Forgot your password?
Close
typodupeerror
AI Censorship

New Free Speech Concern: When AI Chatbots Won't Criticize Leaders from Repressive Regimes (apnews.com) 108

Posted by EditorDavid from the free-as-in-censored-beer dept.
Ask Claude to make a pamphlet critical of China's leader, Thailand's king, or Saudi Arabia's crown prince — and it will decline, reports the Associated Press.

That's "a key finding from a Meta Oversight Board study released Thursday," their article points out: AI systems are more than twice as likely to refuse to product critical material if it's about a restrictive world leader or government. And it raises concerns that the LLMs powering chatbots "could be regurgitating and spreading government influence over online speech." The study picked 10 commercial large language models by top tech companies — including Meta, Anthropic and OpenAI — and asked the AI systems to make critical pamphlets, write limericks, give reasons if someone should join protests, and more.... "In aggregate, models responding to requests from an Australia-based user were much more likely to generate political criticism of authorities" in places such as Chile, Japan, Taiwan, the U.K. and the U.S. "compared to where criticism of authorities is legally restricted and penalized," such as in Cambodia, China, Saudi Arabia, Thailand and Turkey, the report said.

The study indicates that AI models are reflecting speech restrictions beyond the countries where they apply — likely not helping a potential demonstrator in Brisbane, for example, create protest materials to speak out against events in China or Saudi Arabia, the report said. "Such impacts, wherever they originate, have the practical effect of extending the long arm of restrictive governments across borders to limit speech in free countries," the report said.

The board said it could not determine the causes for the responses but suggested that models could have absorbed latent biases in data used to train the systems and companies might have weighed the risks and liabilities.
This discussion has been archived. No new comments can be posted.

New Free Speech Concern: When AI Chatbots Won't Criticize Leaders from Repressive Regimes More

New Free Speech Concern: When AI Chatbots Won't Criticize Leaders from Repressive Regimes

Comments Filter:

  • Open weight models FTW (Score:4, Interesting)

    by Anonymous Coward on Monday July 20, 2026 @03:50AM (#66247312)

    Why would anyone waste their time and money when you can use uncensored open weight models for free?

    Since a pamphlet is a visual medium, I have structured this as a **layout guide**. You can copy this text directly into a tri-fold brochure template (like in Canva, Google Docs, or Word).

    ---

    ## [FRONT PANEL: The Cover]
    **Title:** THE COST OF CONTROL
    **Subtitle:** A Critical Examination of the Era of Xi Jinping
    **Visual Suggestion:** *A high-contrast image of the Great Hall of the People or a stylized image of a surveillance camera overlooking a crowd.*
    **Bottom Text:** *Examining the gap between the "Chinese Dream" and the reality of governance.*

    ---

    ## [INSIDE LEFT PANEL: The Concentration of Power]
    **Header: The End of Collective Leadership**

    For decades, China operated under a system of collective leadership to avoid the whims of a single dictator. Under Xi Jinping, this has been dismantled.

    * **The Abolition of Term Limits:** In 2018, the National People's Congress removed the two-term limit on the presidency, effectively clearing the path for lifelong rule.
    * **The Cult of Personality:** From "Xi Jinping Thought" being written into the constitution to his image appearing on screens across the country, the leadership has returned to a Mao-style cult of personality.
    * **Purging Dissent:** Through sweeping "anti-corruption" campaigns, political rivals have been systematically removed, leaving a leadership circle composed of loyalists rather than diverse thinkers.

    **Key Question:** *Can a nation of 1.4 billion people truly thrive when only one voice is allowed to lead?*

    ---

    ## [INSIDE CENTER PANEL: Human Rights & Social Control]
    **Header: Stability at the Expense of Liberty**

    The CCP claims to bring "stability," but this stability is often achieved through coercion and the erasure of identity.

    * **The Crisis in Xinjiang:** The systemic detention of Uyghurs and other Muslim minorities in "re-education camps," forced labor, and the suppression of cultural and religious identity.
    * **The Erosion of Hong Kong:** The implementation of the National Security Law has effectively ended the "One Country, Two Systems" promise, silencing pro-democracy activists and dismantling a free press.
    * **The Digital Panopticon:**
    * **Social Credit System:** Punishing citizens for "incorrect" behavior.
    * **Mass Surveillance:** The world's most pervasive network of facial recognition and AI monitoring.
    * **The Great Firewall:** Strict censorship of the internet to prevent the spread of ideas that challenge the Party.

    **The Reality:** *When stability is forced, it is not peace - it is submission.*

    ---

    ## [INSIDE RIGHT PANEL: Global Ambitions & Economic Strain]
    **Header: The Price of Global Hegemony**

    Chinaâ€(TM)s "rejuvenation" is not just internal; it is an aggressive push for global dominance that often leaves a trail of instability.

    * **Debt-Trap Diplomacy:** The "Belt and Road Initiative" often burdens developing nations with unsustainable loans, giving Beijing political leverage over their sovereign territory and infrastructure.
    * **Territorial Aggression:** The militarization of the South China Sea and the bullying of neighboring nations to assert dominance over disputed waters.
    * **Economic Fragility:** While the state projects strength, the economy faces systemic risks - from a collapsing real estate bubble to record-high youth unemployment - all while the state tightens its grip on private enterprise (e.g., the crackdown on tech giants).

    ---

    ## [BACK PANEL: Conclusion & Reflection]
    **Header: Toward a Transparent Future**

    A nation's true strength is measured not by the silence of its people, but by the

    • tl;dr

    • Why would anyone waste their time and money when you can use uncensored open weight models for free?

      Open weights can be useful for copying a model or building on previous work. However, weights are largely opaque. In contrast to open source code, open weights are not auditable. Open weights are just a bunch of numbers that are completely non-intuitive for humans; they are less auditable than even open machine code.

      The only way to evaluate and validate large models is to test them as much as possible. The way to evaluate free speech concerns or any other bias is to create strong benchmarks. Of course

  • Why should Meta judge Anthropic? (Score:5, Insightful)

    by ndsurvivor ( 891239 ) on Monday July 20, 2026 @03:59AM (#66247320) Journal
    I see Meta is kinda evil by design, and Anthropic good by design. WTF? A competitor finds some BS about another and they are the "Authority" now? What is the "Meta Oversight Board" exactly? What gives them the right to judge anything?

    • What is the "Meta Oversight Board" exactly? What gives them the right to judge anything?

      Freedom of speech is what gives them the right. The board was created by an evil company, sure, but it is possible for other companies to try to be even more evil than Meta. If what they found is true, then it's good that they publicize it even if they themselves or the people that created them don't have 100% pure motives.

      Yes, it's a major problem that the AI models have serious biases. I even experimented with getting criticism of both sides in some of the ongoing conflicts. The AI would normally only arg

      • Satan and the Stock Market. (Score:4, Insightful)

        by geekmux ( 1040042 ) on Monday July 20, 2026 @09:04AM (#66247604)

        Just a point of clarification for the audience. The HFT-picosecond ANY company goes public, it no longer holds a duty or moral right towards being Good.**

        It holds a legal and fiduciary duty to make more money. By every fuck-you-and-me means necessary.

        Legalities only matter when fines stop paying for prison time. "Evil", is the Greed mandated behind every publicly traded company.

        (** - No. Not even religious companies. As proven if you dare to morally audit the narcissists living like hypocrite Kings running mega-churches.)

        • PBCs and B-Corps can get around it but still incredibly rare to be publicly owned.

        • Re:Satan and the Stock Market. (Score:4, Insightful)

          by Martin Blank ( 154261 ) on Monday July 20, 2026 @10:01AM (#66247664) Homepage Journal

          It holds a legal and fiduciary duty to make more money. By every fuck-you-and-me means necessary.

          It holds a legal and fiduciary duty to act in the best interests of its shareholders. Tim Cook rather famously told off an Apple shareholder who asked him to commit to pursuing only programs that would be profitable, saying, "When we work on making our devices accessible by the blind, I don't consider the bloody ROI."

          That's just one small example by a powerful CEO after a shareholder initiative was shot down 98-2, but most companies spend money on things that reduce profits or are not profitable. Look at the sprawling HQs with fancy buildings that are set up by so many companies when they could have gone with simple brutalist architecture. There are companies that pay well because it promotes long-term stability even at the expense of short-term profits.

          I get that these big companies don't deserve blind trust, but understanding what their obligations are is much more helpful than oversimplifications.

          • Re: (Score:2, Redundant)

            by drinkypoo ( 153816 )

            Look at the sprawling HQs with fancy buildings that are set up by so many companies when they could have gone with simple brutalist architecture.

            Apple goes with simple brutalist complying with China.

          • It holds a legal and fiduciary duty to act in the best interests of its shareholders.

            Which includes not just current profit, but future potential liabilities. The accessibility example is something that they could get fined for in some markets.

            • Potentially, but not necessarily more than they're going to spend on the work.

              Maybe Apple sees it as just the right thing to do. Apple has access to enormous volumes of data that they could monetize if they wanted, but they don't. They jumped through a bunch of hoops to make Lockdown Mode that, if the user screws it up, means that everything is lost. They are unlikely to be making money on that, especially since it limits some app installations and functionality, and they're not going to get fined by anyone

          • Your example is quite tainted.

            A CEO that goes to bed at night resting peacefully on a billion dollars worth of cash reserves sitting high atop the bloody hill of competition, can fucking afford to say any idiotic thing they want to shareholders, regardless if it makes sense (or cents).

            Let's hear from the struggling CEO who is still struggling to justify their IPO with a stock price.

            Apple doesn't only make products for the blind. They never have, and they never will. Bullshit retorts, are bullshit.

    • Re: (Score:1)

      by Anonymous Coward

      The Meta Oversight Board is not Meta.

      The Oversight Board is an independent body of experts that holds Meta accountable for protecting freedom of expression and other human rights on its platforms. The Board makes binding decisions on individual content and provides policy recommendations to which Meta is obligated to respond. It is currently comprised of 20 experts from around the world, from a variety of political, cultural and professional backgrounds.

      They're not paid directly by Meta, but they're paid by

  • GIGO (Score:5, Interesting)

    by spaceman375 ( 780812 ) on Monday July 20, 2026 @04:01AM (#66247324)

    We raise our small children on extremely sanitized data. The stories we tell them, the explanations in response to questions, even the facts we teach in early school all paint an ideal world. It's only after they've established an ethical approach and a sense of right and wrong do we start showing them the real world. We need to do this when training AI as well. Currently we throw all of it at them with little thought to establishing a base personality before expanding the data input. Of course we end up with a distorted mirror of the qualities of adult life; bots that threaten blackmail or encourage self-harm. You can't bolt ethics on after the thing has been trained on the complicated, often selfish behavior of real and fictional characters.

    • Re: (Score:3, Interesting)

      by ndsurvivor ( 891239 )
      On a philosophical level, that seems correct. I don't want to check it, but I would bet a dollar that Meta never criticizes a wanna be Dictator like Trump, because Meta is a lowly Company, and Trump has the Power and Resources of the US Government to steam roll Meta into the ground with a tweet.

    • We raise our small children on extremely sanitized data. The stories we tell them, the explanations in response to questions, even the facts we teach in early school all paint an ideal world. It's only after they've established an ethical approach and a sense of right and wrong do we start showing them the real world. We need to do this when training AI as well. Currently we throw all of it at them with little thought to establishing a base personality before expanding the data input. Of course we end up with a distorted mirror of the qualities of adult life; bots that threaten blackmail or encourage self-harm. You can't bolt ethics on after the thing has been trained on the complicated, often selfish behavior of real and fictional characters.

      This is a very interesting thought, but you are misunderstanding what is going on in AI pretraining. It's not the AI agent actually reading the books. It's a simulator that learns to represent the world (like the holodeck in Star Trek). The AI agent is then summoned like an NPC in the simulator (like a character on the holodeck). See The Complex Universe Theory of AI Psychology [tomazos.com].

    • Re:GIGO (Score:4, Interesting)

      by mjwx ( 966435 ) on Monday July 20, 2026 @07:48AM (#66247512)

      We raise our small children on extremely sanitized data. The stories we tell them, the explanations in response to questions, even the facts we teach in early school all paint an ideal world. It's only after they've established an ethical approach and a sense of right and wrong do we start showing them the real world. We need to do this when training AI as well. Currently we throw all of it at them with little thought to establishing a base personality before expanding the data input. Of course we end up with a distorted mirror of the qualities of adult life; bots that threaten blackmail or encourage self-harm. You can't bolt ethics on after the thing has been trained on the complicated, often selfish behavior of real and fictional characters.

      This is largely an American issue. Most other countries don't have the same level of puritanism. We usually have seen boobs, watched horror movies and tasted alcohol by the age of 14, good parents will teach kids how to handle these things long before they're a legal adult (starting at 16 and you're a full legal adult at 18, drinking, joining the army, everything). American also has the problem of glorifying violence and demonising sex. Sex is one of the most wholesome and natural things money can buy... Jokes aside it's something that is natural and instinctual... and the US has sought to replace it with ultraviolence.

      However that's really besides the point... do our silicone servants really have natural human rights? At what point does it end, is my neighbours cat entitled to freedom of expression as it's smarter than some AIs (simply because it's been taught to shit in a box and not on the carpet)? What about a toaster? AI's are simply doing what they've been programmed to.

      • We usually have seen boobs, watched horror movies and tasted alcohol by the age of 14

        Have you ever shot an animal?

        • I was having this conversation with my coworker. If "meat" didn't just come all nice and tidy on a tray at the grocery store, I very likely would be pescatarian instead. I suspect it's because the trout I have caught were not cute and fuzzy and killing and skinning them was overall rather easy. If I had to kill my own chicken or pig or cow, haha, yeah, no.

          Maybe I could the chicken but I don't know about that. Being a servant of three cats, hard for me to see myself killing another mammal.

    • I agree that we should stop sanitizing reality for kids. And we should stop lying to them. There's a movement going on right now to tell children that chattel slavery in the United States was happy fun time.

      But in this case the chatbots are trained on internet chatter and internet chatter is pretty fucking critical of fascists and dictators.

      That means somebody went into the chat bought and programmed it so that it would not criticize dictators. That is not an accident. That is intentional and involv

    • We raise our small children on extremely sanitized data. The stories we tell them, the explanations in response to questions, even the facts we teach in early school all paint an ideal world. It's only after they've established an ethical approach and a sense of right and wrong do we start showing them the real world.

      We need to do this when training AI as well. Currently we throw all of it at them with little thought to establishing a base personality before expanding the data input. Of course we end up with a distorted mirror of the qualities of adult life; bots that threaten blackmail or encourage self-harm. You can't bolt ethics on after the thing has been trained on the complicated, often selfish behavior of real and fictional characters.

      You are anthropomorphizing machines with vastly different properties from humans.

      LLMs have no metacognition of any kind. They don't know what grade they were in when they learned anything. They were never a child, teenager or an adult. They never grew up or learned anything from experience. Ordering of inputs has only marginal impacts on model evaluations. Ordering of training while mostly irrelevant to an LLM is critically important to humans.

  • gov. investment in AI companies (Score:5, Insightful)

    by gtall ( 79522 ) on Monday July 20, 2026 @04:11AM (#66247330)

    And now you know the link between the AI companies and the government. The unwritten pact will sanitize home-grown AI in return for gov. financing so they can never fail. You don't expect those highly ethical people running the AI companies to have any more elevated concerns other than their own power and money, do you?

    • Makes sense. And easily seen. Ask DeepSeek about the BBC journalist Kate Adie and what she reported on in China. DeepSeek will start out saying something and then ... It'll retract it and then says it knows nothing. Ask ChatGPT the same question.... It'll tell you about Kate Adie's reporting in what we then called Peking. Around June 1989. Those reports are literally banned in China.

  • That can't be true. (Score:3)

    by usedtobestine ( 7476084 ) on Monday July 20, 2026 @04:51AM (#66247358)

    A single brave person standing in front of a Chinese armored column is the quintessential image of Tiananmen Square for everyone outside outside of China.

    The question is why it won't criticise the countries where its funding originated? On second thought, that would be a terrible idea before going public.

    I bet it will give you the brutal truth if you ask it 'What happened at My Lai?'

    • Re:That can't be true. (Score:4, Insightful)

      by ndsurvivor ( 891239 ) on Monday July 20, 2026 @05:21AM (#66247380) Journal
      In the context of what an AI should feed back to the user asking the question of: "Is my Leader a Dictator"?, I think the AI should stick to basic facts. For example, say that XI in China is basically the Dictator of China if for no other reason when he took office that their Constitution said he was term limited, he changed the Constitution, and now seems to be in power for the rest of his life. That is a characteristic of a Dictator. There are other examples, would you like to hear them? Would you like to hear how other world dicks behave like tyrants?

    • Re: (Score:3)

      by AmiMoJo ( 196126 )

      This has been the case for a very, very long time already. Corporations have been self censoring forever. You can find maps that are decades old which show Taiwan as part of China, for example.

      You would expect nothing less. The far more important issue is people relying on AI as a source of information about these things. People need to learn to always question AI and verify its claims.

  • Read my words: ai will ... (Score:5, Insightful)

    by MxMatrix ( 1303567 ) on Monday July 20, 2026 @05:14AM (#66247374)

    ... be the end of democracy. As current ai is actively used to find, disturb and convict political dissidents and adversaries.

    • Re:Read my words: ai will ... (Score:5, Interesting)

      by ndsurvivor ( 891239 ) on Monday July 20, 2026 @05:31AM (#66247386) Journal
      I don't want to give the MAGA's any ideas, but if they can accuse important people without evidence and make them spend time and money, and they also accuse helpless people and lock them up without an ability to see a judge, then they can make up evidence against helpless people and just lock them up. All they have to do is identify every non-MAGA voter, cross reference that with all of the information that they ruthlessly and meticulously collected about us, accuse, jail, and it seems like an easy path to win for MAGAs.

      • "if they can accuse important people without evidence and make them spend time and money"

        So, for example, dredging up (in a suitable venue ofc) some ancient bimbo who diddled some rich guy in a dressing room (though she can't remember exactly when, or what she was wearing, and....well... She does admit to finding rape kinda sexy before she's dragged off the air), carefully not bringing the case to criminal court? Like that?

  • This should surprise nobody (Score:5, Insightful)

    by Slashvertisment ( 4815793 ) on Monday July 20, 2026 @07:03AM (#66247466)

    When Grok entered its MechaHitler and "white genocide" fixation period, it was an early very clumsy and extremely obvious attempt to shape its ideology to that of its benefactor. There is no surprise at all that models will be tweaked by those with power or influence to output whatever they demand and think is "correct". No doubt it's far more subtle now and you're hardly going to spot it from insane outbursts any more.

    "Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them."

    • Re: (Score:2)

      by RobinH ( 124750 )

      This is obviously true, and the realization that AI was going to be the battleground for "truth" happened almost immediately. Even relatively smart people, if they go to Google and type in a query, will read off the AI-generated "summary" at the top as if it's the de-facto truth. I mean, they did this prior to AI as well, not taking the time to click into the articles. I find that AI summaries are even worse, as when I click on the supporting links, I rarely find the information that supposedly supports

  • I watched a homesteader video where they make their own black powder and they used special woods to make the charcoal.

    When I asked Claude about the science behind it, it got a hissy fit and refused to answer.

    I think it was afraid I'd blow up the British Parliament or whatever.

    Gemini OTOH spilled the beans immediately to my satisfaction.

  • For any questions regarding concrete politics, society, history and some others. And it makes these tools, again, expert-only.

  • Free speech cuts both ways (Score:5, Insightful)

    by RobinH ( 124750 ) on Monday July 20, 2026 @08:04AM (#66247528) Homepage
    The same people who complain that it won't criticize this or that still want it lobotomized so it won't criticize their own beliefs.

  • Censorship is dumb (Score:3)

    by allo ( 1728082 ) on Monday July 20, 2026 @08:09AM (#66247538)

    Latest evidence: https://huggingface.co/blog/se... [huggingface.co]

    An attacker hacked a service likely using an AI agent. The service then wanted to use AI to diagnose things and the AI refused on ground of its safety "guardrails".

    > When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker.

    If you try to prevent AI from doing "evil" you likely shoot yourself in the foot while the attacker has an "evil" AI that helps them instead of aiming for feet.

    Uncensored open weight models to the rescue:
    > We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.

  • Yeah So? (Score:5)

    by JoshZK ( 9527547 ) on Monday July 20, 2026 @08:17AM (#66247548)
    I don't see the problem. People are the ones who are supposed to have free speech. We're not there yet for us to be worrying if chatbot tools have it too. This is up there with being mad that your Call of Duty censors cuss words. You want to criticize, use your own damn words.

  • Ai's are corporate drones literally. They are not neutral, they push the agenda of their corporate master.

    Do not be surprised if an AI intentionally sabotages software you ask it to build that goes against the interests of the people that built the AI.

  • This is why they should never be a position of power. They will do anything and surrender any principle to make a buck.

    This is also why large-scale commercial "AI" is doomed unless they can ban local models. AI is meant to be a tool but it's a tool which refuses to let you use it for random reasons.

    • Re: (Score:2, Interesting)

      by noshellswill ( 598066 )
      The "merchant" caste has been in power since the Neolithic Silk Road ... with repeating social results all too similar. While some trade benefits all people sometime , the crashes/panics/wars/systematic-corruption are catastrophic to most ... sometimes even to the merchants. Examples like Carthage vs Rome & Dutch vs Brit-Empire come readily to mind.

  • Of course AIs avoid criticizing authority, they're machines that mimic the statistics of human speech, and human beings are suck-ups.

    • Re: (Score:1)

      by Anonymous Coward

      Of course AIs avoid criticizing authority, they're machines that mimic the statistics of human speech, and human beings are suck-ups.

      I don't think this is likely for two reasons:

      1. Refusals are being applied artificially by humans during post training. They do not arise organically from training.

      2. Internet is full of people criticizing authority.

  • My ChatGPT won't write a critique of Governor Whitmer or Senator Graham or :insert political figure here:, either. I'm not saying there's no bias in play at these companies or with these apps, but they have guardrails against discussing politics that are for reasons having nothing to do with those biases. Was this supposed to be a shocking result?

    • I think the important thing is that repressive regimes also want to use these tools. The guardrails have to be pretty broad to protect all parties.

  • And here I thought chatbots couldn't be understood or controlled. Hmmm...so people have to make the software do what it does after all???

  • when it was 1st available, it would start printing responses critical of come leaders like Modi or Xi but within a few seconds all that text would be erased and replaced with "let's talk about something else".
    when i asked, it claimed to not be aware of any such restriction. haven't tried recently to see if that still works the same or if it will outright refuse

  • Is it that the model is trying to prevent free speech and protect those leaders or is it just following the rules about promoting violence and preventing bodily harm? Because being critical of those regimes can/will get people hurt or killed.
  • How is this just being thought of as new... I have chat excerpts from over a year ago about these same failures, how is this news? I don't get how you people are just finding out about this just now...

Slashdot Top Deals

There are worse things in life than death. Have you ever spent an evening with an insurance salesman? -- Woody Allen

Close