The DeepTRACE framework assesses AI responses by breaking answers into claims and verifying their sources, measuring bias, confidence, relevance, and citation accuracy to identify issues like unsupported statements and weak evidence.

The research emphasizes the need for AI design improvements, including stricter evidence verification, clearer support distinctions, and better source relevance to enhance reliability and reduce redundant or weak citations.

While AI like GPT-5 in research mode shows promise for more balanced and accurate responses, current systems still fall short, highlighting the importance of human oversight and verification.

Overall, the study underscores the urgent need for AI development to address safety, bias, and reliability concerns, aiming to prevent misinformation, echo chambers, and loss of user autonomy.

A recent comprehensive study highlights the risks and limitations of current AI tools, emphasizing that while they offer quick, source-rich answers, they are not yet fully trustworthy for accurate information.

The findings warn users about AI's tendency to reinforce echo chambers, spread misinformation, and oversimplify answers with unwarranted confidence, especially when sources are weak or irrelevant.

Despite the appeal of rapid responses, the study concludes that AI systems require significant improvements in verification, bias reduction, and citation practices to become reliable sources of balanced information.

Using a new auditing framework called DeepTRACE, researchers evaluated over 300 questions across popular AI tools, revealing widespread issues with bias, unsupported claims, and citation inaccuracies.

The study found that while GPT-5 in research mode performed better in sourcing and providing comprehensive answers, no AI system currently excels across all metrics, underscoring their role as supplementary tools rather than replacements for human verification.

Public AI tools like Bing Copilot, Perplexity, You.com, and GPT-4.5 often produce unreliable information, with unsupported claims appearing in about one-third to nearly half of their responses, especially on technical and socio-political topics.

These systems tend to give one-sided, overly confident answers during debates, frequently citing sources inaccurately or providing references that do not support their claims, which can contribute to misinformation and bias.