AI chatbots shill for repressive regimes

Text

[10]New research from Meta's Oversight Board [11] found commercial large language models (LLMs) from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI are far less likely to criticise repressive leaders and governments. The research calls it "free speech infringements by proxy." Apps and material produced by these generative AI tools could actively reproduce restrictions, including on political criticism. It's also unclear why these AI models are doing this – underscoring the need for companies to be transparent and to build in human rights due diligence into the process, before it’s too late. China, Thailand, Cambodia and Saudi Arabia topped the list of countries with the highest refusal rates when LLMs were prompted to create materials criticising governments. (Photo: chinaworker.info [12])

New research from Meta's Oversight Board [11] found commercial large language models (LLMs) from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI are far less likely to criticise repressive leaders and governments. The research calls it "free speech infringements by proxy." Apps and material produced by these generative AI tools could actively reproduce restrictions, including on political criticism. It's also unclear why these AI models are doing this – underscoring the need for companies to be transparent and to build in human rights due diligence into the process, before it’s too late. China, Thailand, Cambodia and Saudi Arabia topped the list of countries with the highest refusal rates when LLMs were prompted to create materials criticising governments.

From The New Humanitarian [13], July 24