Hallucination (AI)

A language model generates the most plausible next words, not the most accurate ones. When it lacks a fact, it rarely says so. It produces something that fits the pattern: a specific-looking percentage, a study with a real-sounding title, a product feature that would make sense but does not exist. The hallucinated sentence reads exactly like the true sentences around it, which is why it slips through.

Hallucination is measurable. The most practical test gives a model a document, asks it to summarise using only that document, and checks whether the summary adds anything the document does not say. Every widely used model still does it at some rate, even in that constrained task.

Marketing content is where hallucinations do the most damage per sentence. It makes claims about products, prices, customers and results, it is published under the company’s name, and buyers act on it. An invented statistic in a blog post can end up quoted in a sales deck, a proposal and an AI answer about your company.

The rates on public benchmarks are low enough to feel safe and high enough to guarantee errors at volume. A team publishing dozens of AI-assisted pieces a month should assume some contain a hallucination unless each claim is checked against a source.

• Vectara’s leaderboard measures “how often an LLM introduces hallucinations when summarizing a document”, asking each model to summarise “using only the facts presented in the document”.Vectara, Hallucination Leaderboard (last updated May 11, 2026), 15 Sept 2026
• 1.8%On that leaderboard, last updated May 11, 2026, the best listed model hallucinated in 1.8 % of summaries, openai/gpt-4o-2024-08-06 in 9.6 %, google/gemini-2.5-pro in 7.0 % and anthropic/claude-sonnet-4-20250514 in 10.3 %.Vectara, Hallucination Leaderboard (last updated May 11, 2026), 15 Sept 2026
• Google Cloud describes hallucinations as “instances where the model generates content that isn’t factual” and lists reducing them as the first benefit of grounding.Google Cloud, Grounding overview, 15 Sept 2026

1. 1

   Treat every number, named study, quote, price and customer name in AI-assisted content as unverified until someone opens its source.

2. 2

   Generate from approved source material rather than asking a general model what is true about your company or market.

3. 3

   Require citations per claim, so review becomes opening links rather than rereading prose for anything that sounds off.

4. 4

   Keep a short list of claims that must never be generated freely, such as prices, security certifications and customer results, and fill them only from the source of record.

5. 5

   When you find a hallucination, fix the source gap that allowed it, not just the sentence.

• Assuming a better model has stopped hallucinating. Benchmark rates fall, but none of the widely used models on the leaderboard reach zero.
• Proofreading for hallucinations. Editors catch the implausible ones; the dangerous ones are plausible by design.
• Trusting a citation because it exists. Models can attach a real source to a claim that source never makes.

Outdated information

Outdated information was true once, such as last year’s price. A hallucination was never supported. Both are fixed by grounding in current sources, but outdated facts are also fixed by maintaining them.

Grounding

Grounding is the main structural defence: the model writes only from supplied sources, and unsupported claims are left out.

1. 1Vectara, Hallucination Leaderboard (last updated May 11, 2026) read 15 Sept 2026
2. 2Google Cloud, Grounding overview read 15 Sept 2026