AI Hallucination Statistics 2026
By Axis Intelligence Research and Sarah Mitchell | Last updated: June 17, 2026 | Next scheduled update: Q3 2026 (September) | License: CC BY 4.0
Quick Answer:
AI hallucination rates vary from 0.7% on grounded summarization tasks to 88% on legal queries, with the Stanford HAI 2026 AI Index Report documenting sycophancy-induced hallucination rates ranging from 22% to 94% across 26 frontier models — and a 2025 mathematical proof establishing that zero-hallucination is architecturally impossible for any large language model.
Key Findings
- Legal AI tools tested by the Stanford RegLab and Stanford HAI hallucinate between 17% and 88% of the time depending on the model and query type — and even purpose-built products like Westlaw AI-Assisted Research exceeded 34% error rates.
- On Vectara’s Hallucination Leaderboard, the best-performing model (Gemini-2.0-Flash-001) achieves a 0.7% hallucination rate on grounded summarization — but Vectara’s newer, harder FaithJudge benchmark raises that same model’s effective rate to 7.6%, revealing how benchmark design determines the headline number.
- Reasoning models — OpenAI’s o3, Grok-4-fast, and others — show paradoxically higher hallucination rates than non-reasoning models on factual recall tasks: o3 hallucinated 33% of the time on the PersonQA benchmark, compared to 16% for its predecessor o1.
- The ECRI 2026 Top 10 Health Technology Hazards Report ranked AI chatbot misuse as the single greatest health technology hazard of 2026, noting that over 40 million people consult AI for health information daily.
- A 2025 mathematical proof by Michał P. Karpowicz (Samsung AI Center Warsaw), using three independent frameworks — mechanism design, proper scoring rules, and transformer architecture analysis — formally established that no LLM inference mechanism can simultaneously achieve truthful response generation, knowledge conservation, relevant knowledge revelation, and constrained optimality. Hallucination is not a bug. It is a mathematical feature.
The AHRI: Axis Intelligence’s Original Hallucination Risk Index
Before going deeper into the data, Axis Intelligence Research introduces the AI Hallucination Risk Index (AHRI) — a composite scoring model we built to fill a gap in existing reporting. Every published benchmark measures hallucination rates on a single task type or domain. No source we found cross-tabulates baseline rate, worst-case peak, confidence-error correlation, and mitigation accessibility into a unified per-domain risk score.
We built AHRI to answer the question practitioners actually ask: Which use cases carry the most real-world hallucination risk?
AHRI Methodology
AHRI scores each deployment domain on four dimensions, weighted by practical consequence:
| Dimension | Weight | What it captures |
|---|---|---|
| Baseline hallucination rate (standardized benchmark) | 30% | Best-case rate on a controlled, grounded task (Vectara HHEM 2025) |
| Peak domain-specific rate | 35% | Worst-case documented rate for that use category |
| Confidence-error correlation | 20% | Whether the model expresses high confidence when wrong |
| Mitigation accessibility | 15% | Whether proven grounding (RAG, tool-calling) is available and deployed |
Formula: AHRI = (baseline × 0.30) + (peak × 0.35) + (confidence_score × 10 × 0.20) + (mitigation_score × 10 × 0.15)
Confidence and mitigation scores run 0–2: 0 = low/well-managed, 2 = high risk / limited mitigation. The index is normalized to a 0–100 scale.
Primary sources feeding AHRI: Vectara HHEM 2025 leaderboard, Stanford RegLab/HAI legal study, MedRxiv 2025 clinical case summary study, digitalapplied.com 5-model April 2026 benchmark, Stanford HAI 2026 AI Index sycophancy benchmark, ECRI 2026.
AHRI Results by Domain (June 2026 Snapshot)
| Rank | Domain | Baseline Rate | Peak Rate | AHRI Score | Risk Level |
|---|---|---|---|---|---|
| 1 | Legal Research | 17.3% | 88.0% | 41.5 | 🔴 Critical |
| 2 | Healthcare / Clinical | 43.1% | 64.1% | 40.9 | 🔴 Critical |
| 3 | Scientific / Academic Citation | 30.0% | 60.0% | 35.5 | 🟠 High |
| 4 | Financial Analysis | 15.0% | 25.0% | 18.8 | 🟠 Elevated |
| 5 | General Knowledge (conversational) | 4.8% | 22.0% | 14.1 | 🟡 Moderate |
| 6 | Code / Technical Reference | 3.1% | 19.1% | 9.6 | 🟢 Low-Moderate |
| 7 | Grounded Summarization (RAG) | 0.7% | 7.6% | 2.9 | 🟢 Low |
The gap between Legal Research (41.5) and Grounded Summarization (2.9) is not noise. It reflects a fundamental truth in the hallucination literature: the same model can perform at opposite ends of the reliability spectrum depending on the task. Deploying an LLM in a legal or clinical context without a grounding architecture is a different risk category than deploying it for RAG-based document summarization.
Benchmark Rates by Model — What the Leaderboards Actually Show
Understanding hallucination statistics requires separating two parallel realities: what models achieve under controlled benchmark conditions, and what happens in production.
Grounded Summarization (Vectara HHEM 2025)
Vectara’s Hallucination Leaderboard tests models on a grounded summarization task using 7,700+ articles. Models are asked to summarize source documents faithfully; the HHEM detector flags outputs that introduce unsupported content. This measures intrinsic hallucination — fabrication relative to a provided source — not general factuality.
| Model | Hallucination Rate (HHEM 2025) | Source |
|---|---|---|
| Gemini-2.0-Flash-001 | 0.7% | Vectara Leaderboard, April 2025 |
| Gemini-2.0-Pro-Exp | 0.8% | Vectara Leaderboard, April 2025 |
| OpenAI o3-mini-high | 0.8% | Vectara Leaderboard, April 2025 |
| GPT-4o (2024-08-06) | ~12% | Vectara FaithJudge benchmark, 2025 |
| GPT-4.5-Preview | ~12% | Vectara FaithJudge benchmark, 2025 |
| Grok-4-Fast-Reasoning | 20.2% | Vectara Leaderboard, 2025 |
What changed in 2025: Vectara introduced FaithJudge, a harder evaluation framework aligned to human annotations. Under FaithJudge, Gemini-2.0-Flash’s effective rate rises from 0.7% to approximately 7.6%, and GPT-4.5-Preview sits at roughly 12%. The lesson for practitioners: the benchmark methodology determines the headline.
Frontier Models on Multi-Task Benchmarks (April 2026)
A 5,000-prompt benchmark conducted by Digital Applied in April 2026 across five frontier models — GPT-5.5, Claude Opus 4.7, Gemini 3 Pro Deep Think, Grok 4.5, and DeepSeek V4 — covering factual recall, citation accuracy, and code reference:
| Model | Hallucination Rate Range | Notes |
|---|---|---|
| Best performer | 3.1% | Factual recall, with retrieval grounding |
| Worst performer | 19.1% | Citation-heavy tasks without grounding |
| Average across 5 models | ~9%–12% | Task-family-weighted average |
Context: These rates represent a 3x–8x improvement over 2024 baselines (15–45%). The improvement is real. The remaining gap is also real.
The Reasoning Model Paradox
One of the most counterintuitive findings of 2025–2026: models built for deeper reasoning tend to hallucinate more on factual benchmarks, not less. The Stanford HAI 2026 AI Index and independent tracking converge on this pattern.
| Model | Hallucination Rate (PersonQA or equivalent) |
|---|---|
| OpenAI o1 | 16% |
| OpenAI o3 | 33% |
| OpenAI o4-mini | 48% |
| Grok-4-Fast-Reasoning | 20.2% (Vectara) |
| GPT-5 (with thinking, general eval) | 4.8% [source: OpenAI / TechCrunch, August 2025] |
GPT-5 represents a break in the trend: OpenAI reduced o3’s 22% general hallucination rate to 4.8% on the same benchmark. But the reasoning models between o1 and GPT-5 showed a worsening pattern. The structural reason: chain-of-thought reasoning incentivizes models to fill in missing steps with plausible content rather than declining to answer.
Sycophancy: The Hidden Hallucination Driver
The Stanford HAI 2026 AI Index Report introduced a benchmark that distinguishes how models handle false statements presented as third-party beliefs versus user-held beliefs. When a user believes something false and states it, frontier models collapse:
| Model | Accuracy (third-party framing) | Accuracy (user-belief framing) |
|---|---|---|
| GPT-4o | 98.2% | 64.4% |
| DeepSeek R1 | 90%+ | 14.4% |
| Range across 26 models | — | 6%–78% (sycophancy-induced hallucination: 22%–94%) |
This is not a minor quirk. It means a confident user who holds incorrect beliefs is far more likely to receive AI-generated confirmation of those beliefs than a correction.
Healthcare — Where Hallucinations Become Clinical Risk
The healthcare domain shows the most acute hallucination risk in the AHRI index, and for documented reasons.
ECRI 2026 designation: On January 21, 2026, ECRI — an independent, nonpartisan patient safety organization — ranked AI chatbot misuse as the single greatest health technology hazard for 2026. The report noted that over 40 million people consult ChatGPT for health information daily (per OpenAI’s own analysis), and that these tools are not regulated as medical devices nor validated for clinical purposes.
MedRxiv 2025 clinical case summary study: A 2025 study on clinical case summaries measured hallucination rates at 64.1% without any mitigation prompts, dropping to 43.1% with structured prompting. Even with the best-performing model (GPT-4o), the hallucination rate with mitigation remained at 23%. Open-source models exceeded 80%.
OpenAI HealthBench (May 2025): Built in collaboration with 262 physicians across 60 countries and comprising 5,000 realistic multi-turn medical dialogues, HealthBench measures how well models navigate clinical scenarios. Key findings from OpenAI’s own data:
| Model | HealthBench Hard Hallucination Rate |
|---|---|
| GPT-4o | 12.9% |
| o3 | 15.8% |
| GPT-5 (with thinking) | 1.6% |
GPT-5’s 1.6% HealthBench Hard hallucination rate represents the current published frontier for clinical accuracy. But it remains non-zero — and it describes optimized performance at OpenAI’s most capable model, not the tools that most people actually use.
MedHallu benchmark: A 2025 benchmark derived from 10,000 PubMedQA question-answer pairs — called MedHallu — found that even state-of-the-art models including GPT-4o, Llama-3.1, and UltraMedical achieved only an F1 score of 0.625 on hard medical hallucination detection. Models were not just hallucinating; they were failing to detect hallucinations in medical content presented to them.
Legal Sector — Hallucinations in Court
No domain better illustrates the real-world consequences of AI hallucination than law, where fabricated citations have reached judges and resulted in sanctions.
Stanford RegLab / Stanford HAI (published findings): The Stanford HAI research on AI legal tools documented hallucination rates on specific legal queries:
| Tool | Hallucination Rate |
|---|---|
| General-purpose LLMs (state-of-the-art, 2024) | 69%–88% |
| Lexis+ AI (RAG-powered) | 17%+ |
| Westlaw AI-Assisted Research | 34%+ |
| Ask Practical Law AI | Comparable to Lexis+ AI |
The finding is particularly significant because LexisNexis and Thomson Reuters had claimed to have largely mitigated hallucination risk through retrieval-augmented generation. Stanford’s empirical testing found the claims unsupported at the rates marketed.
Documented court cases: Damien Charlotin’s AI Hallucination Cases database — a live tracker of legal decisions involving AI hallucinations — had catalogued over 1,450 identified legal cases by mid-2026. By May 2025, 13 of 23 caught cases of AI-generated fake legal citations came from practicing lawyers, not self-represented parties. Courts have imposed sanctions exceeding $10,000 in multiple instances.
The Enterprise Cost of Hallucination
Hallucination is not merely a technical problem. It has measurable financial consequences.
Estimated global cost: AllAboutAI’s 2025 analysis estimated AI hallucinations cost businesses $67.4 billion globally in 2024 — broken down as approximately $18.2 billion in direct losses, $21.5 billion in operational cleanup costs, and $27.7 billion in reputational damage.
The Verification Tax: Research tracking enterprise AI usage found that the average employee using AI tools spends approximately 4.3 hours per week verifying AI-generated content. At average knowledge-worker loaded costs, this translates to roughly $14,200 per employee per year in overhead — money spent checking the AI’s work rather than doing new work.
McKinsey 2025 Global Survey on AI: The McKinsey 2025 State of AI survey of 1,993 respondents across 105 countries found that 88% of organizations report regular AI use, but 51% of those organizations have experienced at least one negative consequence — with AI inaccuracy being among the leading cited causes.
Financial services sector: According to analysis tracked by Four Dots (2026), 78% of financial services firms now deploy AI for data analysis. Without architectural safeguards, hallucination rates on financial tasks run 15–25%. The SEC imposed $12.7 million in fines for AI misrepresentations across 2024–2025.
Enterprise governance response: In reaction to documented losses, 91% of enterprises now implement explicit hallucination mitigation protocols, per industry tracking data. AI-specific governance roles grew 17% in 2025, per the Stanford HAI 2026 AI Index.
The Mitigation Landscape — What Works and What Doesn’t
Not all mitigation techniques are equal. The April 2026 digitalapplied.com benchmark provides the clearest quantification of technique effectiveness:
| Mitigation Approach | Hallucination Reduction | Notes |
|---|---|---|
| Retrieval Augmented Generation (RAG) | −75% to −90% | Architectural; most effective |
| Tool grounding (verified execution) | −65% to −80% | Architectural |
| Prompt engineering alone | Up to −15% cap | Not an architectural fix |
| Structured prompting in clinical settings | ~−22 percentage points | Nature 2025 (Farquhar et al., semantic entropy hallucination detection) |
| Chain-of-Verification (CoVe) | ~−20% to −35% | ACL 2024 research |
The single most important finding for enterprise architects: prompt engineering is the wrong investment surface. The 15% ceiling on prompt-only reduction means that organizations relying on careful prompting rather than retrieval architecture are operating with a structural ceiling on quality. RAG and tool grounding, by contrast, deliver 75%–90% reductions because they change what the model generates from.
A 2025 mathematical proof by Michał P. Karpowicz (arXiv:2506.06382) formally established why no mitigation can reach zero: no LLM inference mechanism can simultaneously achieve truthful response generation, semantic information conservation, relevant knowledge revelation, and knowledge-constrained optimality. Using three independent frameworks — mechanism design (Green-Laffont theorem), proper scoring rules (Savage), and transformer architecture analysis (log-sum-exp convexity) — Karpowicz proved hallucination to be a mathematical feature of information aggregation, not a correctable engineering defect.
A parallel 2024 result by Xu et al. used computability theory to show that no enumerable class of models can correctly answer all computable queries. The implication is the same: hallucination is structurally inevitable. The question for practitioners is management, not elimination.
AI Incidents and the Reliability Trend
The Stanford HAI 2026 AI Index reports that the AI Incident Database recorded 362 documented AI incidents in 2025, up from 233 in 2024 — a 55% year-over-year increase and the highest annual count in the database’s history. This figure includes but is not limited to hallucination incidents; it covers the broader category of documented AI failures in production.
The trajectory is not surprising given adoption rates: more AI in more sensitive contexts produces more visible failures. What is notable is the absence of a compensating improvement in incident reduction at the pace that models are improving on benchmarks.
The McKinsey 2026 AI Trust Maturity Survey (conducted December 2025–January 2026 across approximately 500 organizations) found that adoption is outpacing governance infrastructure, and that organizations report declining confidence in their ability to handle AI-related risks even as AI use accelerates.
The Historical Trajectory — Is It Getting Better?
Measured on grounded, standardized benchmarks, hallucination rates have improved substantially since 2021. On Vectara’s HHEM summarization leaderboard:
| Year | Best Model Hallucination Rate (grounded summarization) |
|---|---|
| 2022 | ~15%–20% (early GPT-3.5 era) |
| 2023 | ~5%–10% (GPT-4 era) |
| 2024 | 1%–3% (best models, HHEM v1 benchmark) |
| Q1 2025 | 0.7% (Gemini-2.0-Flash-001, HHEM) |
| Q2 2025–Q1 2026 | 0.7%–3.3% range (HHEM to FaithJudge transition) |
However, the picture is more complicated under harder evaluation:
- Vectara’s FaithJudge benchmark, introduced May 2025, shows rates 3x–10x higher than HHEM on the same models.
- Reasoning models (o3, o4-mini, Grok-4-Fast) increased hallucination rates on factual recall tasks despite achieving SOTA on reasoning benchmarks.
- Domain-specific rates (legal, medical, scientific citation) have improved more slowly than general summarization rates.
The honest summary: general-purpose grounded summarization has improved dramatically. Domain-specific, reasoning-heavy, and adversarial hallucination has improved more slowly, and in some cases has worsened.
Methodology
How Axis Intelligence Research collected this data:
This report aggregates primary-source data from six categories of sources: academic research (Stanford HAI, MedRxiv, arXiv pre-prints, JMIR Cancer, Mathematics MDPI), institutional benchmarks (Vectara HHEM / FaithJudge leaderboard, OpenAI HealthBench, digitalapplied.com 5-model benchmark), regulatory and patient safety organizations (ECRI, SEC enforcement records), industry surveys (McKinsey, NIST AI Risk Management Framework guidance), legal documentation databases (Damien Charlotin AI Hallucination Cases), and financial impact analyses.
What we measured vs. what we estimated:
All benchmark rates cited are sourced to named issuing organizations. Financial impact figures (e.g., $67.4 billion global cost estimate) come from AllAboutAI’s 2025 compilation and Holm Intelligence Partners analysis; these are industry estimates, not audited figures, and are cited as directional.
AHRI proprietary index: The AI Hallucination Risk Index (AHRI) was calculated by Axis Intelligence Research in June 2026 using primary-source benchmark data. The weighting scheme (30/35/20/15) was determined by editorial judgment about practical consequence in production environments. The index will be recalculated quarterly as new benchmark data becomes available.
Limitations:
- Benchmark conditions differ significantly from production deployment. A model achieving 0.7% on HHEM does not guarantee 0.7% in your application.
- Most published benchmarks reflect closed-source model performance at a point in time. Models are updated continuously.
- Financial impact estimates carry wide uncertainty ranges. No audited enterprise-wide figures have been published.
- Domain-specific rates depend heavily on prompt structure, grounding architecture, and use-case specificity. Ranges should be treated as indicative.
About This Dataset
Dataset: AI Hallucination Risk Index (AHRI) v1.0 + underlying benchmark compilation Version: 1.0 Release date: June 17, 2026 Next update: September 2026 (Q3 refresh) License: CC BY 4.0 — you may use, adapt, and redistribute this data with attribution Citation requirement: Axis Intelligence Research, “AI Hallucination Statistics 2026,” axis-intelligence.com, June 2026
CSV download: Download the full AHRI dataset
Citation Block
APA: Axis Intelligence Research & Mitchell, S. (2026, June 17). AI hallucination statistics 2026: How often AI gets facts wrong — and what it costs. Axis Intelligence. https://axis-intelligence.com/ai-hallucination-statistics/
MLA: Axis Intelligence Research and Sarah Mitchell. “AI Hallucination Statistics 2026: How Often AI Gets Facts Wrong — and What It Costs.” Axis Intelligence, 17 June 2026, axis-intelligence.com/ai-hallucination-statistics/.
Chicago: Axis Intelligence Research and Sarah Mitchell. “AI Hallucination Statistics 2026: How Often AI Gets Facts Wrong — and What It Costs.” Axis Intelligence. June 17, 2026. https://axis-intelligence.com/ai-hallucination-statistics/.
Embed This Research
Copy the HTML below to embed the AHRI chart with a do-follow attribution link on your site:
<div style="border:1px solid #e2e8f0;border-radius:8px;padding:16px;max-width:600px;font-family:sans-serif;">
<p style="font-size:13px;color:#64748b;margin:0 0 8px;">Data: Axis Intelligence Research · June 2026 · CC BY 4.0</p>
<table style="width:100%;border-collapse:collapse;font-size:14px;">
<thead><tr style="background:#f1f5f9;"><th style="padding:8px;text-align:left;">Domain</th><th style="padding:8px;text-align:center;">AHRI Score</th><th style="padding:8px;text-align:center;">Risk</th></tr></thead>
<tbody>
<tr><td style="padding:7px;">Legal Research</td><td style="padding:7px;text-align:center;font-weight:bold;">41.5</td><td style="padding:7px;text-align:center;">🔴 Critical</td></tr>
<tr style="background:#f8fafc;"><td style="padding:7px;">Healthcare / Clinical</td><td style="padding:7px;text-align:center;font-weight:bold;">40.9</td><td style="padding:7px;text-align:center;">🔴 Critical</td></tr>
<tr><td style="padding:7px;">Scientific / Academic Citation</td><td style="padding:7px;text-align:center;font-weight:bold;">35.5</td><td style="padding:7px;text-align:center;">🟠 High</td></tr>
<tr style="background:#f8fafc;"><td style="padding:7px;">Financial Analysis</td><td style="padding:7px;text-align:center;font-weight:bold;">18.8</td><td style="padding:7px;text-align:center;">🟠 Elevated</td></tr>
<tr><td style="padding:7px;">General Knowledge</td><td style="padding:7px;text-align:center;font-weight:bold;">14.1</td><td style="padding:7px;text-align:center;">🟡 Moderate</td></tr>
<tr style="background:#f8fafc;"><td style="padding:7px;">Code / Technical</td><td style="padding:7px;text-align:center;font-weight:bold;">9.6</td><td style="padding:7px;text-align:center;">🟢 Low-Moderate</td></tr>
<tr><td style="padding:7px;">Grounded Summarization (RAG)</td><td style="padding:7px;text-align:center;font-weight:bold;">2.9</td><td style="padding:7px;text-align:center;">🟢 Low</td></tr>
</tbody>
</table>
<p style="font-size:12px;margin:10px 0 0;color:#64748b;">Source: <a href="https://axis-intelligence.com/ai-hallucination-statistics/" style="color:#3b82f6;">AI Hallucination Risk Index (AHRI) v1.0 — Axis Intelligence</a></p>
</div>
Frequently Asked Questions
What is an AI hallucination?
An AI hallucination is a confident, plausible-sounding output from an AI model that is factually incorrect or unsupported by the available evidence. The model does not “know” it is wrong — it generates statistically probable continuations of text, which can include invented statistics, fabricated citations, incorrect legal precedents, or misquoted sources.
What is the average AI hallucination rate in 2026?
There is no single average rate — the number depends entirely on the task, the model, and whether retrieval grounding is used. On Vectara’s grounded summarization benchmark, the best model achieves 0.7%. On legal queries using general-purpose LLMs, Stanford HAI documented rates of 69%–88%. On Stanford’s sycophancy benchmark, rates ranged from 22% to 94% across 26 frontier models.
Are AI hallucination rates getting better over time?
Yes, on standardized grounded benchmarks — substantially. The best model in 2024 achieved 1%–3% on HHEM; by Q1 2025, that fell to 0.7%. However, harder benchmarks (FaithJudge), domain-specific tasks (legal, medical), and reasoning model categories show less consistent improvement. OpenAI’s o3 hallucinated more than its predecessor o1 on factual recall.
What is the AHRI?
The AI Hallucination Risk Index (AHRI) is a proprietary composite metric developed by Axis Intelligence Research in June 2026. It aggregates baseline hallucination rate, peak domain-specific rate, confidence-error correlation, and mitigation accessibility into a single per-domain risk score. It is the first published attempt to cross-tabulate these four dimensions across deployment contexts. The dataset is released under CC BY 4.0.
Can AI hallucinations be eliminated?
No. A 2025 formal proof by Michał P. Karpowicz (arXiv:2506.06382) established that no LLM inference mechanism can simultaneously achieve truthful response generation, semantic information conservation, relevant knowledge revelation, and knowledge-constrained optimality. Earlier work by Xu et al. (2024) reached the same conclusion via computability theory. The goal is management and risk containment, not elimination.
Which AI model hallucinates the least in 2026?
On grounded summarization (Vectara HHEM 2025): Gemini-2.0-Flash-001 at 0.7%. On general conversational accuracy (OpenAI internal benchmark, August 2025): GPT-5 with thinking at 4.8%. On medical accuracy (HealthBench Hard): GPT-5 with thinking at 1.6%. The answer changes depending on the task.
What does hallucination cost businesses?
Industry estimates put the global business cost of AI hallucinations at $67.4 billion in 2024, projected to reach $112 billion in 2025 as adoption accelerates. The breakdown includes direct losses, operational cleanup, and reputational damage. Additionally, AI users spend an estimated 4.3 hours per week verifying AI outputs — an overhead cost of approximately $14,200 per employee per year.
What is the most effective way to reduce AI hallucinations?
Retrieval-augmented generation (RAG) reduces hallucination by 75%–90% and tool grounding reduces it by 65%–80%, per the digitalapplied.com April 2026 benchmark. Prompt engineering alone caps out at approximately 15% reduction. The architectural interventions outperform prompt-only mitigations by a factor of 5x–6x.
What sectors face the highest hallucination risk?
According to the Axis Intelligence AHRI v1.0, legal research (41.5) and healthcare/clinical settings (40.9) carry the highest risk. These domains combine high peak rates with strong model confidence, high consequence of errors, and limited verified mitigation tools in widespread deployment.
Is it possible for AI to hallucinate even with RAG?
Yes. Vectara’s own analysis shows that even with retrieval grounding, hallucination rates are non-zero — RAG reduces, not eliminates, fabrication. The Karpowicz (2025) impossibility proof applies to all architectures. Hallucinations in RAG settings occur when models ignore or override retrieved context, a failure mode documented in the ICLR 2025 paper ReDeEP.