How Many AI Hallucinations Occur Per Second? Global Estimates
Your AI just made something up. Again.
“AI systems produce an estimated 318+ false or hallucinated answers per second.”
Source: Vectara Hallucination Leaderboard and request-volume estimates. View on dashboard →
You ask a model a question. The answer looks confident, reads well, maybe even cites sources. Except it's wrong. Researchers call it hallucination – the system invented something, mixed up a reference, or gave a plausible-sounding answer that simply isn't true.
It happens more than you'd think. Across models and tasks, a few percent of answers are fabricated or incorrect. Scale that to billions of AI requests per day and you get what you see above: hundreds of errors every second, globally. Companies have lost millions to decisions made on hallucinated data. People have cited fake papers, followed wrong medical advice, trusted invented quotes.
The counter isn't an argument against AI. It's a reminder to check.
Key figures
| Time unit | Approx. rate |
|---|---|
| Per second | 318+ false/hallucinated answers |
| Per minute | ~19,100 |
| Per day | ~27.5 million |
| Per year | ~10 billion+ (based on request volume & hallucination rates) |
AI usage exploded in 2022–2023; so did the absolute number of wrong or fabricated answers. The rate of hallucinations (per query) hasn't necessarily gone up—but more queries mean more errors in the world.
Context: global AI request volume has grown sharply; hallucination rates (e.g. Vectara) suggest 3–15% of answers can be wrong or fabricated.
| Period | Estimated scale | Trend |
|---|---|---|
| 2022 | Lower request volume | Early LLM adoption |
| 2023 | Rapid growth | ChatGPT, Bard, others |
| 2024–2025 | ~10+ billion hallucinations/year | Volume up; rate stable |
What are AI hallucinations?
False or fabricated answers that AI systems give with high confidence—invented facts, wrong references, plausible-sounding errors. They affect anyone who trusts AI output without checking, from individuals to enterprises.
How the number is calculated
Exact calculation. We take an estimate of global AI requests per day (e.g. hundreds of millions) and multiply it by a published average hallucination rate (e.g. Vectara's benchmark, around 9.2%). That gives the number of false or fabricated answers per day. We then divide by 86,400 (seconds per day) to get the per-second rate. Hallucination rates vary by model and task; we use a central estimate. The counter multiplies this rate by seconds elapsed since midnight (or since 1 January) for “today” or “this year”.
Documents used for this calculation: Vectara Hallucination Leaderboard (blog & methodology). Full methodology and uncertainties: methodology page.
AI hallucination statistics
- Estimated AI hallucinations per second: 318+
- Estimated per minute: ~19,100
- Estimated per day: ~27.5 million
Tens of millions of AI hallucinations occur every day globally; organizations and users are affected when they rely on unverified output. Based on published hallucination rates and global request volume.
Why AI hallucinations matter at scale
More AI use means more errors in absolute terms. Request volume is rising; model limits—confabulation, no real fact-checking—keep the rate high. For now: verify important answers and treat AI as a tool, not a single source of truth.
FAQ
- How many AI hallucinations occur per second?
- Current estimates suggest over 318 false or hallucinated AI answers are produced every second worldwide. This is derived from published research on hallucination rates and global AI request volumes.
- Why do AI systems hallucinate?
- AI models generate plausible-sounding text by pattern rather than by verifying facts. They can invent references, mix up details, or confidently state false information. Rates vary by model and task; benchmarks like the Vectara Hallucination Leaderboard measure how often it happens.
- How are AI hallucination statistics calculated?
- We combine published hallucination rates (e.g. from Vectara and similar research) with estimates of global AI query volume. The result is converted to a per-second rate. Full methodology and sources are on our methodology page.
- Are AI hallucinations dangerous?
- In high-stakes contexts—medical, legal, financial, or safety-critical—hallucinated output can lead to wrong decisions and real harm. Many organizations already report losses from decisions based on incorrect AI-generated data. Checking important outputs remains essential.
Sources
Documents used for this statistic: Vectara Hallucination Leaderboard (blog & methodology). Full methodology and uncertainties: methodology page.
Explore more: phishing, deepfakes, and the AnythingCounter dashboard. You can also convert and back up your files with Anything Converter.