Transcription
Do AI models still hallucinate in 2026? Yeah, they do.
But using techniques like web search, RAG, and grounding techniques with the latest models, it can get down to 1% on some tasks.
Why don't they just make it say "I don't know"? Because the model doesn't know when it doesn't know something. You can fine-tune it to say "I don't know" more often, but then it starts refusing things it actually knows. There's no clear middle ground just yet.
Why don't they just train it on only correct information? Because even trained on perfect data, it can blend two real facts into one false one. It's not trying to lie. It's just making a prediction. And prediction doesn't care about truth.
Why don't they penalize it when it gets things wrong? They do. It's called reinforcement learning with human feedback. But the problem is that human reviewers can't always verify what's true. So the model learns to match human preferences, not the concrete truth.
Why don't they just make it bigger? Doesn't more data equal more accuracy? Kind of. But bigger models hallucinate too, just in more sophisticated, harder to catch ways.
Why don't they just have another AI check its work? Because the checker has the same problem. It can confidently approve a hallucination or hallucinate on top of a really good answer.
So, how do you actually fix it? You probably can't, not fully, not until the underlying training changes. So, for now, you can use an orchestrator, make some tool calls, add some guard rails, and it does a pretty good job, but it doesn't solve the problem. Follow, and I'll keep you posted.