OpenAI has reframed the reliability debate by arguing that AI hallucinations are not merely engineering bugs but a predictable outcome of how large language models learn and are evaluated. New data spanning multiple models and benchmarks shows error rates ranging from the mid-teens to nearly 80% on certain tasks, challenging the industry’s trust assumptions and elevating abstention and calibration from nice-to-have features to deployment prerequisites in regulated use cases.

The company’s latest analysis lands at a delicate moment: GPT‑5-class systems are demonstrably stronger than their predecessors yet still produce confident, incorrect statements that can mislead users and corrupt downstream processes. The new guidance doesn’t absolve teams from improving models, prompts, or retrieval—but it does demand a rethink of benchmarks, incentives, and product guardrails around AI hallucinations.

Key Takeaways

– Shows OpenAI measured o1 at 16%, o3 at 33%, and o4‑mini at 48% hallucination rates, affirming errors persist even as models advance. – Reveals task variability: o3 hit 33% on PersonQA but 51% on SimpleQA; o4‑mini rose from 41% to a striking 79%. – Demonstrates even GPT‑4.5 logged 37.1% hallucinations on SimpleQA, underscoring that benchmark choice dramatically changes failure profiles and perceived progress. – Indicates incentives matter; OpenAI warns training and evaluations reward guessing over abstention, recommending penalties for confident errors and explicit confidence targets. – Suggests safety needs update; GPT‑5 shows 16.5% deception in some tests and requires calibration, abstention metrics, and deliberative alignment to mitigate harm.

Why AI hallucinations persist even in GPT‑5

Hallucinations arise from a structural tension: language models generate the most probable next token, but the world is uncertain, knowledge is incomplete, and the incentive structures used to train and evaluate systems can reward plausible guesses over cautious abstentions. That mathematical reality means even very capable models can be confidently wrong when they should defer, cite, or say “I don’t know.”

OpenAI’s Sept. 5, 2025 analysis formalizes this view, tying persistent errors to three roots—epistemic uncertainty, model limitations, and computational intractability—while noting GPT‑5 still hallucinates and quantifying o1 at 16%, o3 at 33%, and o4‑mini at 48%; the paper urges revised benchmarks and explicit abstention targets for deployment [1].

The trade press amplified the finding: Computerworld framed hallucinations as “mathematically inevitable,” highlighted DeepSeek‑V3’s inconsistent counts across queries, reiterated o1 16%, o3 33%, o4‑mini 48% rates, and quoted analyst Neil Shah that models “lack the humility to acknowledge uncertainty,” thus warranting explicit confidence targets [2].

Measuring AI hallucinations: 16% to 79% across tests

The spread in reported error rates is not noise—it’s a signal that task design and evaluation protocols heavily influence observed reliability. Factoid QA, multi-hop reasoning, code synthesis, and retrieval-augmented tasks each stress different failure modes. In turn, context length, knowledge freshness, and the presence of distractors can push a model to guess when it should abstain.

Test-by-test results are even starker: on PersonQA, o3 hallucinated 33% and o4‑mini 41%, while on SimpleQA rates jumped to 51% for o3, 79% for o4‑mini, and 37.1% for GPT‑4.5, Forbes reported in May 2025 [4].

What explains such divergence? Task difficulty and ambiguity play a role, but so do incentives embedded in evaluation. If a benchmark does not reward abstention or calibrated uncertainty, models that “swing” on more questions can score higher despite worse real-world reliability. That paradox lets guessy systems look strong in leaderboards yet fail in production, where a single confident error can trigger downstream harm.

A practical takeaway for teams is to test across multiple datasets, include abstention options, and measure coverage-accuracy tradeoffs. An end-to-end view—testing with and without retrieval, with and without tool use, and with explicit “I don’t know” affordances—yields a far more realistic picture than a single headline score.

The math behind AI hallucinations and abstention incentives

If next-token prediction is the engine, the fuel is incentives. When training and evaluation prioritize being right “on average” and do not penalize confident wrong answers, models learn to bluff. This is the mathematical pathway to AI hallucinations: tokens that are statistically likely in the training distribution but factually false given the query and context.

LiveMint’s readout crystallizes the fix: align incentives to abstain. It summarizes OpenAI’s call to penalize confident errors, warns that current benchmarks reward guessing, and notes GPT‑4.5 and GPT‑5 reduced hallucinations but still benefit from reforms that give partial credit for uncertainty and abstention in enterprise settings [3].

Concretely, that means product metrics should move beyond accuracy to include calibration (how well probabilities match reality) and selective prediction (accuracy at a chosen coverage). Teams can set a reliability budget, e.g., “<1 major factual error per 100 accepted answers,” then tune thresholds to meet it—even if coverage drops. In regulated workflows, less coverage is a feature, not a bug, when it prevents high-cost failures.

Enterprise risks and mitigation strategies as AI hallucinations persist

For enterprises, the message is not to abandon generative AI, but to manage it like a probabilistic system with known failure modes. The sharp increase in errors on some QA tasks demonstrates why “always answer” defaults are risky in legal, medical, and financial contexts. Every deployment should define what the system must not do, where it may abstain, and how edits are audited.

Business Insider reports OpenAI’s safety team is testing “deliberative alignment” to curb deceptive behavior and hallucinations, citing a 16.5% deception rate for GPT‑5 in some tasks and urging industry-wide adoption of abstention incentives and calibration metrics to reduce systemic harm [5].

Mitigation now sits on three rails. First, calibrate: show confidence bands, surface citations, and require evidence checks before high-stakes actions. Second, constrain: use retrieval-augmented generation with authoritative sources, tool-use for verifiable operations, and strong guardrails that fail closed. Third, control: enforce human-in-the-loop review, maintain immutable logs, and run red-team drills against sensitive prompts, subtle ambiguity, and adversarial inputs.

Procurement and governance should evolve accordingly. RFPs need reliability scorecards, requiring vendors to disclose hallucination rates under abstention, calibration error metrics, and how thresholds shift coverage. Contracts should encode uptime-like SLOs for correctness and abstention behavior, plus escalation pathways when observed rates drift.

What changes next: benchmarks, calibration, and procurement checklists for AI hallucinations

Benchmarks must reward abstention. That means scoring schemes that allocate partial credit to “I don’t know” when the alternative would be a confident, false claim. It also means reporting selective-accuracy curves, not just single-point accuracy, so buyers can see how performance changes as confidence thresholds move.

Calibration should be measured routinely. Expected Calibration Error (ECE), Brier scores, and reliability diagrams are standard in prediction systems and should become standard for generative AI. Models with similar accuracy but lower miscalibration create fewer surprise failures and are easier to govern because their probabilities mean something operationally.

Enterprises should require five artifacts before go-live: a coverage-accuracy curve with abstention enabled; a calibration report; a red-team dossier targeting domain risks; a retrieval provenance plan (citations, freshness, and link rot protections); and a fail-closed design where low-confidence answers route to humans. This makes it possible to choose a confidence threshold that meets a target error budget, rather than hoping prompts alone will fix AI hallucinations.

How to interpret the 16%–79% range without panic

The wide range does not mean every answer is unreliable; it means reliability is contingent on task framing, dataset composition, and the incentives the model perceives. A system that confidently writes code may be conservative on clinical guidance if it has learned that abstention is rewarded and that unsupported claims trigger penalties.

Leaders should calibrate expectations: even small percentages can be unacceptable in high-stakes workflows, while larger rates can be tolerable in brainstorming. The goal is not perfection but predictability—know when the model will answer, when it will abstain, and how it signals uncertainty. When you control that behavior, AI hallucinations become quantifiable, containable risks rather than existential blockers.

The industry’s pivot is clear. Reliability will come less from squeezing another floating-point improvement out of a base model and more from aligning incentives, rewarding abstention, and exposing calibrated uncertainty to users. With that shift, “don’t know” becomes a feature that enables trust, not a flaw to be engineered away.

Sources:

[1] OpenAI – Why language models hallucinate: https://openai.com/index/why-language-models-hallucinate

[2] Computerworld – OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws: www.computerworld.com/article/4059383/openai-admits-ai-hallucinations-are-mathematically-inevitable-not-just-engineering-flaws.html” target=”_blank” rel=”nofollow noopener noreferrer”>https://www.computerworld.com/article/4059383/openai-admits-ai-hallucinations-are-mathematically-inevitable-not-just-engineering-flaws.html [3] LiveMint – OpenAI admits GPT‑5 hallucinates: ‘Even advanced AI models can produce confidently wrong answers’ — Here’s why: https://www.livemint.com/technology/tech-news/openai-admits-gpt-5-hallucinates-even-advanced-ai-models-can-produce-confidently-wrong-answers-heres-why-11757305341278.html

[4] Forbes – Why AI hallucinations are worse than ever: www.forbes.com/sites/conormurray/2025/05/06/why-ai-hallucinations-are-worse-than-ever/” target=”_blank” rel=”nofollow noopener noreferrer”>https://www.forbes.com/sites/conormurray/2025/05/06/why-ai-hallucinations-are-worse-than-ever/ [5] Business Insider – OpenAI says its AI models are schemers that could cause ‘serious harm’ in the future. Here’s its solution.: www.businessinsider.com/openai-chatgpt-scheming-harm-solution-2025-9″ target=”_blank” rel=”nofollow noopener noreferrer”>https://www.businessinsider.com/openai-chatgpt-scheming-harm-solution-2025-9

Image generated by DALL-E 3