Stanford's 2026 AI Index puts AI hallucination rates between 22% and 94% depending on the task. UnitedHealth is in federal court because their AI denied claims. State insurance regulators in 24 states are auditing AI governance this year. Carriers are now writing AI failure out of liability policies.
AI was never meant to tell you the truth. But you expected it to. Most of us did. Its confident lying can be dangerous. As that realization becomes clearer, controlling what AI is allowed to say will move up every budget and roadmap.
So far the industry response has been to retrofit guardrails onto systems that were never designed to produce verifiable facts. Retrieval bolted on. Confidence scores bolted on. Detectors bolted on after that. None of it fully works, because underneath all of it the model is still guessing the next likely word, and it has to respond with something.
The evidence is not ambiguous. In the first preregistered evaluation of premium retrieval-based legal research tools, Stanford RegLab measured hallucination rates of roughly 17% for Lexis+ AI and 33% for Westlaw AI-Assisted Research, against vendor marketing that had promised hallucination-free citations. Retrieval helped. It did not solve anything.
So I built the thing the retrofits are working around. I call it Gated Truth Architecture, and the idea is blunt: if a system cannot compose an answer at the moment someone asks, it cannot invent one.
An AI reads your documents once, at upload, and drafts candidate question-and-answer pairs from them. A person reviews each one and approves it, edits it, or throws it out. Only approved answers can ever be shown to anyone. When a visitor asks a question, nothing is generated: the system matches the question to an approved answer and returns it word for word, with the exact document, section, and passage it came from. If nothing matches, it says so and logs the question instead of filling the silence.
Human in the loop. Not human after the loop.
Truebe is the live implementation. Upload a document, ask a question from it, see the cited answer, click the citation. Try the demo.
The architecture came out of building Foodbe, a food intelligence platform running on 10,000+ hand-curated, citation-locked fact seeds. Foodbe proved the discipline at scale. Truebe productizes it for anyone with documents and something to lose.
The pattern is vertical-agnostic. The same six rules work for law firms, medical and dental practices, insurance, financial advisory, pharmaceutical patient education, and any business whose website answers questions in public. Different documents. Same discipline.
What changes per vertical is not the architecture. It is the library. The gate is the engineering. The approved answers are the moat.
A gated system will refuse questions a generative one would happily answer. That is the deal, and it is the whole deal. You give up the illusion of universal coverage and you get a system that cannot say something you did not approve. Every refusal is captured, reviewed, and turned into an approved answer, so the library grows around what people actually ask instead of what a vendor guessed they would.
If a wrong answer costs you nothing, you do not need this. If a wrong answer costs you a client, a license, or a lawsuit, no amount of tuning gets a generative system to zero.
The current AI race is a race toward cheaper generation. More tokens, more parameters, more inference, more output. The metric that matters in that race is how much the model can produce.
That metric is wrong for any regulated industry. The metric that matters in healthcare, finance, insurance, and law is how much a system can produce that is verifiably true and provably approved by a person.
An AI hallucination is when a language model generates false, fabricated, or misleading information with complete confidence, presenting invented facts, fake citations, or nonexistent sources as if they were real. The model does not know it is wrong. It is not lying intentionally. It is pattern-matching to what a correct-sounding answer looks like, with no mechanism to verify whether the output is true.
Gated Truth Architecture is a design in which an AI drafts candidate answers from your documents at upload time, a human reviews and approves each one, and only approved answers can ever be served. Nothing is generated at the moment a question is asked, so there is no generative step in which fabrication can occur. The gate is structural: it is not a prompt asking a model to behave, it is the absence of live generation.
No. Retrieval augmented generation hands a model relevant passages before it writes, which meaningfully reduces fabrication. But the model still composes the final answer live, and it can still blend retrieved text with training data or misread the passage it was given. Stanford RegLab measured roughly 17% to 33% residual hallucination in the best-funded RAG products ever tested. A citation attached to a generated answer tells you what the model read, not that the answer is what the source says.
Google's AI Overview told users to put glue on pizza and eat rocks. A lawyer submitted fabricated ChatGPT case citations to federal court and was sanctioned, and more than 1,780 court decisions worldwide have now addressed AI-hallucinated filings. UnitedHealth's AI denied insurance claims using criteria that did not exist in the patient's policy, now the subject of federal litigation. A German court ruled Google directly liable for false AI-generated answers about a person. These are not edge cases. They are the predictable output of systems built to generate rather than verify.
An AI hallucination checker reviews AI output after generation and flags claims it believes are unsupported. Checkers are useful but fundamentally reactive: the wrong answer already exists by the time anything evaluates it, and the checker is itself a probabilistic system with its own miss rate. Gated Truth Architecture moves the checkpoint before publication. There is nothing to check afterward, because nothing unapproved was ever published.
Related reading
If your team ships AI where wrong answers have consequences, legal, health, insurance, finance, reach out.