A recent study indicates that language models are trained to make speculative guesses rather than acknowledge when they lack information.
OpenAI, the creator of ChatGPT, has commented on the ongoing issue where Artificial Intelligence models produce believable yet incorrect statements, which the company terms “hallucinations”.
On Friday, OpenAI released a statement clarifying that models are generally prompted to offer a guess, even if it’s improbable, instead of acknowledging an inability to provide an answer.
The company further noted that this problem originates from the fundamental principles guiding “standard training and evaluation procedures.”
OpenAI has disclosed that instances where language models “confidently generate an answer that isn’t true” continue to affect even the latest and most sophisticated iterations, including their flagship GPT‑5 system.
Findings from a recent study indicate that the issue stems from the current methods used to evaluate language models’ performance. Models that attempt to guess are often rated higher than those that carefully admit uncertainty. Under conventional protocols, AI systems learn that not generating an answer guarantees a zero score, whereas an unverified guess could potentially turn out to be accurate.
The statement concluded by asserting that “Fixing scoreboards can broaden adoption of hallucination-reduction techniques,” though it also conceded that “accuracy will never reach 100% because, regardless of model size, search and reasoning capabilities, some real-world questions are inherently unanswerable.”