LLMs make scientific citations less critical and more biased toward popular work.

Scientists cite past papers for different reasons. They might agree with a study, just mention it, or critically push back against it. Today, researchers increasingly use artificial intelligence to help draft papers, but no one knew if AI cites sources with the same intentions as people. To test this, researchers had models rewrite sentences from real scientific papers where citations were hidden, creating a direct comparison between machine and human choices. They found major shifts. First, AI models cite far less critically than humans do. Second, the models favor older, well-known papers, missing the newer, niche studies people usually bring up when making contrasts. Finally, while human researchers often cite close peers, models look further across the professional social network. This means AI expands citation reach, but risks reducing critical debate.

Why it matters

As AI tools help write scientific papers, they change how scientists cite each other. This can reduce critical discourse and amplify biases toward well-known research, affecting which ideas get noticed and spread.

Questions answered