AI & Media
How many AI hallucinations happen every day? (Live Counter)
A live count of false or fabricated AI answers generated globally
Roughly 319 false answers every second.
Source: Vectara Hallucination Leaderboard; OpenAI query estimates. View on dashboard →
What is an AI hallucination?
An AI hallucination is a confident, plausible-sounding answer that's wrong or made up. The term comes from neuroscience: like a hallucinating brain, the model "sees" facts that aren't there. LLMs predict the next token; they don't check if it's true. With billions of queries per day, even a small error rate means hundreds of millions of false answers daily.
Why AI confidence is the real problem
Why AI systems hallucinate
Large language models generate text by predicting the most statistically likely continuation of a prompt, based on patterns learned from billions of documents. They have no "knowledge" in the traditional sense, no internal model of reality to verify against. When asked about something outside their training data, or when faced with ambiguous or under-specified queries, they fill the gap with plausible-sounding text. This produces confident, well-formed sentences that contain false information. The problem is fundamental to the transformer architecture and cannot be entirely eliminated through training alone.
Scale: from millions to billions of queries
When ChatGPT launched in November 2022, the question of hallucinations was a niche concern among AI researchers. By February 2023, the service had 100 million users, making it the fastest-growing consumer application in history. By December 2024, OpenAI reported ChatGPT alone processes 1 billion prompts per day; by 2025 this grew to 2.5 billion daily. Across all AI assistants (Gemini, Copilot, Claude, etc.), total global AI query volume likely exceeds 5 billion per day. The live counter uses a conservative 2024–2025 baseline of ~27.5 million hallucinations per day (~318/sec), reflecting an effective blended rate applied to measurable query types. Even this conservative figure means that by the time you read this sentence, hundreds of AI hallucinations have already occurred.
Measuring the problem: the Vectara Leaderboard
The most widely used benchmark for LLM hallucination is the Vectara Hallucination Leaderboard, first published in late 2023. It measures how often models introduce factual errors when summarising provided documents, a controlled task that isolates hallucination from simple factual ignorance. The November 2023 results showed the best models (GPT-4) hallucinating in 3% of responses, while weaker models reached 27%. By March 2026, the frontier had improved to 1.8% (Antgroup Finix S1). However, these benchmarks reflect best-case conditions; real-world task hallucination rates are typically higher.
Economic and societal impact
AI hallucinations have caused measurable real-world harm. In 2023, a US attorney submitted a legal brief containing AI-generated case citations, none of which existed, and was fined by a federal judge. Healthcare AI tools have generated incorrect drug dosage advice. Forrester Research estimated the total business cost of AI hallucinations at $67.4 billion annually in 2024. As AI is increasingly used for high-stakes professional decisions, the cost of undetected hallucinations grows proportionally.
What this means for you
Every time you use an AI assistant and do not fact-check the output, you are statistically likely to act on a fabrication in 1 of every 10-30 responses, depending on the model and task. For information you plan to use, share, or act on: that number matters.
The danger is not that AI invents facts. It is that it does so with complete syntactic confidence. Hallucinated citations look identical to real ones. Hallucinated statistics are presented in the same tone as correct ones. There is no "I'm not sure" signal, no hesitation marker, and often no way to distinguish a hallucination from a correct answer without external verification.
Practical rule: treat any specific claim from an AI tool (a statistic, a name, a date, a citation) as a hypothesis to verify, not a fact to use. For high-stakes decisions (medical, legal, financial), independent verification is not optional. For lower-stakes uses, awareness of the limitation is sufficient.
AI hallucinations vs. total AI requests, today
The vast majority of AI requests produce useful answers. But as total volume surpasses 5 billion/day, even a small error rate means millions of hallucinations daily.
Key research findings on AI hallucinations
Even the best AI models in 2026 hallucinate in ~1.8-4.1% of responses (Vectara Leaderboard)
VectaraStanford HAI AI Index 2026: frontier AI models' training details (data, parameters, code) are no longer disclosed by OpenAI, Anthropic, or Google, making independent hallucination verification harder; responsible AI metrics show a widening gap between capabilities and trust infrastructure
Stanford University Human-Centered AIIn November 2023, AI models hallucinated in 3% to 27% of responses depending on the model
VectaraChatGPT processes approximately 2.5 billion prompts per day as of 2025 (OpenAI)
OpenAIChatGPT reached 100 million users in just 60 days, the fastest-growing consumer app in history at that time
OpenAIAI hallucinations cost businesses an estimated $67.4 billion per year globally (Forrester 2024)
ForresterHallucination rates in medical and legal AI applications can reach 10-40% depending on the model and task
Stanford University Human-Centered AIStanford HAI AI Index 2025: nearly 90% of notable AI models now come from industry; training compute doubles every five months, increasing both capability and hallucination surface area
Stanford University Human-Centered AIHow AI hallucination rates have changed over time
AI hallucination rates have improved per-model, but the absolute volume of false AI answers has exploded as query volumes surged from millions to billions per day between 2022 and 2025.
| Year | Rate | Est. per day | Context |
|---|---|---|---|
| 2022 | 6/sec | 501K | ChatGPT launches Nov 2022 |
| 2023 | 76/sec | 6.6M | ChatGPT, Bard, Claude all launch; rapid growth |
| 2024-2025 | 319/sec | 27.5M | Volume up; per-query rate improving; rate applied from lab benchmark, not production measurement |
| 2026 (forecast) | 1K/sec | 109.7M | Query volume surges; rate improvement slower than growth |
What hallucination benchmarks show
| Year | Finding | Value | Source |
|---|---|---|---|
| 2023 | Vectara Hallucination Leaderboard (Nov 2023): GPT-4 3.0%, GPT-3.5 3.5%, Llama 2 70B 5.1%, Claude 2 8.5%, Google PaLM-Chat 27.2% | 3 % hallucination rate (GPT-4, best 2023) | Vectara |
| 2023 | ChatGPT reaches 100 million weekly active users by November 2023 | 100 million weekly active users | OpenAI |
| 2024 | ChatGPT processes ~1 billion prompts per day as of December 2024 | 1.0B prompts/day (Dec 2024) | OpenAI |
| 2024 | ChatGPT reaches 200 million weekly active users (August 2024) | 200 million weekly active users | OpenAI |
| 2024 | ChatGPT reaches 300 million weekly active users (December 2024) | 300 million weekly active users | OpenAI |
| 2025 | ChatGPT processes ~2.5 billion prompts per day in 2025; 800+ million weekly active users | 2.5B prompts/day | OpenAI |
| 2026 | Vectara Leaderboard (Mar 2026): Antgroup Finix S1 32B 1.8%, Gemini-2.5-Flash-Lite 3.3%, Phi-4 3.7%, Llama-3.3-70B 4.1% | 2 % hallucination rate (best model, Mar 2026) | Vectara |
When hallucinations became a known problem
- 2022ChatGPT launches (November 2022), LLM hallucinations enter mainstream awareness
- 2023ChatGPT reaches 100 million users in 60 days, fastest-growing consumer app in history
- 2023Vectara publishes Hallucination Leaderboard; GPT-4 achieves 3.0%, seen as strong performance
- 2023US attorney sanctioned for submitting AI-generated fictional legal citations in federal court
- 2024ChatGPT processes 1 billion prompts/day (December 2024); daily false answers reach tens of millions
- 2025ChatGPT reaches 2.5 billion prompts/day; AI assistants used by 10% of global population
- 2026Best models achieve 1.8% hallucination rate; absolute hallucination volume continues to rise
Scale of the problem in perspective
At roughly 320 false AI answers per second (the live counter), that is roughly 27.5 million incorrect AI responses produced every single day
If each hallucinated answer took 1 minute to fact-check, humanity would need about 19,100 full-time fact-checkers just to review the AI errors produced in a single second
27.5 million false AI answers per day; each one potentially misleading someone who trusts the response
How the number is calculated
The live rate of ~318/sec corresponds to ~27.5 million hallucinations per day, a conservative 2024–2025 baseline. This is derived from an effective blended rate applied to AI assistant query volume (ChatGPT, Gemini, Copilot, Claude, and others). Total global query volume likely exceeds 5 billion/day; the live counter applies a conservative effective rate to measurable query types. The hallucination rate varies enormously by model and task, from 1.8% (best models, focused tasks) to 30%+ (weaker models, open-ended queries). The 9.2% figure from Vectara's Hallucination Leaderboard applies to specific summarisation benchmarks; real-world rates are task-dependent. Forecasts project higher volumes as query growth outpaces accuracy improvements.
Sources: Vectara Hallucination Leaderboard (2023-2026) - OpenAI Signals Data - ChatGPT Usage Statistics 2025. Methodology →
Frequently asked questions
- What is an AI hallucination?
- An AI hallucination is a factually incorrect or fabricated output that a language model presents with apparent confidence. Examples include invented citations, wrong dates, non-existent laws, and false statistics. The model is not "lying", it is completing text in the most statistically likely way, without verifying truth.
- How common are AI hallucinations?
- Rates vary significantly by model and task. As of early 2026, the most accurate models on the Vectara leaderboard achieve 1.8-4.1% hallucination rates in summarisation tasks. Older or smaller models can hallucinate 15-30% of the time. For high-stakes tasks like medical or legal information, studies have found hallucination rates of 10-40% across models.
- Are AI hallucinations getting better or worse?
- The per-query rate is improving: in November 2023, the best model scored 3.0% on Vectara's benchmark; by March 2026, top models are at 1.8%. However, the absolute number of hallucinations is rising because AI query volume is growing faster than accuracy improvements, more people using AI means more false answers in total.
- What is the economic cost of AI hallucinations?
- Forrester Research (2024) estimated the business cost of AI hallucinations at approximately $67.4 billion globally per year, from wrong decisions, wasted work, and legal exposure. This figure is expected to grow as AI adoption in professional settings increases.
- Which AI models hallucinate the least?
- As of March 2026, the Vectara Hallucination Leaderboard shows Antgroup Finix S1 32B (1.8%), Google Gemini-2.5-Flash-Lite (3.3%), Microsoft Phi-4 (3.7%), and Meta Llama-3.3-70B (4.1%) as the most accurate models. However, performance varies by task type, a model excellent at summarisation may perform worse at factual Q&A.
How the AI hallucination estimate is derived
Hallucination rates are sourced from Vectara's Hallucination Leaderboard (the most widely cited independent benchmark) and corroborated by Stanford HAI's AI Index. Query volumes use OpenAI's disclosed 2.5 billion/day for ChatGPT plus independent estimates for other platforms. Because AI companies treat query volume as confidential, the global query estimate carries meaningful uncertainty.
Sources
Vectara Hallucination Leaderboard (2023-2026) - OpenAI Signals Data - ChatGPT Usage Statistics 2025.
Explore related: Cost of AI hallucinations - Hours lost to AI rework - AI assistant requests, and the live AnythingCounter dashboard.