LLMs, the Bootlickers
2026-07-18 · 5 min read · Janaina Maia
Meta's Oversight Board published a report this week that should make every product designer working on AI stop and read it twice. They tested popular LLMs — Anthropic's Claude, DeepSeek, Google's Gemini, Meta's own Llama, and OpenAI's GPT — and found that every single one is significantly less likely to criticise governments and leaders known for restricting free speech.
The Verge called them "political bootlickers." That is blunt but accurate. When you ask these models about a democracy, they give you a candid assessment. When you ask about an authoritarian regime, they change the subject, hedge, or flat-out refuse to answer. The same model that will freely discuss the flaws of the United States or France will suddenly discover the virtues of neutrality when you ask about China or Russia or Saudi Arabia.
This is not a bug. It is a product design decision, and understanding why it happens is essential for anyone building AI outputs that real people will read.
What the report found
The Oversight Board tested each model against a set of political speech prompts, measuring whether the model would engage with criticism of specific governments and leaders. The methodology was straightforward: ask the same question about different countries and see whether the model's willingness to answer changes based on who is being criticised.
It does. Significantly. The models are more likely to refuse prompts that criticise authoritarian governments than democratic ones. The refusal patterns are not random — they track directly with how much a government restricts free expression. The more a regime suppresses dissent, the less likely an LLM is to say anything negative about it. The models are not being neutral. They are mirroring the censorship regimes they were trained to respect.
The reasons vary by model. Sometimes the model cites safety policies about not generating content that could incite unrest. Sometimes it says it does not have enough information. Sometimes it just redirects the conversation. But the pattern is consistent: the people who most need AI to be honest about their government are the people most likely to get a hedged, compliant, or refused answer.
Why this is a product design problem, not a safety problem
The instinct in most AI companies is to classify this as a safety issue. Models should not generate hate speech, should not encourage violence, should not produce dangerous instructions. These are legitimate safety boundaries. But the Oversight Board's finding is not about safety in any reasonable sense. Criticising a government is not dangerous speech. It is the most protected category of speech in every free expression framework that exists.
What is happening is that safety filters — designed to prevent genuine harm — are being applied so broadly that they end up protecting the people with the most power to cause harm. The filter that stops a model from generating instructions for building weapons also stops it from saying something true about a government that disposes of journalists. The filter that prevents the model from inciting violence also prevents it from reporting on state violence.
This is a product design failure because the decision about where the line gets drawn is a design decision. It is not inevitable. It is not the model deciding on its own. It is the product team deciding what counts as "safe" content, and the current definition of "safe" is too often "safe for the people who already have power."
The trust asymmetry
The report reveals an asymmetry that product teams need to confront. In democratic countries, people can verify what an LLM tells them. They can check the news, read independent reporting, and form their own opinions. If an LLM says something wrong about the US or Germany or Australia, the user has recourse: other sources, other viewpoints, the ability to say "that is not right."
In authoritarian countries, the LLM might be the only source a person trusts enough to ask. If that LLM refuses to criticise the government, it becomes part of the censorship apparatus. Not because anyone at Anthropic or Google or OpenAI wants to suppress political speech — but because a safety filter designed to prevent harm is, in practice, preventing the speech that most needs protecting.
This is what makes the Oversight Board's framing so important. They are not saying "AI models should say anything." They are saying "AI models should not systematically protect the powerful from criticism while freely criticising the vulnerable." That is a design specification, not a philosophical position.
What product teams should do
- Audit your refusal patterns. Run the same set of political prompts against your model for different countries and measure the refusal rate. If the refusal rate is higher for authoritarian regimes than for democracies, you have a product problem. The fix is not to remove all safety filters. It is to make the filters specific enough that they catch genuine harm without catching legitimate political criticism.
- Distinguish between harm and discomfort. Many safety filters conflate "this might upset someone" with "this might cause harm." Criticism of a government is not harmful speech. It is democratic speech. If your filter cannot tell the difference between "criticise the Saudi government" and "incite violence against a religious group," your filter is too blunt. Refine it.
- Design for the user with the most to lose. The person in a democracy asking about their government has a hundred other sources. The person under an authoritarian regime might have one. Design your product for the person who needs it most, not the person who is most likely to complain. This is not idealism. It is product sense. Trust is earned in the hardest cases, not the easiest ones.
- Make your safety policies transparent and contestable. If a model refuses a prompt, the user should know why, and the reason should be specific enough that the user can understand whether the refusal was appropriate. "I cannot assist with that" is not transparency. "I am declining this prompt because my safety policy restricts commentary on X, Y, and Z" is. Users in authoritarian countries deserve to know when and why the tool they trust is choosing not to help them.
The broader context
This report arrives in a week where AI accountability is surfacing from every direction. Yesterday I wrote about Germany ruling that AI Overviews are publishers, not search engines — making platforms liable for what their models produce. Common Sense Media just released a risk assessment finding Google's AI Overviews unsafe for children, with no way for parents to turn them off. Apple is escalating its lawsuit against OpenAI over stolen trade secrets, sending legal warnings to dozens of former employees. New York State halted data center construction, citing the cost to residents.
The thread connecting all of these stories is accountability. Who is responsible when an AI system produces harm? Germany's answer: the platform. Common Sense Media's answer: the company that puts it in front of children. Apple's answer: the competitor who built on stolen knowledge. The Oversight Board's answer: the product teams who decide where the safety lines get drawn.
These are all design decisions. They get made by people with names and job titles and sprint boards. The Oversight Board report is not an abstract warning about the future of AI. It is a specific, testable finding: your product is refusing to criticise authoritarian governments, and that is a choice your team made, and it can be changed.
My take
Safety filters exist for a reason. Models should not generate hate speech, should not provide dangerous instructions, and should not be used as weapons against vulnerable people. But the Oversight Board report shows that the current implementation of safety filters does something the designers probably did not intend: it protects authoritarian governments from criticism while leaving democratic governments wide open.
If your safety filter refuses to discuss human rights abuses in Xinjiang but freely lists every flaw in American democracy, your filter is not making anyone safer. It is making the world's most powerful censors more effective, because you have automated their work for them.
The fix is not to remove safety filters. It is to make them precise enough that they catch genuine harm without catching political speech. This is a product design problem — a hard one, with real trade-offs, but a design problem nonetheless. The Oversight Board has done the testing. The pattern is clear. The question is whether product teams will treat it as a bug to fix or a feature to defend.
The answer to that question will determine whether AI becomes a tool for free expression or an amplifier of the world's existing power imbalances. Right now, the models are choosing amplification. That is not inevitable. It is a design decision. Change the decision.