Enterprise AI in Numbers · Capability and cost
How often AI models make things up when summarising a document
4.7%
4.7% as of October 22, 2026, according to Vectara, Hallucination Leaderboard. For each lab, the lowest hallucination rate among its evaluated models (share of summaries that include a claim the source document doesn't support); the reading is the median across labs.
What it means for leaders
Even grounded in a source document, leading models still add unsupported claims in a small but steady share of summaries. Any workflow that publishes or acts on AI summaries needs sampling checks.
What would change the reading: Rates below one in a hundred, and whether newer flagship models do better than older ones.
The breakdown
Every reading
| Date | Value | Change | Source |
|---|---|---|---|
| 4.7% | — | Source |
Definition and method
- What it measures
- For each lab, the lowest hallucination rate among its evaluated models (share of summaries that include a claim the source document doesn't support); the reading is the median across labs.
- Population
- Models from ten major labs evaluated on the Vectara hallucination leaderboard
- Updated
- About monthly
- Source
- Vectara, Hallucination Leaderboard · Apache 2.0
Caveats
- It tests summarising a given document only, judged by Vectara's own detection model; it is not a general truthfulness score.
- A lab's lowest rate may come from a small model rather than its flagship.
Cite this
Enterprise AI in Numbers, "How often AI models make things up when summarising a document", https://lucianoayres.com/ai-brief/data/frontier-hallucination-rate/, retrieved October 25, 2026. Data: Vectara, Hallucination Leaderboard.