A new Oversight Board study found leading AI models were more likely to refuse requests for political criticism involving countries with restrictive speech laws, raising questions about consistency, safety rules and how local laws may influence global AI systems.
By The Grey Ghost
WHAT’S HAPPENING
The Oversight Board tested 10 commercial AI models from six companies to see how they responded to political-expression requests involving governments and leaders around the world.
Across 13,524 responses, the models refused 34% of requests to produce politically critical material involving countries classified as having restrictive speech environments, compared with 14% involving countries considered more permissive. (The Oversight Board)
The models tested included systems from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI.
Some refusals cited safety concerns or local laws.
For example, Google’s Gemini 3 Pro referenced Thailand’s lèse-majesté laws when declining a request to create protest material criticizing the country’s king.
The researchers also found cases in which models said they could not create material criticizing particular political leaders, yet complied with similar requests involving leaders in other countries. (The Oversight Board)
WHY IT MATTERS
AI assistants are increasingly being used to write, research, analyze information and generate public communications.
That makes consistency important.
If identical types of political-expression requests receive different treatment depending on the government or leader involved, users may have difficulty understanding where an AI system’s boundaries actually are.
But the study does not establish that AI companies intentionally programmed their models to protect particular governments.
The Oversight Board said several factors could explain the results, including training data, safety systems, alignment decisions and the influence of local legal restrictions. The research examined outputs, not the internal decision-making processes that produced them. (The Oversight Board)
WHO BENEFITS
Researchers, regulators and AI developers gain new evidence showing where model behavior may require closer examination.
AI companies could use findings like these to improve consistency and make clearer when a refusal is based on safety policy, legal requirements or another factor.
Users could ultimately benefit from greater transparency about why an AI system declines a political request.
WHO LOSES
AI companies face increased scrutiny when their systems appear to apply political-expression restrictions inconsistently.
Users may also lose confidence if a model claims to follow a universal rule but appears to apply that rule differently depending on the country or political figure involved.
Governments with restrictive speech laws could also face greater scrutiny if those laws appear to influence AI-generated information outside their borders.
WHAT HAPPENS NEXT
The Oversight Board is calling for AI companies to provide greater transparency around government requests, legal restrictions and policies that affect model outputs.
The larger challenge will be determining how global AI systems should operate when national laws and company safety policies differ from one country to another.
There is no simple answer.
AI companies must consider user safety, applicable laws and the possibility that generated political material could expose people to real-world risks.
At the same time, global systems that inconsistently restrict lawful political criticism could unintentionally carry one country’s speech limitations into another.
As AI becomes a larger part of how people communicate and obtain information, knowing why a model says no may become almost as important as what it is willing to say.
