"How accurate is AI?" is one of the most searched questions about AI — and one of the hardest to answer honestly, because accuracy varies by model, task, domain, and how you measure it. A frontier model can hit 97% speech recognition accuracy on a phone call and 88% hallucination rate on a legal research query in the same afternoon. This page compiles the most current, primary-sourced AI accuracy statistics across hallucination rates, benchmark performance, medical diagnostics, coding, customer service resolution, and real-world deployments.

Quick answer: Frontier AI hallucination rates on factual tasks have fallen from 15–45% in 2024 to 3–19% in 2026 on standardised benchmarks (Digital Applied, April 2026). But hallucination rates on domain-specific tasks remain far higher: 17–88% for legal research queries (Stanford RegLab, 2024–2025) and up to 64% for medical case summaries without mitigation (MedRxiv, 2025). A 2025 mathematical proof confirmed that zero hallucination is architecturally impossible for any current large language model.

Top AI Accuracy Statistics for 2026 (Editor's Picks)

• 3.1%–19.1% — frontier AI hallucination rates in 2026 across five leading models on standardised factual, citation, and code tasks — down from 15–45% in 2024. — Digital Applied 5-model benchmark, April 2026

• 22%–94% — hallucination rate range across 26 top models on Stanford HAI's sycophancy benchmark, where models are presented with false statements the user appears to believe. — Stanford HAI AI Index 2026

• 362 AI incidents documented in 2025, up from 233 in 2024 — a 56% increase year-over-year. — AI Incident Database, cited in Stanford HAI AI Index 2026

• 96% — AI accuracy in diabetic retinopathy detection, outperforming specialists by more than 10 percentage points. — 2025 clinical trial syntheses

• 97%+ — speech recognition accuracy for English in 2026 production voice AI deployments. — Nuance/Microsoft / AInora benchmarks, 2025–2026

• 92%–96% — call resolution rates for well-configured AI voice agents on standard scenarios (booking, information, routing). — AInora, April 2026

• ~63% — average SWE-bench Verified score across 83 evaluated AI coding models as of April 2026, up from ~40% at end of 2024. — BenchLM.ai / Scale AI SEAL Leaderboard, April 2026

• $67.4 billion — estimated global financial losses tied to AI hallucinations in 2024. — cited in About Chromebooks / multiple sources, 2026

1. AI Hallucination Rates: The Overall Picture

Hallucination — confidently stated but false output — is AI's most documented accuracy failure. Rates vary enormously by task type, model, and whether mitigation (retrieval-augmented generation, structured prompts) is applied.

• 3.1%–19.1% — range of frontier model hallucination rates on standardised factual, citation, and code reference tasks in April 2026, based on 5,000 prompts across five models. This is substantially better than the 2024 baseline of 15–45%. — Digital Applied benchmark, April 2026

• 22%–94% — hallucination range across 26 frontier models on Stanford HAI's sycophancy accuracy benchmark, where a false statement is presented as something the user believes. Performance collapses in the user-belief condition even on models that handle the same false statement well when attributed to a third party. — Stanford HAI AI Index 2026

• ~9.2% — average hallucination rate across all models for general knowledge questions on standardised benchmarks. — About Chromebooks, citing Vectara / Hugging Face Leaderboard, 2026

• A 2025 mathematical proof confirmed that hallucinations cannot be fully eliminated under current large language model architectures — zero-hallucination is not achievable by design. — cited in multiple benchmark analyses, 2025

• $67.4 billion — estimated global financial losses tied to AI hallucinations in 2024. — cited in About Chromebooks / multiple sources, 2026

• 96% improvement in best-model hallucination rates from 2021 to 2025: the top model went from 21.8% hallucination to 0.7% on the Vectara summarisation benchmark over four years. — About Chromebooks / Vectara, 2026

• AI hallucination rates across models decline by roughly 3 percentage points per year on standardised benchmarks, per analysis of the Hugging Face Hallucination Leaderboard. — About Chromebooks, 2026

• 51% of organisations using AI have experienced at least one negative consequence from AI, with inaccuracy as the leading cause. — McKinsey State of AI, 2025

2. AI Hallucination Rates by Task Type

The same model can perform at opposite ends of the reliability spectrum depending on the task. Task type is a stronger predictor of hallucination risk than model choice alone.

• 0.7%–1.5% — hallucination rate on grounded summarisation tasks for top models in 2025, when the model is given a source document and asked to summarise it. — SQ Magazine / benchmark aggregations, 2026

• 10%–20% — hallucination rates on closed-domain question-answering tasks. — BIG-bench evaluation, cited in SQ Magazine 2026

• 20%–35% — proportion of incorrect AI outputs attributable to hallucination on the BIG-bench evaluation. — BIG-bench, cited in SQ Magazine 2026

• 33%–48% — hallucination rates for OpenAI's o3 and o4-mini reasoning models on PersonQA (person-specific factual questions). o3 hallucinated 33% of the time — double the rate of its predecessor o1. — OpenAI system card, cited in About Chromebooks 2026

• 40%–80% — hallucination rates on open-ended generation tasks (the highest-risk category). — benchmark aggregations, cited in SQ Magazine 2026

• 60%+ — incorrect answers from eight generative search tools on news-citation queries tested by the Columbia Journalism Review. — Columbia Journalism Review, cited in Suprmind 2026

• Reasoning models hallucinate 2–3x more than their non-reasoning equivalents on certain tasks, despite higher accuracy scores — indicating a real trade-off between reasoning depth and factual calibration. — CodingFleet hallucination analysis, June 2026

3. AI Hallucination Rates by Domain

Domain-specific accuracy gaps are among the most practically important AI accuracy statistics — the gap between a summarisation task and a legal or medical task is not incremental, it is categorical.

Medical & Healthcare

• 64.1% — hallucination rate in medical case summaries without mitigation prompts, per a 2025 MedRxiv study. — MedRxiv, 2025

• Structured prompts reduced medical hallucination rates by 33% in clinical research settings. — cited in SQ Magazine / benchmark analysis, 2026

• Open-source models show hallucination rates above 80% on some medical tasks, lagging significantly behind proprietary models. — SQ Magazine, 2026

• The best model on MedHallu (a 2025 benchmark built from 10,000 PubMedQA-derived question-answer pairs) reached only 0.625 F1 on the hard hallucination category — GPT-4o, Llama 3.1, and other leading models all struggled. — MedHallu benchmark, 2025

Legal Research

• 17%–88% — hallucination rates for AI on legal research queries, depending on model and query type. Even purpose-built legal AI tools (Lexis+ AI, Westlaw AI-Assisted Research) hallucinate at this range. — Stanford RegLab / Stanford HAI, 2024–2025

• Citation fabrication rates reach as high as 94% in adversarial testing of AI legal research tools. — benchmark analysis, cited in SQ Magazine 2026

General Knowledge & Citations

• ~18% of wrong answers on the MMLU benchmark are attributable to hallucination. — MMLU benchmark analysis, cited in SQ Magazine 2026

• TruthfulQA — once the gold standard for hallucination testing — has been partially compromised: a simple decision tree achieves 79.6% accuracy on its multiple-choice format without reading the questions, by exploiting structural patterns. It should no longer be cited as a reliable hallucination benchmark for 2025–2026 models. — Suprmind analysis, 2026

• A 2026 UC San Diego study found AI-generated product summaries hallucinated 60% of the time, influencing purchase decisions. — cited in SQ Magazine 2026

4. AI Coding Accuracy

• SWE-bench Verified average score: ~63.4% across 83 evaluated AI coding models as of April 2026, up from roughly 40% at end of 2024 and near 0% at the start of 2024. — BenchLM.ai / Scale AI SEAL Leaderboard, April 2026

• 77.2% — Claude 4 Sonnet's SWE-bench Verified score as of October 2025; GPT-5 reached 74.9% at the same time. — CodingFleet, 2026

• SWE-bench Verified has known contamination issues: OpenAI's internal audit found verbatim text overlap between frontier models and benchmark problems, indicating partial memorisation. OpenAI stopped reporting Verified scores in early 2026 and now recommends SWE-bench Pro. — CodingFleet / OpenAI, February 2026

• SWE-bench Pro (1,865 tasks across 41 repos, 123 languages) is the current harder standard. Even the best AI systems resolve only a small fraction of Pro tasks in a single run, highlighting that fully autonomous software engineering remains unsolved. — Scale AI / ICLR 2026

• AI-generated code runs at least 3x slower and uses far more memory than human-written solutions in controlled performance tests, even when functional correctness is similar. — International AI Safety Report, 2025

• AI coding models show 53% accuracy on medium-difficulty tasks and 0% on hard tasks on LiveCodeBench Pro — a contamination-resistant benchmark — when external tools are unavailable. — LiveCodeBench Pro, cited in International AI Safety Report 2025

• GitHub Copilot users complete coding tasks 55.8% faster than without AI assistance. — GitHub research, 2024

5. AI Accuracy in Medical Diagnostics

• 96% — AI accuracy in diabetic retinopathy detection, outperforming specialists by more than 10 percentage points, per 2025 clinical trial syntheses. — Uvik / SQ Magazine compilations, 2025

• 93% — AI-powered cancer diagnostic tools match rate with expert tumour board recommendations. — Scispot / clinical research compilations, 2026

• 90%+ pooled sensitivity and specificity for AI fracture detection on radiographs, per meta-analyses published 2022–2024. Reader studies show improved sensitivity without loss of specificity when radiologists use AI. — NCBI / PMC review, 2025

• 52.1% — overall diagnostic accuracy of generative AI models across a meta-analysis of 83 studies (2025 JMIR Medical Informatics systematic review), comparable to non-expert physicians but significantly lower than expert physicians. — JMIR Medical Informatics, 2025

• 93% on USMLE Step 2 CK — DeepSeek's score on the United States Medical Licensing Examination, outperforming ChatGPT and other models tested. — NCBI / PMC, 2026

• Over 1,250 AI- or ML-enabled medical devices had been cleared or approved by the U.S. FDA by May 2025, with radiology dominating the regulatory landscape. — Uvik, citing FDA data, 2026

• AI medical scribes allow physicians to spend up to 83% less time writing notes, according to multiple hospital system reports. — Stanford HAI AI Index 2026

• AI-powered imaging solutions are expected to prevent up to 2.5 million diagnostic errors annually. — Frost & Sullivan, cited in OneReach.ai 2026

6. AI Accuracy in Customer Service & Voice

• 92%–96% — call resolution accuracy for well-configured AI voice agents on standard business scenarios (booking, information, routing). — AInora, April 2026

• 97%+ — speech recognition accuracy for English in 2026 production voice AI deployments; 94%+ for most European languages. — Nuance/Microsoft benchmark / AInora, 2025–2026

• 87% — caller intent detection accuracy across all industries for AI voice agents in 2025–2026, rising to 94% in domains with well-trained knowledge bases. — Nuance/Microsoft benchmark, cited in AInora 2026

• 71% of callers could not reliably distinguish between AI and a human receptionist in blind tests using 2026-generation voice synthesis. — University of Michigan HCI Lab, 2025

• 680ms — median end-to-end response latency (caller finishes speaking to AI begins responding) for production voice agents in 2026, down from 1,200ms in 2024. — Retell AI benchmark data, 2026

• 89% — average accuracy for AI triage systems in correctly categorising and routing support tickets in real time. — AllAboutAI compilations, 2026

• Top-quartile AI customer service deflection rate: 58.7%, with the bottom quartile at 22.4%. Year-over-year improvement was +9.6 percentage points against the 2025 median of 31.6%. — Salesforce State of Service 2026

• 55%–70% first contact resolution rates for AI-native customer service platforms, versus 14% for traditional self-service channels. — Lorikeet CX / Gartner, 2026

• 4.2% average call abandonment rate when AI answers within 2 seconds, compared to 23.7% when callers wait 30+ seconds on hold. — ContactBabel, 2025

Brilo AI's voice agents are built for real resolution, not just deflection. The platform handles inbound and outbound calls with caller intent detection, structured conversation flows, and CRM integration — achieving resolution rates consistent with the top-quartile benchmarks above. Plans start free at brilo.ai.

7. AI Accuracy: Trust, Oversight & Human Verification

• 27% of organisations review all AI-generated content before use; a similar share reviews less than 20% of outputs. The majority operate without consistent human verification. — McKinsey State of AI, 2025

• 85% of consumers double-check AI answers against other sources before acting on them. — Eight Oh Two study, cited in Instant Press 2026

• A 2025 MIT Media Lab study found that people significantly overtrust AI-generated medical advice despite low accuracy — a documented mismatch between perceived and actual AI reliability. — MIT Media Lab, 2025, cited in multiple sources

• 48% of business leaders are confident in AI's accuracy for personalising customer service responses. — Master of Code / industry surveys, 2026

• 92% of businesses report improved CSAT after implementing AI customer service — suggesting that well-implemented AI satisfies customers regardless of stated accuracy concerns. — multiple surveys, 2026

• 362 documented AI incidents in 2025, up 56% from 233 in 2024. The rise reflects both increased deployment and improved incident documentation. — AI Incident Database / Stanford HAI AI Index 2026

8. How to Interpret AI Accuracy Statistics

No single benchmark is definitive. MMLU measures breadth of general knowledge, SWE-bench measures real-world coding ability, and Vectara measures hallucination on summarisation — none captures overall AI accuracy. Citing one benchmark number as "the" accuracy rate of an AI is misleading.

Task type dominates. The gap between grounded summarisation (0.7% hallucination) and open-ended generation (40–80%) or legal research (17–88%) reflects the architecture of the problem, not just model quality. Match the task to the model's documented strengths.

Contamination affects coding benchmarks. SWE-bench Verified has known contamination — models partially memorise benchmark answers during training. Scores on contamination-resistant benchmarks (LiveCodeBench Pro, SWE-bench Pro) are significantly lower.

Reasoning models trade accuracy for calibration. OpenAI o3 and o4-mini hallucinate 33–48% on person-specific questions despite high scores on reasoning tasks. Extended thinking improves some accuracy metrics while worsening others.

Mitigation works but is underused. Structured prompts reduce medical hallucinations by 33%; retrieval-augmented generation brings hallucination rates on summarisation tasks to under 2%. Only 27% of organisations consistently review AI outputs before use.