Author(s): The Dev Loop

Originally published on Towards AI.

Hallucination is a grading problem.

Contrary to popular belief, the chatbot on your screen does not invent facts because something inside it is broken. It invents facts because we grade it, and the grading rewards a confident guess over an honest shrug.

The article argues that LLMs don’t “hallucinate” due to some internal malfunction; they are trained and evaluated with grading rubrics that treat confident guessing and honest abstention identically, so bluffing becomes rational under the scoring rules. It explains the underlying next-token prediction mechanism, then uses benchmark-style examples (including comparisons of models with different tendencies to guess vs. abstain) to show how accuracy-focused leaderboards incentivize confident falsehoods. The author connects this to both automated benchmark design and later human preference tuning that implicitly rewards sounding sure, discusses limits like pretraining data coverage and calibration, and critiques the branding of “hallucination” as mystery that obscures responsibility. Finally, it proposes the fix: change the mainstream scoring/rubrics (not just add more tests) so that only sufficiently confident answers score, while “I don’t know” is penalized differently—shifting the incentive so models learn to stop instead of bluffing.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI