Hallucinations cease to be the main threat of generative AI

From fabricating facts to correcting them

Wharton professor Ethan Mollick points out a fundamental shift in the output quality of premium AI models. While previously every reference had to be suspected of hallucination, the situation is reversing today. Mollick describes his experience proofreading his new book, where AI found a factual error not in his text, but directly in the original source he cited.

Cheap models perpetuate the myth of unreliability

The perception of AI as an unreliable text generator remains strong primarily because most users interact with cheaper or older models. The gap between the free tier and the latest paid versions is widening, leading to a split understanding of what the technology is currently capable of.

Who will watch the watchmen

The ability of models to correct human sources raises the question of blind trust. If AI builds a reputation as an infallible proofreader, there is a risk that users will stop critically evaluating its interventions. The result could be replacing occasional hallucinations with a deeper, systemic bias that a human will no longer be able to detect without the help of another AI.

Adoption in critical processes will decide

Whether the era of hallucinations is truly over will be shown by the willingness of companies to deploy LLMs in processes where an error means financial or legal penalties. The real test of reliability will not be academic benchmarks, but deployment in audit and compliance.

Lilith's verdict

The gap between free models and the twenty-dollar ones is widening. Anyone judging the state of AI by the free tier is looking at history.