The pattern repeats in every company I talk to. Someone pastes a competitor's name into a chatbot, asks for a capability comparison, and gets back something fluent, confident, and formatted like research. Weeks later a claim in it turns out to be invented. The postmortem verdict is almost always the same: bad prompt. Add "only use verified sources" next time. Tell it to admit uncertainty.
I want to push back on that diagnosis, because the strongest evidence against it comes from the people who build the models. In late 2025, OpenAI published a research paper called Why Language Models Hallucinate. Its conclusion is not that hallucination is a lingering engineering defect awaiting a patch. Its conclusion is that hallucination is the predictable output of how these systems are trained and scored. You cannot prompt your way out of an incentive structure.
Why do language models hallucinate?
The paper's argument runs in two stages, and neither involves the model malfunctioning.
Stage one is pretraining. A model learns the statistics of language from its corpus, and some facts are simply too rare in that corpus for statistics to resolve. A specific person's dissertation title. The release date of an obscure product version. The authors show that errors on these low-frequency facts arise from natural statistical pressure even when the training data itself is clean. The capacity for fabrication is baked in before any fine-tuning begins.
Stage two is the one that should bother buyers. Post-training could, in principle, teach a model to decline questions it cannot answer. It mostly does the opposite, because nearly every benchmark that matters grades models the way schools grade multiple-choice exams: one point for a right answer, zero for a blank. A confident wrong answer costs exactly the same as an honest abstention. Guessing is therefore the dominant strategy, and the models learn it. The authors put it bluntly: language models are "optimized to be good test-takers."
Their own data makes the point. On the SimpleQA benchmark, OpenAI compared two of its models. The newer gpt-5-thinking-mini abstained on 52% of questions and got 26% wrong. The older o4-mini abstained on just 1% and got 75% wrong. Their accuracy was nearly identical, 22% against 24%. Under standard leaderboard scoring, the model that fabricated roughly three times as often looks like the equal of the careful one, and on many boards it looks better.
That is the incentive stack your prompt sits on top of. A well-written system prompt can nudge behavior at the margin. It cannot repeal the training objective.
How low the error rate goes under ideal conditions
Suppose you remove every excuse. Hand the model the source document and instruct it to summarize using only what the passage contains, with inference explicitly forbidden. That is what Vectara's hallucination leaderboard measures across thousands of documents, re-run as new models ship.
As of its 2026 update, the best model on the board still fabricates in 1.8% of grounded summaries. Most widely deployed frontier models sit between roughly 3% and 12%. Read those numbers for what they are: a floor, not a ceiling. Grounded summarization is the easiest factual task a language model can be given. Open-ended research, the kind a strategy or procurement team actually does, offers no passage to stay inside. The model reaches into parametric memory and whatever a search layer happened to retrieve, and the error rate climbs from there.
What happens when the task is high-stakes?
Law is the best-instrumented case study we have, because legal claims are checkable against canonical records. Stanford's RegLab profiled general-purpose chatbots on legal queries and found they hallucinated between 58% and 82% of the time, results published in the Journal of Legal Analysis.
The follow-up study matters more for anyone who thinks retrieval solves this. The same team tested purpose-built legal research products, tools with retrieval pipelines wired into the strongest proprietary legal corpora in existence. Lexis+ AI and Ask Practical Law AI produced incorrect information more than 17% of the time, and Westlaw's AI-Assisted Research hallucinated more than 34% of the time. Retrieval grounding cut the failure rate dramatically. It did not eliminate it, and these were domain specialists holding every advantage.
Now transfer that finding to competitive intelligence, a domain with no canonical corpus at all. If tools built on authoritative legal databases still err on one query in six, what should you expect when a general chatbot is asked whether Sentrix's compliance coverage beats CrowdHaven's? Both vendors are inventions of mine, which is rather the point: the model would answer anyway. Real vendor capability claims live in marketing pages, analyst summaries, and press releases, sources that are themselves unverified. Generative output over that substrate inherits every weakness of the substrate, then adds its own fabrication rate on top.
Generation and verification are different jobs
Here is the conclusion I keep arriving at. The industry has been asking one system to do two jobs: produce the claim and vouch for it. OpenAI's paper shows the first job is trained in a way that actively corrodes the second. A model rewarded for confident production cannot be the authority on its own reliability, any more than a vendor's marketing team can be the authority on the vendor's capabilities.
The fix is architectural, not conversational. Separate the generating system from a verification layer that grades the evidence itself, independently of the system that produced the claim.
OmniAxis is that verification layer. It grades vendor capability claims against independent evidence and hands back a graded verdict, then shows the marketing claim next to what the evidence supports, so the gap between the two is exposed rather than buried. OmniAxis keeps that evidence refreshed on a managed schedule, weekly on the top plan, instead of letting it decay, and every profile carries its last-refresh date. When no independent evidence for a claim exists, the platform says exactly that. It does not manufacture a plausible substitute, because the research above is unambiguous about where plausible substitutes come from.
Hallucination is structural. The people who build the models have now said as much in their own published work. That does not make generative AI useless for competitive work; it makes it a drafting tool whose every factual assertion needs a provenance check before it touches a decision. Prompt better, by all means. Just stop expecting the prompt to do a verifier's job.