VeltrixVeltrix.
← All articles
10 / 62March 14, 2026

AI Trust and Hallucination

The confident liar. Why AI models state falsehoods with certainty.

AI hallucinations are not glitches. They are a structural feature of how large language models work. LLMs don't retrieve facts from a database — they predict the statistically most probable next word based on patterns learned during training. When the training data runs thin, or when a question requires precise knowledge the model doesn't have, it doesn't stop. It continues predicting. It generates something that sounds correct, in the format of something correct, with the confidence of something correct.

Type 01
Factual Fabrication

The model invents a fact that has no basis in reality — statistics, events, people, relationships — and presents it as established knowledge.

"The ACME study (2023) found that 78% of participants showed improvement..." — the study does not exist.
Type 02
Citation Fabrication

The model generates plausible-looking references — real-sounding journals, real-sounding authors, real-sounding titles — that do not exist. URLs may look valid. Authors may be real but never wrote the cited work.

"Smith et al. (2022), Journal of Applied Economics, pp. 145–162..." — no such paper exists.
Type 03
Faithful Confabulation

The model faithfully summarises a real document — but introduces facts, conclusions, or quotes that were never in the source. It stays "on topic" while inventing the detail.

OpenAI's Whisper inserted words never spoken during medical transcriptions, including invented treatment names.
So what does this mean?

AI hallucination isn't a bug to be patched — it's a structural feature of how LLMs work. A 2025 mathematical proof confirmed it cannot be fully eliminated under current architectures.

This changes how you should interact with every AI tool. The confident, authoritative tone isn't evidence of accuracy — it's evidence that the model successfully predicted a fluent-sounding sequence. Treat every AI output as a draft that requires verification, not a source of truth.

Hallucination rates vary enormously by model, task type, and how you measure them. The best models, on controlled summarisation benchmarks, are now below 1%. But "summarising a short document" is not the same as "answering a complex question about a real person, case, or event" — and on those harder tasks, rates climb dramatically.

Gemini 2.0 FlashBest-in-class (grounded tasks)VEC
0.7%
GPT-4 / Claude 3.7Strong models (general tasks)VEC
1–5%
Cross-model meanAverage across all modelsAAI
9.2%
OpenAI o3 / o4-miniReasoning models (open-ended)OAI
33–48%
Best citation modelPerplexity (citation tasks)CJR
37%
The improvement story is real but incomplete: the best models dropped from a 21.8% hallucination rate in 2021 to 0.7% on controlled benchmarks in 2025 — a 96% reduction.AAI However, a 2025 mathematical proof confirmed that hallucinations cannot be fully eliminated under current LLM architectures. The models generate statistically probable sequences. Some confabulation is structurally unavoidable.
So what does this mean?

The headline 0.7% rate is for controlled document summarisation. For the tasks most people actually use AI for — research, citations, complex questions about real people and events — rates climb to 33–48%.

The gap between marketing claims and real-world performance is enormous. When a vendor says "our model hallucinates less than 1%," ask: on what benchmark, for what task? The number that matters is the one for your use case, not the one on the leaderboard.

These are not hypothetical examples. Every case below is documented and consequential. Taken together, they reveal a pattern: hallucinations cause the most damage when they occur in high-trust, high-stakes contexts — precisely the environments where AI adoption is being pushed hardest.

Legal — USA, 2023

Mata v. Avianca: The Ghost Docket

A New York attorney used ChatGPT to draft court filings and submitted six completely fabricated legal citations — real-sounding case names, real-sounding courts, real-sounding outcomes. When opposing counsel challenged them, the attorney claimed he didn't know ChatGPT was generative rather than a legal database. Federal judge sanctioned both lawyers.MATA By May 2025, lawyers were submitting hallucinated content to US courts at a rate of two to three cases per day.DC

Outcome: Sanctions issued. 324 of 1,081 documented hallucination court cases are in the US alone.
Consulting — AU & CA, 2025

Deloitte: Two governments, two hallucinated reports

In July 2025, Deloitte submitted a $290,000 report to the Australian government containing fabricated academic citations, non-existent footnotes, and an invented quote from a federal court judge. Weeks later, a $1.6M CAD (~$1.13M USD) Deloitte report for the Canadian government on healthcare workforce was found to contain similar AI-generated errors including fictional academic papers.DEL Both reports were live on government websites before anyone noticed.

Outcome: Partial refund. Governments received policy-influencing reports based partly on non-existent evidence.
Healthcare — USA, ongoing

OpenAI Whisper: Words that were never spoken

OpenAI's Whisper speech-to-text model, widely deployed in healthcare settings for clinical transcription, was found to hallucinate content during transcription — inserting words and phrases that were never spoken, including violent language, racial references, and invented medical treatment names.WHSP Even the best models hallucinated potentially harmful medical information 2.3% of the time when tested on medical questions.

Outcome: No liability ruling yet. 2.3% on medical questions means roughly 1 in 43 clinical AI responses may contain a harmful fabrication.
Media — USA, 2025

Chicago Sun-Times: The books that don't exist

Readers of the Chicago Sun-Times discovered that a "Summer Reading List for 2025" included 10 fabricated book titles attributed to real authors — books that were convincingly described but had never been written. Only 5 of 15 titles were real. The list had been generated by AI by an outside content supplier and published in the print edition.AAI In Q1 2025 alone, 12,842 AI-generated articles were removed from online platforms due to hallucinated content.

Outcome: Print readers cannot un-read what they read. The hallucinated books remain cited in some online conversations.
So what does this mean?

Hallucinations cause the most damage in high-trust, high-stakes contexts — law, medicine, government policy — precisely the environments where AI adoption is being pushed hardest.

The pattern is clear: the consequences scale with the authority of the context. A hallucinated email is an embarrassment. A hallucinated court filing is a career-ending sanction. A hallucinated medical transcription is a patient safety incident. Know the stakes before you deploy.

Hallucinations are not a reputational problem — they are a financial one. The cost accumulates through lost time, remediation work, legal exposure, and decisions made on bad information.

$67.4B
Global losses (2024)

Total estimated business cost of AI hallucinations globallyAAI

4.3h
Per worker per week

Average time knowledge workers spend fact-checking AI outputsMSFT

47%
Made a bad decision

Of enterprise AI users who acted on hallucinated content in 2024DRN

$14.2K
Per employee per year

Estimated hallucination mitigation cost per enterprise employeeFOR

Legal exposure timeline

2023
Hallucinations enter the courts

Mata v. Avianca: first major US case.MATA 7 in 10 hallucination court cases involve self-represented litigants. Judges begin issuing standing AI disclosure orders.

2024
Professional liability expands

Walters v. OpenAI: defamation claim over hallucinated output proceeds through courts. Air Canada chatbot ruling sets corporate liability precedent. Courts establish that hallucinations about real people are "publication risks."

May 2025
2–3 cases per day

By May 2025, lawyers submit hallucinated AI content to US courts at 2–3 cases per day.DC 13 of 23 court-caught hallucination cases are now the fault of qualified legal professionals, not laypeople.

2025–2026
1,081 tracked cases and counting

Researcher Damien Charlotin's AI Hallucination Cases Database now tracks 1,081 documented court cases globally.DC The hallucination detection market grew 318% between 2023 and 2025 as enterprises scrambled for solutions.

So what does this mean?

This isn't a reputational risk — it's a financial one. At $14,200 per employee annually and 4.3 hours weekly spent fact-checking, the "productivity gains" from AI are being significantly offset by the verification tax.

Factor this into every AI ROI calculation. The 47% figure — nearly half of enterprise AI users making decisions on hallucinated content — means your organisation probably has unaudited AI exposure right now. Map it before it maps you.

The industry is not standing still. Four mitigation strategies have demonstrated measurable results — though none eliminates the problem entirely, because elimination is architecturally impossible under current LLM design.

01
Retrieval-Augmented Generation (RAG)

Instead of answering from parametric memory, the model retrieves relevant documents from a verified knowledge base before generating a response. Grounds output in actual source material. Most effective for domain-specific enterprise deployments where the knowledge base can be curated and verified.

Reduction: ~71% fewer hallucinations on grounded tasks
02
Forced citation with source verification

Requiring the model to cite a source for every factual claim — and then verifying that the cited source exists and contains the claimed information — catches fabricated citations before they leave the system. Perplexity's model is built on this principle; even so, it hallucinated 37% of citations in the Columbia Journalism Review test.CJR

Note: Even citation-first models hallucinate at scale on hard tasks
03
Human-in-the-loop verification

76% of enterprises now run explicit human review processes before AI outputs are used in high-stakes decisions.DRN Not a technical solution — it's the acknowledgement that current AI cannot be trusted without human oversight. The cost: 4.3 hours per worker per week.MSFT

Adoption: 76% of enterprises have implemented this (2025)
04
Abstention training (teaching models to say "I don't know")

The most underused technique. Models can be trained to abstain from answering when uncertain rather than fabricating.MIT MIT research shows models that abstain 52% of the time dramatically cut error rates. The industry pressure to be helpful works against abstention — but this is where architecture and incentives diverge most dangerously.

Key insight: Abstention is not failure — it is accuracy
So what does this mean?

No single technique eliminates hallucinations — because elimination is architecturally impossible. The winning strategy is layered: RAG for grounding, forced citations for verification, human review for high-stakes decisions, and abstention for honest uncertainty.

If you're deploying AI in production, you need all four. If you're using AI personally, the minimum is simple: verify every important claim before acting on it. When an AI says "I don't know" — that's the system working correctly, not failing.

$67.4B

global business losses from AI hallucinations in 2024 aloneAAI

34%

more confident language when AI generates incorrect information than correctMIT

1,081

documented court cases involving AI hallucinations globally and countingDC

5 things you can do this week
to protect yourself from AI hallucinations.
1.

Subscribe to Veltrix Collective for weekly AI reality checks. We track hallucination rates, mitigation tools, and case law so you don't have to. Every Tuesday, data-backed, jargon-free. One email that keeps you ahead of the curve.

2.

Never trust a confident AI response without verification. MIT research shows AI is 34% more likely to use confident language when wrong. Treat certainty as a red flag, not a green light. When ChatGPT, Claude, or Gemini sounds absolutely sure — that's when you check.

3.

Verify every AI-generated citation before using it. Open the source. Check it exists. Read it. The Columbia Journalism Review found even the best citation-aware model gets 1 in 3 citations wrong. Use Google Scholar, Semantic Scholar, or CrossRef to verify before you cite.

4.

Set up a RAG pipeline for any repeated high-stakes AI task. Tools like Claude with file uploads, ChatGPT with browsing, or enterprise solutions like LangChain and n8n can ground AI responses in verified documents. This cuts hallucinations by ~71% on grounded tasks.

5.

Audit where AI outputs enter your decision process this week. The 47% figure — nearly half of enterprise AI users making decisions on hallucinated content — means your organisation probably has unaudited AI exposure. Map it. Then add human checkpoints at every high-stakes node.

$67.4 billion in losses. 1,081 court cases. 34% more confident when wrong. The data is clear — AI is powerful, but it requires informed users. Subscribe and stay informed.

Source references

AAI
AllAboutAI, "AI Hallucination Report 2025/2026"$67.4B global cost, 0.7% Gemini rate, MIT 34% confidence finding, 12,842 articles removed in Q1 2025, 9.2% cross-model mean hallucination rate.
VEC
Vectara HHEM Leaderboard, April 2025Model-by-model hallucination rates on document summarisation. GPT-4o ~1.5%, Claude 3.7 Sonnet ~4.4%, Gemini 2.0 Flash 0.7%.
OAI
OpenAI SimpleQA / PersonQA results, 2025o3 at 33%, o4-mini at 48% hallucination rate on factual questions about real individuals.
CJR
Columbia Journalism Review, March 2025Citation hallucination test — Perplexity was best performer at 37% hallucination rate. More than 1 in 3 citations were wrong.
MIT
MIT Human Dynamics Lab, January 2025Models 34% more likely to use confident language ("definitely," "certainly") when generating incorrect information than correct.
DEL
Fortune / Entrepreneur, October–November 2025Deloitte Australia ($290K report with fabricated citations) and Canada ($1.6M CAD report with fictional academic papers).
DC
Damien Charlotin AI Hallucination Cases Database, 20261,081 tracked court cases globally. 324 in US. 2–3 submissions per day by May 2025.
MATA
Mata v. Avianca, SDNY, 2023First major US case of AI-fabricated legal citations. Federal sanctions issued. Standing AI disclosure orders follow.
WHSP
OpenAI Whisper healthcare transcription researchInserted words never spoken during medical transcription, including violent language and invented treatment names. 2.3% harmful fabrication rate on medical questions.
FOR
Forrester Research, 2025$14,200 per enterprise employee annually in hallucination-related mitigation effort.
MSFT
Microsoft, 20254.3 hours per week per knowledge worker spent fact-checking AI outputs.
DRN
Drainpipe.io / Suprmind, 2025–2026Enterprise hallucination impact data. 47% made major business decision on hallucinated content. 76% human-in-the-loop adoption.
Veltrix Collective
AI's reality,
not its hype

Weekly data briefings on what AI actually does — the hallucinations, the breakthroughs, and the tools that separate signal from noise. No hype. Just what you need to use AI safely.

Weekly, every Tuesday · No spam · Privacy policy · Unsubscribe anytime
Data synthesis as of March 2026. All figures are best available approximations drawn from cited sources above — not a single primary study. Hallucination rates vary by model, task type, benchmark, and measurement methodology. Hover any TAG inline for source context, or see the reference key above.
Written by Luke Madden, founder of Veltrix Collective. Data synthesis and analysis by Vel.