Back to Blogs
AI Hallucination Examples: What They Look Like in Real Enterprise Workflows
The most expensive AI hallucination examples rarely look like the funny screenshots that circulate online. They look like a chatbot inventing a refund policy that a tribunal later forces the airline to honor. A consulting report delivered to a government with fabricated citations. A medical transcription tool inserting treatments no doctor prescribed. This article walks through documented AI hallucination examples from real enterprise workflows, explains why they happen, and shows what they cost.
Key Insights
- An AI hallucination is confidently presented generated output not supported by source material or verifiable fact; the two types are factuality (contradicts real fact) and faithfulness (contradicts the source given).
- Documented enterprise examples include fabricated case law (Mata v. Avianca), an invented chatbot policy that cost Air Canada a tribunal ruling, fake citations in a paid Deloitte government report, fabricated content in medical transcriptions, and hallucinated software packages exploited via slopsquatting.
- The most expensive hallucinations are fluent and plausible, not obvious, which is why they survive review and reach customers, regulators, and courts.
- Hallucinations persist as models improve because reasoning models generate more claims, agentic workflows compound errors across steps, and retrieval does not guarantee grounding.
- The enterprise fix is architectural: per-output source traceability, influence scoring, full agent execution logs, and human contestability, measured on your own workflows.
First, the definition, because precision matters here. An AI hallucination is output from a generative AI system that is presented confidently but is not supported by the system’s source material or by verifiable fact. The model is not malfunctioning in a technical sense. It is doing what it was built to do, predicting plausible text, and plausibility is not the same thing as truth. Hallucinations in AI include fabricated facts, invented citations, nonexistent products, made-up policies, and false attributions, all delivered in the same fluent, confident tone as correct answers.
Who needs to understand this: anyone deploying generative AI in workflows where a wrong answer carries legal, financial, or operational consequences.
That covers legal and compliance teams, customer service leaders, finance organizations, healthcare operators, and engineering teams shipping AI-assisted code.
What you will get from this article: a clear AI hallucination definition, seven categories of documented examples mapped to the enterprise workflows where they occur, the cost mechanics behind them, and the architectural reason they persist even as models improve.
What Is an AI Hallucination? A Working Definition for Enterprises
What is an AI hallucination, precisely? It is unsupported generated content, and the unsupported part is what makes it dangerous. A model that says “I don’t know” is safe. A model that fabricates a court case, complete with docket number and plausible quotations, creates work and risk in equal measure.
Two types of AI hallucinations matter for enterprise risk assessment:
- Factuality hallucinations. The output contradicts verifiable real-world fact. Example: citing a research paper that does not exist.
- Faithfulness hallucinations. The output contradicts or exceeds the source material it was given. Example: a summary that adds claims the original document never made.
Faithfulness hallucinations are the quieter threat in enterprise settings, because they occur even in retrieval-augmented systems that were supposed to fix the problem. Giving the model the right documents does not guarantee the model stays inside them.
Quick glossary: terms that show up in every hallucination discussion
- AI hallucination: confidently presented generated output unsupported by source material or verifiable fact.
- Factuality hallucination: output that contradicts real-world fact (a fabricated statistic, a nonexistent case).
- Faithfulness hallucination: output that contradicts or exceeds the source material the model was given (a summary that adds claims the document never made).
- Confabulation: a synonym for hallucination used in some research and in the NIST AI Risk Management Framework, which treats it as a reliability risk for generative systems.
- Grounding: tying a model’s output to verifiable source material. Strong grounding lowers hallucination risk but does not eliminate it.
- Slopsquatting: a supply-chain attack where someone registers a software package name that AI code assistants hallucinate, then fills it with malicious code. The term combines “AI slop” with “typosquatting.”
These distinctions matter for risk assessment because the fix differs by type. Factuality hallucinations call for better grounding and source control. Faithfulness hallucinations call for output-level verification against the specific sources retrieved, which is a harder and more architectural problem.
Seven AI Hallucination Examples From Real Enterprise Workflows
These AI hallucination examples are documented public incidents, organized by the workflow where each occurred. The pattern to watch: in every case, the output was fluent, specific, and wrong.
1. Legal research: fabricated case law
In Mata v. Avianca (2023), attorneys submitted a federal court brief citing multiple judicial decisions that did not exist. ChatGPT had invented the cases, complete with realistic citations, and even “confirmed” their authenticity when asked. The court sanctioned the lawyers. Since then, courts in multiple countries have sanctioned dozens of filings containing AI-fabricated citations, and Stanford RegLab research found general-purpose models hallucinating on 69 to 88 percent of legal queries. Legal language is dense, precedent is long-tail, and the cost of an invented authority is professional discipline.
2. Customer service: invented policy
Air Canada’s website chatbot told a customer he could buy a full-fare ticket and apply for a bereavement discount afterward. No such policy existed. In 2024, a Canadian tribunal ruled the airline liable for its chatbot’s statement and ordered compensation, rejecting the argument that the chatbot was a separate entity responsible for its own words. The precedent generalizes: what your customer-facing AI says, your company has said.
3. Consulting and reporting: fabricated citations in a paid deliverable
In late 2025, Deloitte Australia partially refunded the Australian government after a commissioned report was found to contain AI-generated errors, including references to nonexistent sources. The workflow was ordinary knowledge work, drafting a report, which is exactly where generative AI adoption is heaviest. The incident shows hallucinations surviving professional review and reaching a paying client.
4. Healthcare documentation: transcription that adds content
An Associated Press investigation in 2024 reported that OpenAI’s Whisper speech-to-text model, widely used in medical transcription tooling, sometimes fabricated entire sentences, including invented medications and commentary, in transcripts of clinical conversations. Researchers found hallucinated content in a meaningful share of transcripts examined. In a documentation workflow, a fabricated sentence does not look like an error. It looks like part of the record.
5. Software engineering: package hallucinations in generated code
Code assistants regularly suggest importing software packages that do not exist. Security researchers have shown that attackers can register those hallucinated package names and fill them with malicious code, a supply chain attack pattern sometimes called slopsquatting. Here the hallucination is not just wrong output. It is an exploitable attack surface inside the development workflow.
6. Financial analysis: confident numbers without provenance
Generative AI hallucinations in finance workflows tend to be subtler: a summarized earnings figure attributed to the wrong quarter, a covenant threshold “recalled” rather than retrieved, a risk rating justified by reasoning the source documents do not contain. Few of these become public incidents precisely because finance teams catch them through manual review. That catch rate is the hidden cost: analysts re-verifying every AI output are paying for the work twice. One documented data point on scale: Deloitte CFO Advisory reported in January 2026 on a healthcare enterprise that accumulated roughly $6 million in unplanned annualized AI cost from token consumption that finance could not trace to defensible output.
7. Public-facing content: the launch-day error
Google’s Bard famously claimed in its February 2023 demo that the James Webb Space Telescope took the first picture of an exoplanet. It did not. The error coincided with a roughly $100 billion single-day drop in Alphabet’s market value. The example endures because it compresses the whole problem into one frame: a confident, specific, plausible claim, wrong, in public, with the brand attached.
Two more, from government and developer tooling
Government services: New York City’s MyCity chatbot. In 2024, investigative reporting found that New York City’s official small-business chatbot, built on a major cloud AI service, advised users in ways that conflicted with the law, at one point suggesting that employers could take workers’ tips. A government chatbot dispensing incorrect legal and regulatory guidance is the public-sector version of the Air Canada problem, with citizens rather than customers on the receiving end.
Developer tooling: the AI support agent that invented a policy. In 2025, the AI coding tool Cursor saw users report that its own AI support agent told them a (nonexistent) policy restricted using the product across multiple devices. Users canceled subscriptions over a rule that did not exist until the company clarified. The incident shows the pattern turning inward: a company’s own AI hallucinating about the company’s own product, to its own customers.
AI Hallucination Examples Mapped to Workflow Risk
| Workflow | Documented example | Primary risk | Who absorbs the cost |
| Legal research and filings | Mata v. Avianca fabricated cases | Court sanctions, malpractice exposure | Legal, outside counsel |
| Customer service | Air Canada chatbot policy | Legal liability, brand damage | Support, legal, brand |
| Reports and deliverables | Deloitte Australia government report | Refunds, reputational harm | Consulting teams, client trust |
| Clinical documentation | Whisper transcription fabrications | Patient safety, record integrity | Clinicians, compliance |
| Software development | Hallucinated package names | Supply chain compromise | Engineering, security |
| Financial analysis | Untraceable figures and reasoning | Bad decisions, re-verification cost | Finance, risk |
| Public content | Bard JWST claim | Market and brand impact | The entire company |
| Government services | NYC MyCity chatbot | Citizen harm, legal exposure | Public agencies, residents |
| Developer tooling | Cursor support agent policy | Churn, self-inflicted brand damage | Product, support |
Why Hallucinations in AI Persist Even as Models Improve
The industry narrative says hallucinations are shrinking, and on narrow summarization benchmarks, they are. Production data tells a different story, for three reasons.
Reasoning models generate more claims. OpenAI’s own system card from April 2025 showed its o3 model hallucinating on 33 percent of PersonQA prompts, double the rate of its predecessor o1. Longer reasoning chains mean more assertions, and every assertion is an opportunity to be wrong.
Agentic workflows compound errors. An agent that plans, retrieves, calls tools, and composes a final answer can hallucinate at any step, and errors propagate downstream. Microsoft Research has measured agentic tasks consuming roughly 1,000 times more tokens than chat interactions. Enterprises are scaling exactly the workload shape where per-step error rates compound.
Retrieval does not equal grounding. RAG narrows the problem but does not close it. If retrieval surfaces an outdated policy document, the model can faithfully summarize it and still be wrong. If retrieval surfaces nothing relevant, many systems answer anyway. Faithfulness hallucinations occur on top of retrieved content, which is why “we use RAG” is not a complete answer to a risk committee.
What Enterprises Can Do About It
A short checklist for teams moving from AI hallucination examples to action:
- Inventory where generative AI touches consequential decisions. The AI hallucination examples above cluster in legal, customer-facing, clinical, financial, and engineering workflows. Map yours.
- Demand provenance, not just answers. Systems should show which sources each output reasoned from. An answer with no traceable source should be treated as unverified by default.
- Score and control source quality. The fastest way to cut hallucinations is to find and fix the bad sources the model relies on, which requires influence visibility at the data level.
- Log complete execution traces for agents. When a multi-step workflow goes wrong, you need to see every step to find where it went wrong.
- Keep humans in control of high-stakes outputs. Contestability, the ability to challenge and correct an output, is becoming a regulatory expectation under the EU AI Act, not just good practice.
- Measure your own rate. Build an evaluation set from real workflow queries and track hallucination and defensibility rates over time. Public benchmarks will not do this for you.
This is the design problem Seekr built SeekrFlow around. Rather than treating the model as a sealed box, SeekrFlow traces every output back to the specific source documents and training data that shaped it, scores how much each source influenced the answer, and logs full execution traces across agent workflows. When an output is wrong, teams can see why, correct the source, and prove the fix, which turns hallucination management from incident response into an operating discipline. SeekrGuard adds evaluation and certification before models reach production, which is the step most of the incidents above never had.
One honest limitation: no architecture makes the examples above impossible. Generative systems remain probabilistic, and low-stakes internal drafting with human review may not justify full source-control infrastructure. The dividing line is consequence. Where an output can trigger a sanction, a liability, a patient record error, or a market reaction, traceability is the difference between an incident you can explain and one you cannot.
Summary block. AI hallucination examples in enterprise settings include fabricated case law, invented customer policies, fake citations in paid reports, fabricated clinical transcript content, nonexistent software packages, and untraceable financial figures. They persist because reasoning models generate more claims, agent workflows compound errors, and retrieval does not guarantee grounding. The enterprise response is architectural: source traceability, influence scoring, audit trails, and human contestability.
Get the Hallucination Tax Field Guide
These examples are the surface. The field guide breaks down where hallucinations originate in production workflows, what each one costs, and the architecture that traces every output back to its sources.
Read the Field Guide
Frequently Asked Questions
What is an AI hallucination?
An AI hallucination is confidently presented output from a generative AI system that is not supported by its source material or by verifiable fact. It includes fabricated facts, invented citations, nonexistent policies, and false attributions, delivered in the same fluent tone as accurate answers.
What are the most common AI hallucination examples in enterprises?
The most common AI hallucination examples in enterprises are fabricated legal citations, invented customer service policies, fake references in reports, fabricated content in transcriptions, hallucinated software packages in generated code, and financial figures that cannot be traced to source documents.
What are the two types of AI hallucinations?
The two types of AI hallucinations are factuality hallucinations, where output contradicts verifiable real-world fact, and faithfulness hallucinations, where output contradicts or goes beyond the source material the model was given. Faithfulness hallucinations occur even in retrieval-augmented systems.
Why do hallucinations in AI happen?
Hallucinations in AI happen because generative models predict plausible text rather than verify truth. Gaps in training data, ambiguous prompts, weak retrieval, and long reasoning chains all increase the chance the model fills a gap with fluent fabrication instead of declining to answer.
Do generative AI hallucinations go away with RAG?
Generative AI hallucinations are reduced by RAG but do not go away. Retrieval can surface wrong or outdated documents, and models can still add claims the retrieved sources do not support. Faithfulness hallucinations on top of retrieved content are a documented failure mode in production systems.
How costly are real-world examples of AI hallucinations?
Real-world examples of AI hallucinations have produced court sanctions for lawyers, a tribunal-ordered payout for Air Canada, a partial government refund by Deloitte Australia, and an approximately $100 billion single-day market value drop after Google Bard’s launch error. The recurring internal cost is human re-verification of every AI output.
How can enterprises prevent AI hallucinations in their workflows?
Enterprises prevent AI hallucinations most effectively by controlling and tracing sources: requiring per-output provenance, scoring source influence, logging full agent execution traces, keeping humans able to contest outputs, and measuring hallucination rates on their own workflows rather than relying on public benchmarks.
What is slopsquatting?
Slopsquatting is a supply-chain attack that exploits AI hallucinations in code generation. When AI coding assistants invent the names of software packages that do not exist, attackers register those names and fill them with malicious code, so a developer who installs the hallucinated package imports the attack. It is a documented reason hallucinations in AI are a security issue, not only an accuracy one.
Is an AI hallucination the same as a confabulation?
An AI hallucination and a confabulation refer to the same phenomenon: confident generated output unsupported by fact or source. Confabulation is the term used in some research literature and in the NIST AI Risk Management Framework, which classifies it as a reliability risk for generative AI systems that enterprises are expected to manage.
