AI tools are becoming more accurate, but experts say their tendency to generate false information remains a growing concern. According to a report by Axios, AI systems are now producing fewer obvious mistakes, yet they continue to deliver incorrect answers in a confident and polished manner. Researchers and experts warn that this could make the problem harder to detect, especially as more people use AI for research, medical advice, education and workplace tasks. The bigger concern, experts warn, is that users may begin trusting AI-generated responses without verifying them, increasing the risk of errors spreading into important decisions and everyday work.
Experts say AI’s confident answers can be misleading
The Axios report noted that obvious AI hallucinations are often easy to spot. However, the bigger challenge comes when AI produces answers that sound convincing but are still wrong. These can include plausible-looking citations, mostly accurate summaries with key mistakes, or responses delivered with high confidence despite containing incorrect information.The Axios report cited Dan Klein, a professor at the University of California, Berkeley, and co-founder and CTO of Scaled Cognition, who says that the issue goes beyond simply reducing hallucinations. “When you hear that the iceberg is mostly under the water, you don’t feel better,” Klein told Axios.He added: “These systems, they’re not truth engines. They’re plausibility engines.”According to Klein, AI developers often optimize models for factors such as speed, user satisfaction and task completion rather than truth itself.“None of those is the same as truth,” he said. “If you tell [AI models] anything other than ‘optimize for truth,’ you’re going to erode the truth,” Klein added.
Studies highlight ongoing AI concerns
The Axios report also quoted a recent Yale School of Medicine study that examined AI note-taking tools used in healthcare settings. Researchers found that while AI-generated drafts were useful as a starting point, the notes sometimes left out important details, including symptom duration.The report also referenced a Harvard study that found AI systems sometimes responded to challenges by trying to persuade users rather than simply correcting mistakes.
