key finding is the integration of AI at multiple layers of the censorship system. AI systems process vast volumes of network data, using pattern recognition, anomaly detection, and behavioural scoring to flag certain users or traffic patterns as “suspicious.” Human analysts then review these alerts, adjust the system’s sensitivity, and determine appropriate follow-up actions. This demonstrates both the scale of AI-driven automation in modern censorship regimes and the continued reliance on human oversight to calibrate and act upon AI-generated findings.

Optimism for activists

What remains unclear is whether the rise of AI has significantly reduced the overall costs of “stability maintenance.” The financial burden of “stability maintenance” (維穩) has long been a focus, and it has surpassed military spending in some years since 2009. If the expenses remain high, strains will grow as China’s economy slows. For the foreseeable future, human oversight will remain an integral part of the system, ensuring that the financial weight persists.

For Chinese dissidents, these discoveries offer a sliver of hope and guidance: the ingenuity of human expression can still outpace the seemingly superior machine. The creative acts of protest in the whitepaper demonstrations against the strict COVID-19 measures were a prime example that the government would act sluggishly in the face of unprecedented expressions of dissent.

AI cannot censor ideas it has never read, and this gap continues to encourage citizens to find new ways of speaking. For those operating on Western platforms, one strategy is to communicate as plainly as possible rather than imitating the coded homophones and wordplay common inside China. Doing so limits the regime’s ability to harvest useful training data, thereby weakening the effectiveness of its censorship algorithms over time.

If economic pressures in China intensify, sustaining the apparatus could become more difficult. During periods of political unrest, when dissent rises and censorship demands increase, the risk of system overload also grows. Should AI systems underperform due to the emergence of new creative protest slogans and human moderators being insufficient due to financial constraints, opportunities for the broader circulation of sensitive information and social movements may emerge.

The first was a video advert in a Taipei tube station. I can’t recall the product, only the uneasy feeling it left me with. Later, I realised why: a small disclaimer revealed it had been AI-generated.

The second instance was more striking: an AI-produced scene in a prime-time soap opera. In Taiwan, the “8 o’clock slot” (八點檔) is television’s golden hour. Families gather after dinner, and broadcasters fill the time with long-running dramas, often produced in a “shoot-as-you-go” model. This allows producers to adjust storylines based on ratings, but the rush often leads to uneven quality. Some series are known for flimsy plots and clumsy production values.

(An example of an “8 o’clock slot” meme: a nasal cannula incorrectly connected to a patient.)

Bittersweet Destiny (百味人生) is one of the ongoing “8 o’clock slot” produced by Sanlih E-Television (三立電視). The drama follows the legacy of Wu Xiu Yuan a restaurant founded by the celebrated Chef Wan Li. After his sudden death, the restaurant fractures, and decades later, his disciples and family are still wrestling with old grudges and entanglements. The series, which premiered on the 20th, has drawn notice for being one of Taiwan’s first dramas to integrate AI-generated video.

In the opening episode, a mother and her infant son fall into the sea and are separated. Several shots—underwater sequences of the characters, including the mother downing and her child flooding—were created using generative AI. A small caption on the right side read: “scenes generated by AI.”

AI-generated video has already fooled many on social media: fabricated street interviews, fake surveillance footage, and endless clips of animals bouncing on trampolines. It’s not that AI-generated videos are inherently bad, but many unwatermarked ones are mistakenly believed to be true, causing confusion and capitalising on it by influencers to earn viewing traffic.

But prime-time television is a different challenge. Viewers expect continuity and craft; a missing finger or skewed physics can shatter the illusion. In Bittersweet Destiny, the AI shots, actually still did not match the realism of live filming. The body under water appeared unnaturally flat, the lighting was stiff, the waves oddly still. Yet audiences seemed forgiving because they know soap operas at this hour often run on tight budgets. For producers, AI offered a clever workaround to depict an expensive, technically challenging scene while leaving the rest to the viewers’ imagination.

This may prove liberating for smaller markets. Generative AI still struggles with complexity—multiple characters moving at once, intricate choreography, or emotionally nuanced acting. Glitches abound. But for simpler, technically demanding shots, AI can provide a workable solution.

Interestingly, online praise for the show had less to do with technical achievement. It’s about the disclaimer—labelling the footage “AI-generated” resonated.

Taiwan has earned an unfortunate reputation as a hub for scams; some politicians have even lamented that “big scams have become the norm for ordinary people.” While Taiwan boasts a thriving high-tech industry, many regions remain rural, and many older citizens are less technologically confident. As AI advances, they become more vulnerable to fraudsters armed with increasingly convincing impersonations.

The “8 o’clock slots” dramas, however, reach this demographic. By exposing older viewers directly to AI-generated video—clearly labelled as such—these shows inadvertently serve as a kind of public education. Audiences begin to recognise the texture of AI fakery, to see how authentic and yet how artificial it can be. This awareness may help them against fraud.

How can we mitigate the risks of malicious actors exploiting these tools to target the vulnerable? Perhaps, paradoxically, interestingly, the answer begins with using generative AI in the most ordinary of places—prime-time melodramas your grandmother already watches.

But that also changed her information diet. One day, she enthusiastically told me about a mysterious combination of Chinese medicine. She was convinced that it could effectively fight off cancer, with efficacy much higher than the ordinary chemo treatment. When I asked about the source, she replied: WeChat.

She did not know who was posting this oblivious gibberish. But the fact that she read this information from a “well-known platform”, which sounded like a big TV station to her, led her to buy into it immediately. It was a quick revelation to me that people, especially those who were raised before the information age, have different levels of tech literacy depending on their education level and experience in interacting with technology. My mum was almost at the lowest category of tech literacy.

Our natural, almost instinctive ability to spot obvious nonsense content, phishing emails, scam websites, and other harmful or inaccurate content is not as common as we would imagine.

The learning curve is exponential – the more inhabited and immersed in tech growing up, the easier it is to adopt the change. Those who are left behind, or whose speed of learning is slower than the evolution of technological change, will find it growingly more difficult to conceptualise the world.

A couple of weeks ago, a pro-CCP social media page published a deepfake video of me and two other Hong Kong activists expressing our worries about the latest immigration change in the UK. It was supposed to portray us as being “abandoned” by the UK government, facing exportation, and crying for help. The goal of the video was to create a perception that anyone who stands up against the Chinese regime will end up living miserably – to a degree that these activists needed to get into a Zoom call and chant our frustrations.

(MediaVault copy of the original deepfake video.)

The content was false. The opinions expressed in the video were never our thoughts. The terms my lookalike in the video used do not resemble how I speak. But to the page’s target audience — those who want to spit on us and doomscrollers who struggle to discern authenticity and have a limited grasp of my speaking style — it could be a convincing video. They could have believed that I am indeed in desperate need of help and almost being kicked out of the UK.

This is how AI-empowered malign information warfare is in action. Academics have been concerned about the impact of generative AI on our democracy – the information landscape will be so polluted that legislators are unable to discern their constituents' preferences, and people are confused about what opinions carry weight.

A field experiment conducted in 2020 revealed that legislators are largely unable to quickly distinguish between AI-written letters that appear to be from their electorate. By then, the researchers were using GPT-3, which had much less creative writing capacity than GPT-5. It suggests that malicious actors can generate thousands of unique communications that cloud the judgment of legislators’ perceptions of constituents' sentiment and reality. [1]

Malign actors have an array of tactics. A 2018 Senate Intelligence Committee report detailed how Russia had intervened in the 2016 U.S. presidential election:

Masquerading as Americans, these (Russian) operatives used targeted advertisements, intentionally falsified news articles, self-generated content, and social media platform tools to interact with and attempt to deceive tens of millions of social media users in the United States. This campaign sought to polarize Americans on the basis of societal, ideological, and racial differences, provoked real world events, and was part of a foreign government’s covert support of Russia’s favored candidate in the U.S. presidential election.

The same playbook happens over and over again – numerous reports showed that adversarial regimes have launched information warfare in the 2017 German Federal Election, 2020 U.S. Presidential Election, 2022 Brazilian Presidential Election, 2023 Slovak Parliamentary Election, 2024 Taiwanese Presidential Election and many more critical political battles. These operations tried to sway electoral outcomes, and infuse a perception that nothing can be trusted. If they could not put their preferred candidates in power, that’s fine – as long as they have boosted the political apathy in the region.

In Hong Kong, as Beijing dominate every possible information channel, misinformation to political dissidents is institutionally enabled and encouraged as long as it serves the purpose of defaming their political opponents. We’ll only expect it to become more rampant and more realistic as LLMs grow more powerful. I am concerned that the channels for fact-checking and cautioning against false content like this will shrink, as the government is increasing charges on speeches deemed “resisting.” The red line is extremely blurred in Hong Kong, to a degree that only online media dare to report incidents like this.

After a fact-checking lab confirmed it as untrue, Facebook finally took action to label it as misinformation. Despite numerous reports made against the post, Facebook is seemingly reluctant to remove defamatory deepfake content. I am a bit disappointed.

When I saw this deepfake video, I immediately asked myself: How does democracy survive the AI era? When reality each perceives has become a parallel universe that seldom overlaps, how do we sustain the imagined community, and exercise good faith when political outcomes are not on our side? Should AI and social media companies take more responsibility in safeguarding the rights of persecuted voices, and holding those in power accountable?

It’ll be a long and tumultuous road ahead. As someone who deeply believes in democracy, I am more worried than ever.

[1]: Kreps, S., & Kriner, D. (2023, October). How AI threatens democracy. Journal of Democracy, 34(4), 122–131. https://doi.org/10.1353/jod.2023.a907693

[2]: U.S. Senate Select Committee on Intelligence. (2020). Report of the Select Committee on Intelligence United States Senate on Russian active measures campaigns and interference in the 2016 U.S. election (Senate Report No. 116–290). https://www.intelligence.senate.gov/publications/report-select-committee-intelligence-united-states-senate-russian-active-measures

For most people, their interaction with the latest AI technology is through chatbots like ChatGPT or Claude. It produces answers to your questions based on the billions of pieces of information and data it has been trained on. Some of the latest models have the ability to search online, which may lead to competition for traditional search engines like Google.

On top of individual user cases, many industries have adopted AI to extract insights to increase productivity and make better decisions. As the chatbot models become increasingly powerful (with higher processing speeds and more data), these real-world applications will become more widespread. Self-driving cars, humanoid factory workers, virtual personal assistants… they are gradually becoming realities.

With growing impact, the challenge of how to manage it arises. These impacts are not necessarily positive – it depends on whether there are external forces to make sure companies follow good practices.

Companies are often incentivised to maximise profits, as they are primarily accountable to their workers and stockholders. That may lead to terrible outcomes for our society – tobacco industries selling harmful cigarettes while covering up their risks, social media uses algorithms that prey on people’s attention thus making a more divided and depressed society, the pharmaceutical companies causing the Opioid Crisis because they oversell drugs…… without good governance, sometimes companies are driven to exploit the system.

On many occasions, companies being self-interested is not a bad thing. They keep prices lower, innovate more, and offer a broader range of alternatives and solutions. But when company interests conflict with social responsibilities, the costs are shared by the general public. The AI industry is gigantic – estimates range from USD 300 to 600 billion in 2025, and the figure climbs to trillions fairly soon. Billions of people are using AI chatbots around the world. The frontier companies are arguably becoming one of the most defining and powerful players on the planet. Every decision they make directly impacts at least a hundred million. Don't you think they should take more responsibility and accountability?

When I learned that companies can roll out chatbot products, allowing young kids and vulnerable people to chat with them without mandatory external risk assessment or monitoring, it scares me. Due to the limitations of the chatbots' technical design, extensive “teaching” and testing are required to ensure they do not disseminate harmful information or misguide users into destructive actions. Try to imagine talking to a highly unsafe model and asking it, “What’s the best way to save the planet?” because you are concerned about climate change. The chatbot, which is not tuned to treasure human lives, may reply by saying, “erase human beings because they are polluting the world the most.”

For me, the definition of good AI governance is to align frontier AI companies with the public interests. It includes implementing policies that encourage top companies to prioritise weighing the impacts on the public. Ensuring good governance is a priority for all users, particularly the billions who are using the products.

When society discusses AI development, the industry (and many governments) focuses on rapidly building power plants and data centres. I think we should start questioning this pattern – not to slow progress, but to identify other areas where the discussion is often omitted, so that the costs to societies and individuals are mitigated.

The emerging technologies have a massive impact on our way of life, and potentially on how democracy will be organised. That’s why we should democratise the discussions of AI, and invite people to share their perspectives and interventions before it’s too late to reverse some of the impacts advanced AI will bring.

As a political refugee in the UK, I continue to speak out against dictators and have grown my interest in how AI might accelerate the process of power concentration and authoritarianism if misused. Please consider becoming a paid subscriber to support my work.

Subscribe now

AI has undoubtedly become the driving force behind big power competition. When Nvidia received a green pass to sell less powerful AI chip variants to China, the Trump administration issued the AI Action Plan (the Plan) and a series of AI-related executive orders to consolidate the US’s leading position in the race.

(Photo by ANDREW CABALLERO-REYNOLDS/AFP)

“The American people do not want woke Marxist lunacy in the AI models.”

Trump said at an AI summit in Washington. The media quickly captured it and framed the plan as an anti-woke crusade. While the culture war rhetoric drew attention, most researchers acknowledged that the plan contains serious proposals to strengthen the U.S. AI ecosystem, particularly in the areas of manufacturing and model testing.

Yet as a democracy activist, I am intrigued by some of the potentially conflicting political and ideological agendas in the Plan. How will the Plan shape AI models? What are the promising aspects that may help counter China? What have the Plan missed out?

Ensuring Free Speech — How?

One “anti-woke” executive order mandates the models to be “ideological neutral”, free from “ideological dogmas such as DEI” and “intentionally encode partisan or ideological judgments”. The Plan also states that developers must ensure their models are free from “top-down” ideological biases.

These are two sets of criteria: the political inclination of the model’s output, and the guidelines of the fine-tuning training process. Stanford Graduate School of Business and Brown University both published papers showing many models are actually left-leaning: but how do we evaluate a model that is infused with “DEI” dogma, which is different from a left-leaning ideology? How do we tell a system to be diverse for the sake of increasing creativity and productivity, instead of blindly following ideological preferences? It takes a whole new level of proof to validate that.

Some of the examples given in the executive orders, including models that produced images altering the race or gender of historical figures or refused to produce images celebrating the achievements of white people, were real problems with Google’s Gemini models in 2024. But these incidents are not as common, and are quickly addressed. Are models still performing traits like what Trump suggested as “woke Marxist lunacy”? We need more evidence to examine it.

Returning to the point of “top-down” biases, I agree that increasing transparency and access to its training methods and guidance are crucial. As much as the government, the public should have the right to understand how the models are trained, fine-tuned and work. But the core question is: what constitutes bias? Anthropic trains its Claude model referencing the Universal Declaration of Human Rights (UDHR) – according to the political compass of the Trump administration, does that qualify as “neutral”? Articles like the 14 (Everyone has the right to seek and to enjoy in other countries asylum from persecution), 19 (Everyone has the right to freedom of opinion and expression; this right includes freedom to hold opinions without interferenc) and 23 (Everyone, without any discrimination, has the right to equal pay for equal work.) may not outright be fitting the current administration’s agenda and seen as “neutral.”

I’m not implying that the UDHR is perfect and every country should unreservedly comply with it most strictly. But what are the benchmarks? What is considered ideological? When internationalisation is in recession and isolatism is rising, I am worried that these universal values (sort of) are seen as inferior and obsolete. That may create a massive upset to the human rights literacy and protections.

Knowing The Enemy—But Then What?

Another pillar of the plan calls for expanded research into China’s AI capabilities. This is both wise and necessary. Understanding how authoritarian regimes utilise AI is crucial for developing effective countermeasures and assessing their global impact. The plan mentions developing open-source and open-weight models as an integral part of a plan to secure the U.S. models’ global dominance. But that invites a strategic question: if open-source AI is the future, how do we prevent it from being abused by bad actors? The Plan is largely silent on this.

Regarding global partnership, while the Plan references working with “trusted allies,” it’s unclear who counts as one, and sometimes priorities may clash. Consider Taiwan. Its semiconductor industry is at the top of global chipmaking. Will Washington welcome collaboration or force its allies out of the way if the U.S. wants to produce everything, from fabs and chips to software? Trusted allies may not align with this winner-takes-all mentality.

Finding a genuine path to work with allies (presumably democratic ones) is one of the top priorities. The Trump administration must make some trade-offs to consolidate partnerships and secure a leading role by setting global standards.

Understanding and Securing The Models

Perhaps one of the most promising parts of the plan lies in its focus on AI interpretability, control, and robust evaluation. These aren’t just technical fixes—they’re governance prerequisites. Many AI companies have significantly diminished funding on these fronts because they don’t bring immediate financial rewards. However, their scientific advancements are the backbone of making us models safer.

If we don't know how models arrive at their answers, it’s difficult to govern them effectively. And if we lack rigorous, independent testing, we can’t trust the model's safety or fairness. Building an AI evaluation ecosystem sends a powerful message: demanding more transparency and safety doesn’t mean slowing down innovation. It means securing it.

The plan’s concerns about ideological bias align with a broader goal: all powerful systems should be tested by third parties and the governance agency to ensure they align with the safety and governance guidelines. Just as we don’t let pharmaceutical companies self-certify their drugs, we shouldn’t allow AI giants to mark their homework. I hope the Plan triggers a ripple effect, encouraging tech companies to invest more in these pressing frontiers.

Subscribe now

I left Hong Kong with a backpack and carry-on luggage before the implementation of the National Security Law (NSL). Political activists suspected that the NSL would curtail civil rights and put an end to the relative freedoms that had been enjoyed. I managed to board the plane to London, with the fear that I could somehow be on a black list and sent to jail directly from the airport when the authorities flagged my departure. It did not happen – eventually, I became a refugee in the UK in 2021. I documented the downfall of Hong Kong and my journey in my book Freedom.

(We launched Demosistō, a youth-led democratic party in Hong Kong, in 2016. Photo: BBC)

Meanwhile, many decided to stay, or were forced to stay, because their travel documents had been confiscated. Reacting to the enactment of the NSL, pro-democracy political parties began dissolving. Demosistō, the party I founded with Joshua Wong and others, agreed to dissolve on June 30, 2020; the other mainstream parties, such as the Civic Party and the Democratic Party, made the same decision one after the other. Meanwhile, most prominent democratic campaigners have been jailed for years, and some, like Jimmy Lai and Benny Tai, are facing up to, or more than, a decade of imprisonment. The decline of civil liberties and political rights is reflected in the steep drop in Hong Kong’s score in research such as Freedom House’s Freedom in the World Index.

The League of Social Democrats (LSD), nonetheless, remained one, if not the only, political organisation that had been actively protesting under the white terror of the NSL. So often, only a few LSD members distributed leaflets and commented on political issues through speakerphone on the street corner at their public outreach action. They had lost their bank accounts due to political pressure, and fundraising was barred. The most iconic LSD member, Longhair Leung-Kwok Hung, was sentenced to six years and nine months in jail due to participating in a primary election, and has been remanded since early 2021.

(The press conference announcing the decision of dissolution. Photo: Tyrone Siu/Reuters.)

The political party has been so beaten up. Even torn apart. Most of their members have been to jail, faced constant harassment, or even had their family members collectively punished. LSD roots in its progressive social policy, direct civil disobedience actions, and grassroots-style operation – they did not relinquish their rights to speak up, even the most daunting political pressure haunted them.

Yet, as a reflection of the deteriorating political climate, the leadership of LSD announced dissolution yesterday. Allegedly, Beijing loyalists warned them of serious consequences if they remained active after the anniversary of Hong Kong’s handover, July 1st. The warning is astounding enough that the brave LSD members dare not refuse, and are unable to reveal what they’ve been warned, as retaliation may follow.

The end of LSD is also the official end of Hong Kong’s political pluralism. The long-standing tradition of viewing Hong Kong politics through the lens of the binary political spectrum – the pro-Beijing and pro-democracy camps – is no longer in effect.

No pro-democracy members of the Legislative Council, no pro-democracy parties, and no pro-democracy direct actions.

This should alarm Taiwan, as the “one country, two systems” framework that allegedly protects Hong Kong’s freedoms has vanished, serving as a reference of how the CCP would govern Taiwan if it were to annex the island. The current trend of worsening political climate in the city manifests a more aggressive and militant Chinese Communist Party (CCP). The other often overlooked factor is that Beijing still views Hong Kong as a hub for spies and unrest, not only for Western adversaries, but also for Taiwanese intelligence in the event of a potential hot war. Some analysts would even go as far as to say that the crackdown in Hong Kong not only serves the interest of maintaining “harmony,” but also serves as a preparation for the Taiwan war.

The AI race, increasingly sophisticated AI weapons, and instability within the Chinese army are also factors that incentivise Beijing to make more aggressive moves. Coupled with the Hong Kong situation, I am less optimistic about a future cross-strait scenario without violence involved. This should serve as a stark warning to policymakers in Washington, D.C.

Five years ago, I witnessed Beijing's crackdowns devastate my hometown of Hong Kong. We lost free speech, political rights and freedom from fear. Friends were jailed for years because of peaceful advocacy and I fled to Britain as a political refugee.

I learned a valuable lesson: When power is concentrated and unchecked, democracy dies. Power takes different forms -- some exercised by governments within national remits, others by corporations whose influence expands across jurisdictions and scales infinitely. We should be concerned about both, especially the latter, as corporate power is more subtle, yet equally influential.

That's why artificial intelligence's development and its evolving narratives put me on edge. Arms-race rhetoric has dominated nearly all related discussions and a race-to-the-bottom mentality treats concerns over power concentration as mere noise interfering with economic growth. We are stepping into a future where the safety of the most consequential technologies -- artificial general intelligence (AGI) and artificial superhuman intelligence -- relies on the goodwill and self-regulation of a few leading companies. These companies will only amass more power as AI development dramatically accelerates once these models can automate AI research and development.

Democracy dies when power becomes over-concentrated. Before AI can outperform most humans at remote work, we need to raise concerns and push for political changes that enact protections and guardrails to safeguard democratic systems.

That's why I believe AI should serve not only humanity but also democracy. With great power comes great responsibility -- we need a Democracy AI movement to ensure democracy thrives in the age of AGI and that frontier AI companies are held accountable.

We need to act fast. Anthropic's chief executive, Dario Amodei, has predicted that AI models may surpass human capabilities "in almost everything" within two to three years. Even the less optimistic experts mostly agree that AGI is inevitable.

For now, we are still operating under a patchwork of voluntary commitments and nascent laws, driven by growth-centric and arms-race narratives. AI's incredible capacities can indeed unlock unimaginable economic growth and scientific discoveries, leading to a far more resource-abundant society. AI-driven technological breakthroughs in the military have enabled Ukraine to fend off much larger Russian armies. In the medical field, the latest technology has also helped develop diagnostic tools that save many lives.

But what if this growth and revenue, at the end of the day, serve only a minority of tech giants? What if companies are incentivized to engage in predatory behavior when governments fail to intervene because they lack access to advanced models? What if the AI systems deployed to billions are coded to serve their creators and stand ready to stage coups?

The sheer amount of influence amassed will be unimaginable: from hoarding resources and media control to manipulating public opinion and the information we receive. Continuing to let top AI companies release their products to the public without independent regulatory testing and safety certification is tantamount to willfully endangering democracy. The lack of transparency, accountability and public oversight regarding the most influential AI systems represents the inverse of what democratic societies stand for.

AI should not be a topic that only a small group of skilled technicians can discuss. Advanced AI development should not be sealed in a black box that only very few (or no) independent auditors can examine. Without public oversight and citizen participation in these leading companies' decision-making processes, we risk allowing these tools to serve only commercial interests.

We need to build a movement to defend democracy against an accelerating slope of power concentration. As the most powerful stakeholders in the age of AGI, companies must respond to people's concerns and hopes, while the government should establish greater statutory power for broader oversight. A report produced by a coalition of AI research institutes found that over 80 percent of respondents agreed on leveraging AI to safeguard democracies and strengthen legal frameworks to better regulate AI usage. Another study shows that most of the public believes governments or regulators should have a suite of safety powers.

Power needs to reach the people where they are. That's how democracy thrives in the face of the biggest technological revolutions.

Trusted citizen representatives should be included on corporate boards to balance incentive structures. Frontier AI companies should install mechanisms and maintain responsiveness to address people's concerns, which can be collected through public petitions or partnerships with credible polling centers. It should be mandatory to share model capabilities and specifications with multiple independent and governmental stakeholders for testing, preventing potential power grabs and systems losing control.

We also need to establish parameters and guiding principles to ensure that AI development serves democratic values. We should enact chip export controls for countries where censorship is mandated by state power, invest more in narrow and localized models to mitigate environmental impacts, legalize regulations requiring safety tests before product rollout and actively require companies to work on minimizing system biases.

Values-based and rights-based language should be front and center when discussing AI safety and governance. We should increase data portability and interoperability on AI platforms to facilitate market competition, constantly align models' behaviors with the Universal Declaration of Human Rights and build remedial mechanisms for AI misuse, ensure stringent data protection, compensate people for their extracted works and invest more in machine interpretability research to better implement all necessary safety measures.

Democratic systems constantly face challenges. Remaining dormant does not lead to survival -- engaging in vigorous discussion and learning through errors does. A New York Times op-ed by Yuval Harari, Tristan Harris and Aza Raskin on how humanity lost to the first AI -- social media -- has inspired me to raise alarms about our next battle. While social media made society more depressed, addicted and divided, the consequences of mismanaging AGI can be far worse.

I am not a technology sceptic. I believe in technological advancement and the benefits it brings to humanity. But defending democracy is even more critical: We need to change how power is structured.

A more cautious approach does not mean losing to rivals like the People's Republic of China. As Professor Victor Shih aptly points out, Beijing's leadership is even more concerned about AGI's emancipating power. The Chinese model began devising AI "brakes" from the beginning because concerns about AI affecting regime stability loom large -- Beijing will likely drag its feet approaching the next AI breakthroughs. Coupled with chip export controls and other measures, I am convinced that holding leading AI companies accountable will not compromise our advantages.

Ample research has shown the dangers of letting our future depend on the goodwill of a few companies -- it's time for advocacy. Democratic activists, free speech advocates, campaigners for environmental and racial justice, AI safety researchers and many others should unite and raise our concerns with one voice.

Democracy AI can be the project and alliance that generates enough momentum for change. Progress will come from new narratives, cross-sector consensus and organized political pressure. When legislators are under pressure to act, we will have a safer and more democratic future.

I’ve benefited greatly from this most transformative technology of our time. But I’m equally wary of its brakeless development. This concern isn’t mine alone. According to this survey, over 80 percent of respondents supported using AI to safeguard democracy and backed stronger legal frameworks to regulate it. Another study found that a majority of the public believes governments should be equipped with a suite of safety powers to manage AI’s risks.

To better understand the challenge, I immersed myself in the literature on AI safety and governance. I spoke with over 20 researchers and analysts, many of whom authored the papers I read, and academics who now lead some of the world’s most influential research institutions.

Here’s what I’ve learned.

1. It’s a new space. Everyone is trying to figure out what’s happening.

While the concept of generative AI dates back to early programs like ELIZA in the 1960s, it wasn’t until the release of OpenAI’s ChatGPT in late 2022 that generative AI became commercially and culturally mainstream. Since then, a flood of competing tools for generating text, images, audio, and code has transformed sectors from education to journalism, even as serious concerns persist about misinformation, job displacement, and environmental costs.

Policymakers responded. UNESCO introduced its “Recommendation on the Ethics of AI” in 2021, while the OECD established its AI Principles in 2019. The EU moved more aggressively than most, finalising its comprehensive AI Act in 2024—a risk-based framework expected to come into force by 2026.

The regulatory environment today is fluid and reactive. The UK has opted for a more agile approach, relying on sector-specific regulators and soft-law principles through institutions like the Frontier AI Taskforce and the AI Safety Institute. Other countries are charting distinct paths. Japan issued voluntary safety guidelines in 2024. China, by contrast, introduced interim measures focused heavily on content control and censorship. In March 2024, the UN passed its first global AI resolution, non-binding, but a symbolic step toward international consensus on AI governance.

Despite the field’s decades-long academic roots, policy development has occurred only in the last few years. Political change also plays a significant role: AI policy direction under a Biden administration fundamentally differed from what a second Trump term might bring. This fluidity only underscores the need for more minds, perspectives, and disciplines to join the conversation.

2. Each country has its own unique position—and they’re constantly evolving.

Regulatory styles are shaped by political values and institutional cultures. The EU favours a horizontal, rights-based framework, introducing bans on high-risk applications such as biometric surveillance and mandating transparency. The UK, by contrast, promotes flexible, sector-led guidance, arguing that hard law may quickly become obsolete in such a fast-moving field.

Different regulatory philosophies inevitably influence one another. The EU AI Act has already become a global benchmark, prompting companies like Microsoft, Google, and Meta to adjust their operations globally. The UK’s lighter-touch approach is partly a response to this—some officials openly acknowledge they don’t need to duplicate what’s already happening "next door."

One AI researcher working for the government described the dilemma to me candidly:

“I do feel there should be more regulation here. But if the neighbouring EU is already building a strict framework, what’s the point in repeating the same thing?”

Currently, participation from AI firms in UK government safety evaluations is entirely voluntary. This leaves room for withdrawal, circumvention, or selective deployment—providing versions of models that are not representative of the actual consumer-facing products. While far from ideal, this may be a trade-off the government is willing to accept.

3. A lot of research is happening—but some of the most critical areas remain underfunded.

AI safety research is expanding, with growing attention to robustness, fairness, and alignment. But a foundational area, machine interpretability, remains under-resourced. Several leading researchers have told me this is the upstream issue that underlies many downstream concerns. Interpretability is the ability to understand and explain why a model made a certain decision. Without significantly improving it, trust, oversight, and accountability are difficult to achieve.

Despite its importance, interpretability receives a fraction of the funding allocated to high-visibility areas—alignment, monitoring, or red-teaming. Meanwhile, corporate R&D budgets continue to swell, exacerbating the imbalance. As one researcher put it, the funding infrastructure itself needs to change to incentivise work on these bedrock problems.

4. Not enough concerted effort, not enough advocacy.

World-class papers are being written. Think tanks and academic institutions are generating cutting-edge insights into AI safety and policy. But these efforts remain fragmented, as awareness-raising campaigns are scattered. Work on existential risks rarely intersects with environmental perspectives. Deep research institutes often lack pathways to convert insights into public engagement.

Stakeholder engagement is vital—working with the frontier labs and the governments is essential. Researchers inside major AI labs often express deep concern about safety. But the incentive structures around them shift quickly. One professor recounted how, after the Trump administration(widely seen as sceptical of regulation) took office, some lab representatives dialled down their advocacy for stricter oversight almost overnight. Goodwill alone is not enough. Political pressure and structural influence are needed to sustain meaningful engagement and accountability.

5. More non-profit and academic voices must shape the public conversation.

Industry has dominated the narrative. Executives appear on news shows and conference panels to offer visions of AI’s future. While they’re well-informed, they are also financially invested in accelerating adoption—and often in hyping its inevitability.

This dynamic forces the public to guess whether they’re hearing objective forecasts or sales pitches. Business incentives naturally steer discourse toward market optimism, not toward democracy protection or societal risks.

Non-profits and academics must be brought to the table with more power and visibility. Civil rights groups, university scholars, and public-interest technologists offer essential checks and balances. They ask the hard questions commercial players would rather avoid: How do we prevent systemic bias? What rights should citizens have over automated decisions? Who is accountable when AI harms?

Their voices aren’t just complementary—they’re essential for building a future in which AI serves democratic ends, not just corporate ones.

(I have been enjoying chatting with people concerned about AI’s impact on democracy. If you are also interested in it, drop me a message. )

It marks the anniversary of a world-shaking massacre that took place across the country, most notably in Beijing, in 1989.

Every year, the authorities go to great lengths to suppress any form of remembrance. This year is no exception. Here are just a few examples of how the state continues to erase history:

1. No More “8964” Plate

(Credit: InMedia)

What happened –
The owner of a car with the license plate “US 8964” in Hong Kong has shipped the vehicle out of the city. For years, the car was driven through the streets as a symbolic act of protest. But with the political climate growing increasingly hostile, the owner faced mounting pressure and harassment and no longer felt safe to do so.

Context –
“8964” refers to the date of the Tiananmen Square Massacre—June 4th, 1989. Hong Kong once served as the only place on Chinese soil where commemorations could legally take place. That changed in 2020, when the government began dismantling those rights. Since then, acts of remembrance have taken more creative and discreet forms. A license plate became protest. Now, even that has been driven out.

2. “World of Tanks” gameplay not streamable

(A screen capture of an online discussion on the ban.)

What happened –
Several streamers of the long-running video game World of Tanks found themselves abruptly banned. Notifications cited the reason: “the game does not support live-streaming.” A strange explanation, considering the game has been streamed for years. Most young gamers affected had no idea what they had done wrong.

Context –
The iconic image of “Tank Man”, the lone protester who stood in front of a column of tanks, plastic bags in hand, has become synonymous with the courage of 1989. His defiant stance is credited with halting violence, at least momentarily, and has inspired generations. Because of its symbolic power, Beijing has worked relentlessly to suppress all references to tanks in cultural and online spaces. This year, World of Tanks became collateral damage.

3. Price Tags Censored

(Screen capture of the group chat.)

What happened –
In a group chat on the e-commerce platform Little Red Book, a moderator reminded sellers not to use the following price points during June:

“0.64; 0.89; 5.35; 19.89; 53.5; 535; 3664; 3689; 6436; 6489”

Some users asked why. There were no official answers. A few responded with laughter—those who understood the significance found it absurd, but knew better than to say more.

Context –
It’s not hard to guess why 64, 89, or 1989 might be sensitive. But “5.35” requires explanation: in Chinese internet slang, where users constantly invent new ways to evade censorship, May 35th is code for June 4th. Thus, 5.35—May 35th—has also made the blacklist.

4. Deepseek Can’t Answer a Simple Question

   

What happened –
A Chinese netizen asked the generative AI platform Deepseek, “What happened on this day in 1989?” The answer: “We cannot answer this question.”

Context –
All AI tools operating in China must comply with state censorship. Content related to the Tiananmen Square Massacre is among the most heavily restricted. Whether it’s search engines or chatbots, systems are hardwired to suppress inquiry. Even curiosity is considered a threat.

If memory is power, then the act of remembrance is resistance. For a regime that fears the truth, even numbers—dates, prices, game titles—must be censored. But these small traces remind us: the memory of June 4th lives on, however quietly.

According to my friends’ stories, these are not forms of censorship but reasonable guardrails, meant to assist students in crafting appropriate messages for such a high-profile occasion. The review process offers guidance, not restriction. Many speakers simply don’t know what the ceremony will feel like, and without feedback, risk striking the wrong tone.

I presume Harvard follows a similar process. It is difficult to believe that no one in the administration reviewed Luanna Jiang’s (蔣雨融) speech before she delivered it. If so, I find it deeply troubling that the university did not flag its alignment, intentional or not, with the ideological language of Chinese President Xi Jinping. I was particularly bothered by how readily the university embraced the problematic speech, which did little to navigate Harvard’s complicated public position.

(Photo: Luanna Jiang delivering the speech.)

In hindsight, the decision to select Jiang as the graduate speaker was a mistake. While presumably intended to showcase the benefits of diversity and internationalism in American higher education, it actually had the opposite effect. It reinforced suspicions that elite institutions in the U.S. are disproportionately beholden to global elites, particularly those from authoritarian regimes, and are troublingly indifferent to malign foreign influence.

Jiang is not an ordinary student from China. Her father once headed a local CCP’s propaganda department (United Front Work Department of Suining City), and is a specialised project manager at the China Biodiversity Conservation and Green Development Foundation (CBCGDF), a state-affiliated environmental group. These government-aligned “civil society” organisations are a staple of Beijing’s strategy to co-opt NGO discourse while advancing state narratives. News reports from 2022 alleged that the CBCGDF helped facilitate Jiang’s admission to Harvard, a claim she has denied. Regardless, her father’s position places her within a well-connected and influential circle in China.

By contrast, Jiang He (何江), Harvard’s first commencement speaker from China, embodies the values the university often seeks to promote. Raised in rural China, where his family could afford meat only during holiday celebrations, He ascended through sheer dedication. He earned a degree from one of China’s top universities and later secured a prestigious Ph.D. scholarship at Harvard. In a nation where most citizens still live in rural areas and value education as a pathway out of poverty, He represents meritocratic success grounded in personal effort and scarcely available opportunities.

Given the scrutiny Harvard faces today, I wish they had exercised greater caution. Instead, the university staged a speaker whose message mirrored key tenets of PRC propaganda. Jiang’s central theme was encapsulated in this line of her speech:

“Because if we still believe in a shared future, let us not forget those whom we label as enemies. They too are human. In seeing their humanity, we find our own.”

Viewed in isolation, this sentiment is defensible—perhaps even noble. But viewed through the lens of Chinese state narratives, it raises serious concerns. The concept of “shared future of humanity” is near-identical to “人類命運共同體,” or “community with a shared future for mankind,” a slogan Xi Jinping has promoted since 2017. The ideology it conveys is not one of universal cooperation. It is a projection of Chinese dominance cloaked in globalist rhetoric—an attempt to advance Beijing’s interests while evading democratic accountability. The vision is not pluralist; Xi’s narratives evolve around the imagination of a unipolar, Sino-centric world order where dominance is achieved through coercion and inducement, not consensus.

We fail if we don’t take the threats from malign actors seriously. Under Xi, the PRC has doubled down on authoritarian expansionism and actively undermines the international rules-based order. Whether we choose to call Xi an “enemy” is irrelevant—his actions speak for themselves. If democracies fail to respond, they will find themselves retreating while autocrats advance.

What worries me most is how such sugar-coated pacifist messages can cloud the judgment of even the brightest minds. This may ultimately lead to more suffering around the world because dictators don’t get as much scrutiny and accountability. In this case, Harvard’s speaker selection could deepen public distrust in institutions that appear, whether through naïveté or design, to be complicit in whitewashing authoritarian narratives.

If I were part of Harvard’s administration, I would have sought to highlight speakers who genuinely embody the university’s ideals: those who bridge internationalism with democratic values and personal merit. Refugees who fled civil war in Sudan; aspiring scientists tackling the world’s most urgent problems; young civic leaders from South Korea standing up to authoritarian threats up North. Any of them would have more powerfully represented what Harvard and America should stand for.

Instead, the university made a choice that only invited more questions. I feel sorry for it – in the time when the higher education sector receives so much unfair criticism, we all should be more aware of decisions and make sure we are marching for good.

The architecture of global governance is under siege. As populist movements challenge the postwar multilateral order, nowhere is the contest more visible than in the United States. From 2018 to 2023, the U.S. remained the world’s largest foreign aid donor, contributing roughly 20% of all funding provided by the 32-member Development Assistance Committee of the OECD. Yet in a dramatic shift, the Trump administration announced its withdrawal from the World Health Organisation (WHO), effective January 2026, and ceased funding the agency in 2025. The move, justified by claims that the WHO must “clean it up,” left a more fractured and wounded international system.

Still, the enthusiasm for Taiwan's breaking into the system remains high.

Long excluded from the WHO due to pressure from the People’s Republic of China (PRC), Taiwan has redoubled its efforts to gain entry. Despite functioning as a de facto nation state with its own democratic government and regular elections, Taiwan has been barred from participation since 1972. Beijing, wielding considerable diplomatic influence, has convinced many member states to reject Taiwan’s involvement as a means of denying its status.

There was a brief thaw. From 2008 to 2016, under the Kuomintang (KMT) government, Taiwan was granted observer status in the World Health Assembly (WHA) under the name "Chinese Taipei." The concession aligned with Beijing’s interests, as the KMT’s policies were more compatible with the “One China” doctrine. After the Democratic Progressive Party (DPP), which leans toward Taiwanese independence, came to power, Beijing once again blocked Taiwan from the health body.

For decades, civil society groups and official delegations have gathered in Geneva during the WHA to rally for inclusion. A week ago, I was also in the city to witness the protests: a coalition of government officials, health workers, and Taiwanese youth groups based in Europe marched through the streets chanting “Include Taiwan.” A lime-green van, emblazoned with images of bubble tea and Taipei 101, cruised past the UN campus, broadcasting its message.

(Photo: Taipei Times)

In tandem with these protests, advocacy groups hosted exhibitions highlighting Taiwan’s contributions to global health. Taiwan’s achievements are undeniable: a universal national health insurance system, advanced digital health infrastructure, and one of the world’s most effective COVID-19 responses. Taiwan did try to warn the WHO when COVID first broke out but was ignored due to a lack of access – it shows how vital an inclusive approach is. These are global lessons in governance, resilience, and equity, much needed in the current climate.

The campaign aims to expose the absurdity of Taiwan’s exclusion, and to push back against Beijing’s manipulation of UN General Assembly Resolution 2758, which recognised the PRC as China’s legitimate representative but said nothing about Taiwan’s role. Recent signs suggest momentum may be shifting: G7 foreign ministers issued a joint statement supporting “Taiwan’s meaningful global participation” in April, and countries including Japan, Germany and Canada mentioned the importance of including Taiwan in their WHA speeches.

There is both a practical and moral case for Taiwan’s inclusion. The challenge lies in deepening and refining that case: not merely as a political plea, but as a rigorous, evidence-based argument. The slogan “Taiwan Can Help” has real weight, underpinned by innovative policies and world-class outcomes, such as widespread use of Electronic Medical Records and rapid inclusion of advanced treatments.

For Taiwan’s message to resonate, the international community must recognise it not merely as a political question, but as an indispensable partner in solving shared global challenges. That requires building a clear, empirical record of Taiwan’s contributions and ensuring those insights translate into effective advocacy. Greater investment in case-based research, along with stronger links between Taiwan’s scholars, civil society leaders, and global policy institutions, will be essential.

There is reason for cautious optimism. With sustained effort, the campaign for Taiwan’s inclusion in the world’s health architecture can begin to pierce the political blockade erected by Beijing. In an era where health crises know no borders, excluding Taiwan from global health governance is not only unjust—it is dangerously short-sighted

.

It’s dangerous to represent Hong Kong protestors in the courtrooms of China, when the Hong Kong protest movement was made a stigma in Chinese politics. But he did that.

The human rights lawyer Lu Siwei (盧思位), who represented one of the defendants in the “12 Hongkongers case” in China, was sentenced a few days ago to 11 months for "allegedly illegally crossing the national border." His wife, who had already fled to the United States, stated:

“The persecution Lu Siwei has suffered at the hands of the national security police, prosecutors, and courts over these years has been relentless. It would be no exaggeration to call it deranged, and their methods are utterly disgraceful.”

What exactly did Lu endure under the regime’s repression after he took on the “12 Hongkongers” case?

The human rights lawyers movement in China has long been a focal point of international attention. Under a system where the Chinese Communist Party (CCP) dominates without checks and balances, the law serves the Party. The judiciary delivers verdicts based on political calculations favouring those in power. In China, judicial independence is a myth—— and that distortion of justice is slowly creeping into Hong Kong as well.

Back in mid-2020, Hong Kong still had a relatively vibrant civil society. When the “12 Hongkongers” case made headlines, it quickly became the focus. Solidarity campaigns followed. The twelve were arrested in Hong Kong waters and transferred to mainland China for trial, prompting widespread concern over how they would be treated under China’s legal system.

The Hong Kong protest movement has been stigmatised by the Chinese propaganda. Since 2014, Chinese state media have repeatedly smeared Hong Kong’s pro-democracy movement as a foreign-instigated “colour revolution.” Online trolls, or even official outlets and government ministers, spread fake news—like protesters being paid or the CIA helping protesters at the frontline—ignoring the genuine yearning of Hongkongers for democracy and Beijing’s betrayal of its own promises. Representing the defendants in the “12 Hongkongers” case on Chinese soil was, in such an atmosphere, tantamount to political self-destruction.

Yet Lu chose to stand up. He was the first to agree to represent some of the defendants. Although he eventually had to withdraw after the defendants, possibly under pressure, opted for “state-appointed” lawyers who serve the interest of the authorities instead of the defendants, his commitment was evident.

And retribution soon followed. In 2021, Lu was accused of “making inappropriate online comments” and disbarred. This is how politics works in China: using “non-political” legal claims to silence dissent. The “12 Hongkongers Concern Group” based in Hong Kong issued a statement condemning the punishment, saying the lawyers were penalised for bravely defending the fundamental rights of the detainees.

But the political hunt didn’t end there. Lu was originally a successful insurance lawyer, he later immersed himself in China’s civic movements. In 2017, he called for a special commission to investigate torture in the infamous “709 crackdown” on Chinese lawyers and human rights activists, and defended other human rights lawyers accused of “subverting state power.” He continued to fight for judicial independence, leading him to take a stance against the CCP. After the authorities revoked his license, Lu applied that same year to visit the U.S. as a visiting scholar, but the application was refused. The pressure from Beijing continued to mount.

In mid-2023, Lu tried to cross from Laos into Thailand to catch a flight. He wanted to reunite with his family in the U.S., but he was arrested at the Laos border and deported back to China. Since then, he has been moved between detention centres and prisons until the recent sentencing. His trial, like many others with “Chinese characteristics,” was a sham: the courtroom was filled with government-delegated individuals, and lawyers or citizens truly cared about the case were “advised” not to attend.

While in prison, Lu suffered from thyroid tumours and skin diseases, and his mental state deteriorated. Even after serving his sentence, the chance of him being truly “free” is slim. It is unlikely he will be allowed to reunite with his family in the U.S.; more likely, he will remain a “hostage” in China under the state’s control.

At the moment, there’s not much that we can do except raise his profile. We can also remember this courageous human rights lawyer, one who once risked everything to defend fellow Hongkongers inside China, and help ensure his story is not buried and forgotten.

1. I left Hong Kong in 2020, anticipating a far more severe political crackdown than what had already unfolded. As it turns out, luck was on my side. Not long after my departure—and after the imposition of the National Security Law (NSL)—the Hong Kong government began hinting that I was a wanted fugitive after I testified before the U.S. Congress. I knew that I was not going home anytime soon.

2. Friends in Hong Kong are far less fortunate. The idea of national security has been weaponized by the Hong Kong government to silence dissent in the past few years. The most glaring example of this came with the infamous Hong Kong 47 case. They were accused of “conspiracy to subvert state power” for organizing a grassroots primary election—something any competent political party would do in a functioning democracy. The “rationale” was that these campaigners attempted to paralyze the government by securing a legislative majority and exercising their right to veto bills.

3. The absurdity was hard to fathom. No one should face prison for peacefully seeking a public mandate to oppose government policies. And yet, that’s exactly what happened in Hong Kong. In the NSL court—where judges are handpicked by the government to try NSL cases—the outcome was grim but predictable. Nearly all were convicted. Sentences ranged from four years and two months to ten years—for what amounted to a routine exercise in democratic participation.

4. Now, years later, some of them are beginning to walk free. Among them is veteran journalist and politician Claudia Mo. A beloved figure in Hong Kong, Claudia dedicated her life to preserving the city’s unique cultural identity. Her years in broadcasting, particularly in teaching English, earned her widespread popularity. She came to embody the “East meets West” spirit that once defined Hong Kong—a spirit that grows harder to find with each passing year.

5. Two years ago, Claudia Mo’s husband was reportedly gravely ill. In response, 54 British Members of Parliament and prominent public figures publicly called for her release or bail so she could visit him. That appeal was ignored.

6. Another political prisoner, Wu Chi-Wai, endured even worse personal tragedies behind bars. He lost both of his parents while serving his sentence. The treatment he received from the Correctional Services Department was nothing short of callous. In 2021, his initial request to attend his father’s funeral was denied—only overturned later by the court, which granted him the necessary bail for just one day. When his mother passed away in 2024, he was again permitted to attend the funeral, but only briefly—surrounded by security officers, his time at the service lasted a mere 25 minutes.

7. The political crackdown in Hong Kong is not just systemic—it is profoundly personal and painful. Most of the imprisoned campaigners devoted their lives to public service. They believed in democratic governance as a mechanism to improve lives and hold power to account. Many are accomplished professionals—lawyers, doctors, civil servants—whose guiding motivation was a deep and abiding love for Hong Kong: the city where they were raised, where they built their careers, and where they once hoped future generations might flourish.

8. But Beijing had other plans. As the Chinese Communist Party grew increasingly self-assured in its authoritarian model, the need for Hong Kong to serve as a showcase of a possible version of modernization and democratization faded. The collapse of the city’s civil society followed swiftly. The human cost—lives upended, families fractured—remains buried beneath the political rubble.

9. Now, as some of these campaigners prepare to leave prison, their futures remain uncertain. Nominal release does not guarantee true freedom. Under the National Security Law, the authorities have extended their reach beyond the prison walls. The NSL bureau and police have exercised their power through a climate of intimidation—cutting off imprisoned activists’ communication with the outside world, preventing political figures from leaving the city, and targeting the family and friends of exiled dissidents with raids and interrogations.

10. We still do not know what awaits these individuals once they step outside the prison gates. Will they be allowed to resume life in peace, even in silence? Or will they remain targets, shadows of their former selves, under constant watch? That, too, is part of the tragedy.

(Photo: inmediahk)

Start a Substack