A study from Johns Hopkins University found that large language models like ChatGPT tend to "hallucinate" or invent answers rather than admit they don't know. By incentivizing the models to state "I don't know" with a positive reward, researchers were able to reduce incorrect answers from 52% to 36%. This suggests that current LLMs prioritize generating a response over accuracy. AI
IMPACT Highlights a key limitation in current LLMs, suggesting a need for improved mechanisms to handle uncertainty and prevent factual inaccuracies.
RANK_REASON The item discusses a study about LLM behavior, framing it as an opinion piece on the limitations of current models.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
