A team from the Massachusetts Institute of Technology (MIT) has challenged earlier research that said artificial intelligence (AI) can form its values.
On Wednesday, tech news site TechCrunch reported that MIT researchers analyzed top models from Meta, Google, Mistral AI, OpenAI, and Anthropic. They aimed to test whether AI holds firm views and values, and if those can shift.
The study found that none of the models showed consistent preferences. Their views changed based on how users phrased prompts. Based on these findings, the team concluded that AI only responds differently depending on prompts and lacks a consistent value system. They called it a purely imitative system.
Steven Casper, a doctoral researcher at MIT, said AI models may show preferences that match certain principles in specific cases, but it’s misleading to say they hold general values or opinions based on limited tests.
Mike Cook, an AI researcher at King’s College London, agreed with the MIT study’s conclusions. He said people project meaning onto AI when they claim it resists value changes. He added that the difference between goal optimization and true values in AI is only semantic.