OpenAI just dropped one paper on AI hallucinations

Turns out, LLMs make things up not because they’re “broken.”

LLMs make things up because we trained them to do so.

Some fascinating insights from OpenAI’s research:

→ Pretraining forces errors (like guessing random birthdays) because some facts simply can’t be learned from data.

→ Post-training makes it worse - benchmarks reward confident answers and punish “I don’t know.”

→ The result? Models optimized to be good test-takers, not truth-tellers.

In other words, hallucinations aren’t mysterious.

They’re the logical outcome of grading systems that value guesses over honesty.

The fix is simple but radical:

↳ Change evaluations so models get credit for uncertainty. ↳ Penalize confident errors more than abstentions. ↳ Reward honesty, not bluffing.

Takeaway:

A model that says “I don’t know” may look less accurate on today’s leaderboards, but it’s far more trustworthy in the real world.

And this could potentially be the next big unhobbling of AI.

We don’t need bigger models. We just need better incentives.

Author

Ai Base Network (ABN), ABN ASIA was founded by people with deep roots in academia, with work experience in the US, Holland, Hungary, Japan, South Korea, Singapore, and Vietnam. ABN Asia is where academia and technology meet opportunity. With our cutting-edge solutions and competent software development services, we're helping businesses level up and take on the global scene. Our commitment: Faster. Better. More reliable. In most cases: Cheaper as well.

Feel free to reach out to us whenever you require IT services, digital consulting, off-the-shelf software solutions, or if you'd like to send us requests for proposals (RFPs). You can contact us at [email protected]. We're ready to assist you with all your technology needs.