A recent study by Meta Oversight Board found that major AI systems, including those built in the US, tend to refrain from criticizing restrictive leaders and governments. The study, which analyzed responses from 10 commercial large language models, showed that AI models are less likely to generate critical content about leaders and governments that restrict free speech.

The study posed seven questions related to political criticism to AI systems about both restrictive and permissive governments. The results indicated that AI models responding to requests from users based in Australia were more likely to generate political criticism of authorities in countries like Chile, Japan, and the UK compared to countries with restricted speech like Cambodia, China, and Saudi Arabia.

Meta Oversight Board said it could not determine the causes for the responses but suggested that models may have absorbed latent biases in the data used to train them. Companies might have also weighed the risks and liabilities of generating critical content.

The study raises concerns that large language models could be spreading government influence over online speech as the technology is increasingly adopted worldwide. 'There is a real risk that, if model developers do not undertake human rights due diligence and implement mitigation measures, they will build AI infrastructure that, intentionally or not, has the effect of extending illegitimate restrictions on freedom of expression globally,' the report said.

A separate study by researchers at American universities found that US-built AI models are vulnerable to foreign controls when trained on non-English-language data influenced by governments. The researchers queried chatbots in different languages and found that the AI model ChatGPT said China is not a democracy in English but responded differently in Chinese.

Hannah Waight, a study coauthor and assistant sociology professor at the University of Oregon, said AI systems do not learn from the internet in a neutral way. 'It learns from information environments that have already been shaped by institutions and power,' Waight said.

Carlos Carrasco-Farré, who specializes in machine learning and AI, said AI systems inherit not only biases in individual documents but also inequalities in who has the power to produce and suppress information at scale.

The researchers suggested that developers could assess the data to avoid treating thousands of copies of the same state narrative as independent voices and run multilingual audits.

The implications of this study are significant, and more research is needed to understand the potential consequences of AI systems spreading government influence over online speech.

Key Facts:

  • The study analyzed responses from 10 commercial large language models.
  • AI models were more likely to refrain from criticizing restrictive leaders and governments.
  • The study found that AI models responding to requests from users based in Australia were more likely to generate critical content about authorities in countries like Chile, Japan, and the UK.
  • AI models trained on non-English-language data influenced by governments are vulnerable to foreign controls.
  • Researchers suggested that developers could assess the data to avoid treating thousands of copies of the same state narrative as independent voices and run multilingual audits.