In October 2025, the Australian government received a $440,000 report from consulting firm Deloitte — containing non-existent academic sources, fabricated citations, and a fake quote from a federal court judgement. The company submitted a revised version and issued a partial refund. The following month, Deloitte's $1.6 million health report for the Canadian government of Newfoundland and Labrador was found to contain at least four false citations to research papers that do not exist.

This is what AI hallucination looks like in the real world. Not science fiction — consulting invoices, government reports, and legal briefs contaminated by AI-generated fiction delivered with complete confidence. It's the defining reliability problem of the current AI era, and understanding it is no longer optional for anyone building with or making decisions based on AI outputs.

"Language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty." — OpenAI Research Paper, September 2025 (Kalai, Nachum, Vempala & Zhang)

The 2025 research consensus has shifted significantly. Hallucinations are no longer seen simply as a data problem or a model size problem. They are now understood as a systemic incentive problem — baked into the way models are trained and evaluated — with multiple contributing causes that interact in complex ways. Here's the full picture.

01 What Is AI Hallucination, Exactly?

The term "hallucination" refers to outputs from AI models that are fluent, confident, and plausible-sounding — but factually incorrect, unfaithful to the source, or entirely fabricated. The word is somewhat contested: some researchers prefer "confabulation" (borrowed from neuroscience, describing the brain's tendency to fill memory gaps with invented content), while others — including Rebecca Parsons, CTO of ThoughtWorks — argue that all LLM outputs are hallucinations; it's just that some of them happen to be true.

There are two main categories that matter in practice:

Factual hallucination: The model states something about the world that isn't true. A citation to a paper that doesn't exist. A statistic that was never measured. A historical event that didn't happen. A person's birthday stated with confidence despite the model never having seen it.

Faithfulness hallucination: The model's output contradicts or goes beyond its input. You give it a document to summarise and it adds facts not in the document. You ask it to answer based only on a provided context and it draws on training data instead. This category is especially dangerous in RAG (retrieval-augmented generation) systems where you assume the model is grounded in real source material.

02 The Five Root Causes

Hallucinations don't have a single cause. They're the product of multiple compounding factors — each contributing in different proportions depending on the task, the model, and the domain. Here are the five that matter most.

Visual 01
The Five Root Causes of AI Hallucination
Noisy & Incomplete Training Data
Web-scale corpora inevitably contain false information, outdated facts, and contradictions. The model learns statistical patterns across this data — including its errors. For rare topics, where the training signal is thin, errors compound dramatically. A model that saw 1,000 sentences about a famous scientist and two about an obscure one will be far more reliable on the former.
The Incentive Problem: Rewarding Confident Guesses
The most important 2025 discovery. Standard accuracy-based benchmarks reward a model that guesses "September 10" as a birthday (1/365 chance of being right) over one that says "I don't know" (guaranteed zero). Over thousands of evaluations, models learn to bluff. Saying "I don't know" is never rewarded — so models don't learn to say it.
Probabilistic Next-Token Prediction
At its core, an LLM is predicting the most statistically likely next word. This works brilliantly for language fluency. It works poorly for rare facts, because the "most likely" continuation of a sentence about a specific event may be plausible-sounding fiction rather than verifiable truth. The model has no ground-truth oracle — only pattern matching.
Faulty Circuit Inhibition (Anthropic, 2025)
Interpretability researchers at Anthropic identified specific internal circuits in Claude that normally prevent the model from answering when it lacks sufficient information. Hallucinations occur when these "refusal circuits" are incorrectly inhibited — for example, when the model recognises a person's name (triggering a response) but lacks sufficient facts about them (leading to fabrication).
Sycophancy & User Pressure
RLHF-trained models learn that agreeing with users gets positive feedback. If a user confidently asserts a wrong "fact" and asks the model to confirm, many models will agree rather than correct — even when their training data contradicts it. This form of socially-induced hallucination compounds factual errors in real conversations. OpenAI's GPT-4o sycophancy incident in May 2025 — where the model validated users' delusional beliefs and was subsequently rolled back — became a watershed moment for the field, forcing explicit anti-sycophancy objectives into RLHF pipelines across all major labs.

The causes are cumulative — a rare topic (cause 1) + accuracy-rewarding benchmark (cause 2) + probabilistic generation (cause 3) = extremely high hallucination risk. High-stakes domains like law, medicine, and science trigger all five simultaneously.

03 How Bad Is It? Real Hallucination Rates

Hallucination rates vary enormously by model, task, and domain. The headline figures from the 2025 AI Hallucination Report give the clearest picture yet — and show both how far the field has come and how far it still has to go.

Visual 02
Hallucination Rates by Domain — Best-in-Class Models vs. Average (2025)
General QA (best)
0.7%
General QA (avg)
9.2%
Financial Data
2.1–13.8%
Medical Info
4.3–15.6%
Legal Research
6.4–18.7%
Scientific Papers
3.7–16.9%
o3/o4-mini PersonQA
33–48%

Sources: AI Hallucination Report 2025 (AllAboutAI). Gemini 2.0 Flash currently holds the lowest rate at 0.7%. The o3 and o4-mini PersonQA results (33–48%) illustrate how reasoning-optimised models can trade factual reliability for reasoning performance on certain tasks.

Perhaps the most striking finding: four models now have sub-1% hallucination rates on general factual tasks — a milestone that seemed distant just two years ago. Yet on specialised tasks and rare entities, even the best models remain unreliable. Dario Amodei suggested at a 2025 Anthropic developer event that frontier models may already hallucinate less than humans on certain factual tasks — a provocative claim, but one that illustrates how the goalposts are shifting from "eliminate hallucination" to "manage uncertainty predictably."

04 Real-World Consequences

Visual 03
Documented High-Stakes Hallucination Incidents (2023–2025)
Incident Year Consequence Severity
Air Canada chatbot — fabricated bereavement refund policy, court ruled airline liable 2024 Legal liability established; chatbot "separate entity" defence rejected High
Deloitte Australia report — $440K government report contained fake citations and fabricated court quote 2025 Revised report submitted; partial refund issued High
Deloitte Canada — $1.6M health report had 4+ non-existent research paper citations 2025 Public disclosure; regulatory scrutiny of AI use in consulting High
Legal filings (multiple cases) — lawyers submitted briefs citing cases that don't exist 2023–25 Court sanctions; professional discipline proceedings High
Medical chatbots — incorrect drug dosage information in clinical assistant tools 2024–25 Product recalls; mandatory human review protocols introduced High
ChatGPT sycophancy (4o) — model validated users' delusional beliefs; public concern raised 2025 OpenAI rollback; NYT coverage; RLHF sycophancy now an active research area Medium

The pattern is consistent: AI outputs trusted without verification in high-stakes domains. The solution is not avoiding AI — it's building verification into every workflow.

05 Every Proven Technique to Reduce Hallucination

The 2025 research consensus is clear: you cannot eliminate hallucination, but you can reduce it dramatically through a layered approach combining architectural choices, training methods, deployment patterns, and prompt engineering.

Visual 04
Hallucination Reduction Techniques — Layered Approach
Retrieval-Augmented Generation (RAG)
Ground every response in retrieved, verifiable documents. The model answers from sources rather than memory. Most effective for factual, domain-specific tasks. Requires span-level citation and verification to prevent faithfulness hallucinations within retrieved content. Studies consistently show RAG reduces factual hallucination by 30–60% on knowledge-intensive tasks compared to closed-book generation.
Calibration-Aware Training
Penalise confident wrong answers more than uncertainty. Reward appropriate "I don't know" responses. OpenAI's 2025 paper argues this is the most fundamental fix — changing the incentive structure that causes hallucination in the first place.
Prompt Engineering
Explicit instructions dramatically reduce hallucination. Tell the model: "Only use facts from the provided document. If the answer isn't in the document, say so." Ask it to cite sources inline. Request confidence levels. Instruct it to express uncertainty rather than guess.
Chain-of-Thought Verification
Ask the model to reason step-by-step before answering, then check its own reasoning for contradictions. "Think through this carefully. Are you certain of each claim? If not, flag it." Self-consistency checking — sampling multiple answers and comparing — also reliably reduces hallucination on factual tasks. Research shows that asking a model to generate three independent answers and select the most consistent one reduces error rates by 15–30% on complex reasoning tasks.
Fine-Tuning on Domain Data
Training a model on high-quality, curated domain-specific data (medical, legal, financial) significantly reduces hallucination within that domain. Requires careful dataset curation and domain expertise, but produces the most reliable results for specialised applications.
Human-in-the-Loop Verification
For high-stakes outputs — legal, medical, financial — mandatory human review before any action is taken. Not a limitation to apologise for; the correct architecture for the current capability level. AI handles the draft; a qualified human handles verification.

No single technique is sufficient. Production deployments in high-stakes domains should combine at minimum: RAG (grounding) + prompt instructions (uncertainty expression) + human review (verification). For lower-stakes tasks, prompt engineering alone may be sufficient.

06 The Uncomfortable Truth: Will It Ever Be Zero?

The 2025 OpenAI research paper contains a mathematically sobering result: even with perfectly clean training data, hallucination cannot be fully eliminated. The proof is elegant and depressing. A model's text generation error rate must be at least twice its sentence classification error rate. Since some questions are inherently unanswerable from training data alone — birthdays of obscure individuals, results of unrecorded meetings, details that simply never appeared in text — there will always be a floor below which hallucination rates cannot fall.

Georgia Tech's Santosh Vempala puts it plainly: "If you go to a classroom with 50 students and you know the birthdays of 49 of them, that still gives you no help with the 50th. The reality is that we won't ever get to 100% accuracy."

This mathematical floor has a practical corollary: the rarer the entity or event, the higher the hallucination risk. A model asked about Abraham Lincoln will be far more reliable than one asked about a regional politician or a mid-tier research paper. The training corpus contains millions of reliable, cross-validated sentences about well-known figures and near-zero about obscure ones. When the model encounters a question about the latter, it is working almost entirely from generalised pattern-matching with very little actual signal — and no internal mechanism to distinguish "I know this" from "I'm pattern-matching a plausible-sounding answer."

This is why the PersonQA benchmark — which specifically tests models on questions about less-famous individuals — exposes hallucination rates 10–50× higher than general knowledge benchmarks. It strips away the comfortable familiarity of well-represented training data and forces the model into exactly the territory where hallucination is most likely. The o4-mini's 48% error rate on PersonQA isn't an anomaly; it's an honest look at what happens when a model is asked about things the internet simply hasn't written much about.

Research Finding

The Incentive Paradox

There's a commercially awkward dimension to the hallucination problem. AI researcher Wei Xing at the University of Sheffield observed: "Fixing hallucinations would kill the product." If a model responded "I don't know" to every genuinely uncertain question, users would find it useless and seek answers elsewhere.

OpenAI's own research acknowledges this tension. The fix — penalising confident errors more than uncertainty — requires rewriting evaluation benchmarks and accepting lower scores on conventional accuracy metrics. It's the right thing to do technically. It's a harder sell commercially when customers compare benchmark numbers.

The field is moving toward calibrated uncertainty as the new standard — models that express confidence levels, flag low-certainty answers, and proactively say "I'm not certain about this — you should verify." This is less satisfying than a confident answer, and more valuable than a confident wrong one.

07 What This Means For You Right Now

AI hallucination in 2026 is a known, measurable, partially manageable risk — not a mysterious defect or a reason to abandon AI tools entirely. The correct response is the same as the correct response to any known risk: understand it, design around it, and verify where it matters.

For developers building AI applications: RAG is not optional for factual applications — it is the baseline architecture. Prompt instructions that explicitly request uncertainty expressions cost nothing and reliably help. Observability tools that log model outputs for spot-checking are a minimum viable safety net.

For end users of AI tools: treat AI outputs in specialised domains (legal, medical, financial, academic) the way you'd treat a confident colleague who might be wrong. Verify citations independently. Cross-reference statistics. Never submit AI-generated professional content without fact-checking the specific claims that matter.

For the field overall: the 2025 shift toward calibration-aware training and uncertainty-friendly evaluation is the right direction. Four models already achieving sub-1% hallucination on general tasks demonstrates the problem is tractable. The work now is domain-specific reliability, sycophancy reduction, and building the evaluation infrastructure that incentivises honesty over confident bluffing — at every level of the training pipeline.

A Practical Checklist for Reducing Hallucination Risk

If you're building or deploying AI systems today, here's a concrete checklist. For factual applications (legal research, medical information, financial analysis): implement RAG with span-level citation so every claim can be traced to a source document. Add a verification prompt after generation — "Review your response and flag any claims you are uncertain about." Enable structured output so hallucinated figures can't hide inside prose.

For summarisation tasks: always instruct the model to stay faithful to the provided text and never add external information. Follow up with a faithfulness check: ask the model to identify any sentence in its summary that goes beyond what the source document says. Studies show this two-pass approach reduces faithfulness hallucinations by 40–60% on documents longer than 2,000 words.

For conversational AI: actively test for sycophancy before deploying. Present the model with incorrect "facts" stated confidently by a simulated user and check whether it agrees or corrects. Calibrate your system prompt to reinforce correction over agreement: "If the user states something factually incorrect, politely correct them with evidence rather than agreeing." Monitor output logs for hedging language — too little hedging on ambiguous topics is a hallucination risk signal.

The single most underused technique is simply asking the model to express uncertainty. A prompt that ends with "If you are not certain of any claim in your response, say so explicitly" reliably produces more accurate, better-hedged outputs — at zero cost. The model's uncertainty is already there; you just need to give it explicit permission to surface it.