London, United Kingdom – A recent study has revealed a phenomenon it describes as “shared hallucinations” between humans and artificial intelligence systems, indicating that long and continuous conversations may in some cases lead to the construction of inaccurate information or false conclusions adopted by both parties within the context of the dialogue, despite lacking a factual basis.
Hallucinations shared between humans and the system
The researchers explained that this phenomenon does not indicate a mental disorder,
but rather a situation where the user and the AI system exchange incorrect information or assumptions,
which they then continue to develop during the conversation until they appear as established facts.
The study added that AI models may sometimes produce answers that seem convincing and logical,
but contain errors or fabricated information—a phenomenon known in AI as “hallucination.”
When the user interacts with these answers without verifying them, they may inadvertently contribute
to reinforcing and completing them, creating a state of “shared hallucination” between the user and the system.
Reducing the risk of spreading misinformation
The researchers pointed out that longer conversations increase the likelihood of errors accumulating,
especially if the dialogue relies on undocumented information or unverified assumptions.
They emphasized that excessive trust in AI outputs can lead to
the entrenchment of inaccurate concepts among some users.
The study affirmed that these findings do not diminish the importance of AI technologies,
but rather highlight the necessity of using them as aids,
not as replacements for critical thinking or reliable sources,
particularly in scientific, medical, legal, and historical subjects.
The study’s authors called for the development of more efficient mechanisms
to mitigate the phenomenon of “hallucination” within AI models,
along with raising user awareness about the importance of reviewing and verifying information before relying on it.
This would ensure that users benefit from the capabilities of
these technologies while minimizing the risk of spreading misinformation.
