Hallucination
In one sentence
A hallucination is when an AI states something false with the same confidence it states something true, which is not a lie, because there is no intent, and not a bug, because the system is working as designed.
Not to be confused with Grounding, or Chunking.
Definition
A hallucination is when an AI says something false in exactly the same confident tone it uses for something true.
It is not a lie, because nothing intends it. It is not a bug either, because the system is doing what it was built to do.
Language models generate plausible continuations. Plausibility and truth correlate strongly in training data, which is why models are useful at all, but they are not the same property, and where the two diverge the model follows plausibility. Confidence in the output is not a signal of accuracy, and that is the part users find hardest to internalize.
Five types worth distinguishing, because they have different remedies
- Fabrication. Inventing facts, figures, citations or policies wholesale. Remedy: retrieval grounding.
- Extrapolation. Filling gaps in the retrieved context with plausible additions. Remedy: an explicit instruction to answer only from context, plus verification.
- Conflation. Merging two sources into a single wrong statement, common when a corpus carries regional or tier variations. Remedy: better chunking and reranking.
- Faithful reproduction of a wrong source. The model correctly reports an outdated document. Remedy: corpus hygiene, not model work. This one is frequently misclassified as hallucination when it is a content management failure.
- Instruction drift. Over a long conversation the model loosens its adherence to its original constraints. Remedy: constraint reinforcement and session limits.
The mitigation stack, in rough order of effectiveness
- Retrieval grounding, so answers derive from real passages.
- Retrieval confidence thresholds, so the system abstains rather than generating when retrieval is weak.
- An explicit abstention instruction, so "that is not in my material" is an output the system is allowed to produce.
- Post-generation verification against the retrieved passages.
- Scope guardrails that block whole topic classes regardless of what retrieval returned.
- Corpus hygiene, keeping the source documents current.
- None of these eliminates the problem. Together they reduce it substantially and make the residue traceable.
Common misconception
That hallucination is being solved and will shortly disappear. It is being reduced. It is a property of how generative models work rather than a defect waiting on a patch, and any deployment plan that assumes elimination is resting on an assumption that does not hold.
Why it matters commercially
A voice agent on a company website speaks as the company. A hallucinated price, warranty term, eligibility rule or medical statement is one the business may then have to honor or defend. It is the single most common objection raised against deploying voice agents, and it deserves a serious answer rather than reassurance.
In voice specifically
Voice makes hallucination worse in four specific ways. There is no visible citation, because a text interface can show a source link unobtrusively and speech cannot, so the listener has no cheap way to check. Spoken delivery carries authority, and a confident voice is more persuasive than confident text, which means voice hallucinations are more likely to be believed and acted upon. Transcription error opens a pathway that does not exist in text at all: the model can faithfully answer a question the visitor never asked, and the answer will be internally consistent and completely irrelevant. And speech-to-speech architectures leave no transcript by default, which removes the audit trail anyone would need to establish what happened.
Where AsqVox fits
True RAG over uploaded documents, combined with voice navigation that shows the visitor the source of what they were just told, is the relevant mitigation. Showing the passage is a partial answer to the missing-citation problem that voice interfaces otherwise have.
Visual
Five different failures, one symptom
Confidence is not a signal of accuracy. It never was.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
No reliable, transferable industry figure exists for hallucination rate in deployed business RAG systems. Academic benchmarks exist but do not map cleanly onto business corpora, and vendor figures are measured on vendor data.
-no reliable figureThis needs to lead. Any single percentage claim should be treated as unverifiable, and stating that plainly is more credible than borrowing a number from somebody else's benchmark.
Whisper Large-v3 achieves around 2.7 percent word error rate on clean benchmark audio.
2.7%independentWhisper Large-v3 published benchmark, 2026 - The clean-audio figure, published for the model itself. It is the number quoted in marketing and the one least like production.
The same model lands at roughly 8 to 12 percent word error rate on real-world English.
8 to 12%industry rangeReported real-world word error rate for Whisper Large-v3, 2026 - This is the hallucination pathway unique to voice. A misheard question can be answered flawlessly and irrelevantly, and nothing in the answer will look wrong.
Deepgram Nova-3 reports around 5.26 to 6.84 percent word error rate in production.
5.26 to 6.84%vendor claimDeepgram Nova-3 reported production figures, 2026 - Reported by the vendor on its own traffic. Useful as a second data point on the same pathway rather than as a comparison against the Whisper figures, which were measured differently.
The 2026 Coval and Gradium independent benchmark found a speed against accuracy tradeoff across recognition providers, with no single leader.
No single leaderindependentCoval and Gradium independent benchmark, 2026 - The useful conclusion is that upstream error rate is not a solved variable you can design around. Choosing a faster recognizer moves the hallucination pathway rather than closing it.
EU AI Act Article 50 transparency obligations, including disclosure requirements relevant to conversational and synthetic systems, become enforceable on 2 August 2026.
2 August 2026industry rangeEU AI Act, Article 50, 2026 - Disclosure does not address accuracy. It changes what a person is entitled to know about who they are speaking to, which is a different obligation and should not be presented as a hallucination control.
Examples
In practice
An agent is reported to hallucinate delivery timeframes. Review of the transcripts shows it faithfully reporting an uploaded logistics document that was superseded four months earlier and never replaced. The model behaved correctly throughout. The failure was content management. The team adds document review dates and surfaces stale-document warnings in the dashboard, which is a workflow fix rather than an AI one.
The everyday version
Hallucination is when the system invents something and delivers it in the very same assured tone it uses for the things it genuinely knows. It never sounds hesitant, because it is not hesitant. That is the reason grounding it in your own documents matters, and the reason it has to be able to admit when it does not know.
Usage
Who says it
- Everyone, which is part of the problem. It has become a catch-all for any AI error, including errors that are not hallucinations.
- AI practitioners use it more narrowly, for generated content that no source or training data supports.
- Buyers raise it as their primary objection, usually phrased as what if it says something wrong.
Where it turns up
- Next to accuracy commitments, grounding and source attribution, abstention behavior, audit logging, guardrails and escalation triggers.
- Enterprise buyers increasingly ask for abstention to be demonstrated live. Being asked to make the agent fail on purpose is a good sign about the buyer.
Common misuse
- Using hallucination for every AI mistake. A stale source faithfully reported is a content problem. A misheard question is a recognition problem. Both get labeled hallucination, and both then get the wrong remedy.
- Claiming a system does not hallucinate. Nothing currently available can support that claim, and making it destroys technical credibility.
- Treating grounding as elimination rather than as reduction and traceability.
Questions people ask
Does RAG stop hallucinations?
It reduces them and makes the residue traceable. It does not eliminate them. Retrieval grounding addresses fabrication, but a model can still extrapolate past the passage it was given, conflate two similar passages into one wrong figure, or faithfully report a document that is out of date. Treating grounding as elimination rather than reduction is one of the most common mistakes in the category.
How often do AI voice agents hallucinate?
There is no reliable, transferable figure for deployed business systems. Academic benchmarks exist but do not map onto business corpora, and vendor numbers are measured on vendor data, so any single percentage should be treated as unverifiable. What can be stated is that speech recognition adds a pathway text does not have: real-world word error rates of roughly 8 to 12 percent mean the agent can flawlessly answer a question the visitor never asked.
Why is hallucination worse in voice than in text?
Four reasons. There is no visible citation, because speech cannot show a source link the way a text interface can. A confident voice is more persuasive than confident text, so a false statement is more likely to be believed and acted upon. Transcription error creates a failure mode text does not have. And speech-to-speech architectures leave no transcript by default, which removes the audit trail.
What is the difference between a hallucination and a stale document?
A hallucination is generated content that no source supports. A stale document faithfully reported is the model doing its job correctly over material the business never updated. To a listener they are identical, and they need opposite remedies: one is model and retrieval work, the other is corpus hygiene. Misclassifying the second as the first is why teams end up tuning models that were never the problem.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.