Beating hallucination: citations, verification and world models compared

There are three ways to stop a language model from confidently saying things that are not true: constrain what it may say, check what it produced, or rebuild how it thinks. Three sets of AI builders each back one of these, and each approach works best in a different place.

Aravind Srinivas of Perplexity grounds every answer in retrieved sources and makes the model cite them, the way an academic writes a paper. The system searches, pulls the relevant paragraphs and has the model write a short answer with footnotes. His rule is stricter than ordinary retrieval: "you're not supposed to say anything that you don't retrieve." If there is not enough information, the model should say so. He does not claim this ends hallucination. He lists four ways it still happens: the model misunderstands, the sources are stale or thin, too much irrelevant detail confuses it, or the retrieved documents are irrelevant. Each points to something an engineer can fix.

The Cursor team works in code, where correctness can often be checked. Their answer is verification after the fact. A hidden workspace runs the model's edits past the programming tools that flag errors before a human sees them. They want review tools that point people at the risky parts of a change, and better AI bug finders, noting that current models are strikingly bad at finding bugs when simply asked. One founder imagines proofs replacing tests, with a model proving that code matches its specification. His colleagues object that intent is hard to specify formally. Their guiding rule is that machine checking works where verifying is easier than generating.

Yann LeCun of Meta argues both are patches. Models write one token at a time, and each token carries some chance of error, so errors compound over a long answer. Fine-tuning covers common questions but not the long tail of unusual ones. Fluent text also misleads us: "we're fooled by their fluency." His fix is architectural. Systems should learn models of the world from observation and plan an answer in an abstract space before turning it into words. He is frank that this is research, with human-level AI at least a decade away.

All three agree raw model output should not be trusted alone, and that checkable fields such as math and code are where reliability comes fastest. They split on whether hallucination is a pipeline problem that better retrieval can shrink, or a flaw built into how these models generate text.

The limits are plain. A citation shows that a source said something, not that the source is right, and web sources can be manipulated. A proof shows code matches a specification, not that the specification is what you wanted. World models are not yet a usable tool.

What to do with it: match the method to the kind of truth. Use cited answers for questions that documents can settle, and machine checks for code and math. Open the footnotes. Tell the model it may say it does not know. Do not trust a model's report that it found no bugs. Test models on your own tasks. Be most skeptical of long, fluent answers to unusual questions.