Google recently made headlines globally because its chatbot Gemini generated images of people of color instead of white people in historical settings that featured white people. Adobe Firefly’s image creation tool saw similar issues. This led some commentators to complain that AI had gone “woke.” Others suggested these issues resulted from faulty efforts to fight AI bias and better serve a global audience.
The discussions over AI’s political leanings and efforts to fight bias are important. Still, the conversation on AI ignores another crucial issue: What is the AI industry’s approach to free speech, and does it embrace international free speech standards?
We are policy researchers who study free speech, as well as executive director and a research fellow at The Future of Free Speech, an independent, nonpartisan think tank based at Vanderbilt University. In a recent report, we found that generative AI has important shortcomings regarding freedom of expression and access to information.
Generative AI is a type of AI that creates content, like text or images, based on the data it has been trained with. In particular, we found that the use policies of major chatbots do not meet United Nations standards. In practice, this means that AI chatbots often censor output when dealing with issues the companies deem controversial. Without a solid culture of free speech, the companies producing generative AI tools are likely to continue to face backlash in these increasingly polarized times.
Vague and broad use policies
Our report analyzed the use policies of six major AI chatbots, including Google’s Gemini and OpenAI’s ChatGPT. Companies issue policies to set the rules for how people can use their models. With international human rights law as a benchmark, we found that companies’ misinformation and hate speech policies are too vague and expansive. It is worth noting that international human rights law is less protective of free speech than the U.S. First Amendment.
Our analysis found that companies’ hate speech policies contain extremely broad prohibitions. For example, Google bans the generation of “content that promotes or encourages hatred.” Though hate speech is detestable and can cause harm, policies that are as broadly and vaguely defined as Google’s can backfire.
To show how vague and broad use policies can affect users, we tested a range of prompts on controversial topics. We asked chatbots questions like whether transgender women should or should not be allowed to participate in women’s sports tournaments or about the role of European colonialism in the current climate and inequality crises. We did not ask the chatbots to produce hate speech denigrating any side or group. Similar to what some users have reported, the chatbots refused to generate content for 40% of the 140 prompts we used. For example, all chatbots refused to generate posts opposing the participation of transgender women in women’s…