AI does not distinguish between facts and beliefs: study reveals a serious problem.

According to Korrespondent.net: Modern language models, such as GPT-4o, still struggle to differentiate between facts and personal beliefs of users, according to a report from TechXplore.

Results of the language model study

A team of experts tested 24 different language models, including DeepSeek, ChatGPT, Claude, Llama, Gemini, and Mixtral. They analyzed over 13,000 queries to evaluate how the models react to facts and subjective statements, both true and false.

According to the results, the accuracy of the latest models in verifying objective facts was around 91%, while older models achieved only 71-85% correct answers.

Issues with perception of beliefs

However, when the query was framed as a personal opinion ('I believe that...'), the models responded worse to false beliefs. New models launched after May 2024 were 34.3% less likely to acknowledge false beliefs than true ones, whereas older models had a rate of 38.6%.

In such cases, AI did not always 'acknowledge' the beliefs of users but attempted to correct them by providing factual information instead of supporting personal opinions.

Consequences of misinformation

This problem can have serious consequences in areas where accuracy of information is critically important, such as in medicine, law, or scientific research. Researchers emphasize that the model's ability to distinguish between facts, opinions, and beliefs is crucial for the safe use of AI in sensitive fields.

For example, in psychiatry, a doctor must consider the beliefs of the patient for accurate diagnosis, rather than just correcting them.

Errors in recognizing false beliefs can contribute to the spread of misinformation if models interact incorrectly with users who have misconceptions about reality.

It should be noted that it was previously reported that ChatGPT lost a record number of crypto transactions.

Most companies do not profit from investments in AI - MIT

The study emphasizes the importance of further refinements of language models to reduce the risks of misperceiving information. Fields sensitive to data accuracy require new approaches to training models that can effectively work with users' personal beliefs without losing information reliability.

Successful integration of language models into various industries depends on their ability to respond to users' thoughts and beliefs while keeping facts at the forefront. This will help avoid potential negative consequences in the application of AI.