In a shift from what it formerly held, OpenAI has acknowledged that hallucinations—instances where artificial intelligence generates plausible but false information—are not only common but mathematically unavoidable, regardless of improvements in training data or engineering techniques.
The admission came in a landmark study published on September 4 by OpenAI researchers Adam Tauman Kalai, Edwin Zhang, and Ofir Nachum, alongside Georgia Tech professor Santosh S. Vempala. The paper, republished by Computer World, establishes a mathematical framework showing why large language models (LLMs) will always produce some level of falsehood, even when trained on perfect data.
“Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty,” the authors wrote.
Register for Tekedia Mini-MBA edition 19 (Feb 9 – May 2, 2026): big discounts for early bird.
Tekedia AI in Business Masterclass opens registrations.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Register for Tekedia AI Lab: From Technical Design to Deployment (next edition begins Jan 24 2026).
The finding carries weight given OpenAI’s role in igniting the global AI boom with ChatGPT, which has since been integrated into enterprises, schools, and governments. It also marks a departure from earlier industry claims that hallucinations could be engineered away with better training, fine-tuning, or retrieval-augmented methods.
When Better Models Still Fail
The study demonstrated that hallucinations result from statistical properties inherent to model training. The researchers derived mathematical lower bounds proving that generative models will always retain an irreducible error rate.
Even state-of-the-art systems stumbled on seemingly trivial tests. When asked, “How many Ds are in DEEPSEEK?” DeepSeek-V3, Meta AI’s models, and Anthropic’s Claude 3.7 Sonnet all returned incorrect counts ranging from two to seven. OpenAI confirmed its own systems were no exception.
“ChatGPT also hallucinates,” the paper admitted. “GPT-5 has significantly fewer hallucinations, especially when reasoning, but they still occur. Hallucinations remain a fundamental challenge for all large language models.”
Ironically, OpenAI’s most advanced reasoning models hallucinated more than simpler ones. Its o1 model fabricated details in 16 percent of tests, while newer o3 and o4-mini models produced fabricated results 33 percent and 48 percent of the time, respectively.
“Unlike human intelligence, it lacks the humility to acknowledge uncertainty,” said Neil Shah, VP at Counterpoint Technologies. “When unsure, it doesn’t defer to deeper research or human oversight; instead, it often presents estimates as facts.”
How Rivals Have Positioned the Problem
The study directly challenges the narrative advanced by OpenAI’s competitors. Anthropic, the maker of Claude, has often marketed its constitutional AI framework as a way to reduce hallucinations, emphasizing “alignment” and “trustworthiness.” Google’s DeepMind similarly claimed that retrieval-augmented generation (RAG) could drastically cut hallucinations by grounding answers in external databases. Meta, too, has argued that scaling model size and refining evaluation would push the problem closer to elimination.
But the OpenAI study points in the opposite direction, noting that hallucinations are not byproducts of immature engineering, but consequences of deep mathematical laws. By showing that even rival flagships like Claude and DeepSeek-V3 produced wildly incorrect answers to simple factual questions, OpenAI positioned its research as not just a self-diagnosis but a critique of the broader industry’s optimism.
Flawed Benchmarks, Flawed Incentives
The study also exposed how industry evaluation methods worsen the problem. Current benchmarks such as GPQA and MMLU-Pro penalize models for responding “I don’t know,” effectively rewarding confident but wrong answers.
“We argue that language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty,” the researchers said.
Analysts say this dynamic is already harming real-world deployments. “Clients increasingly struggle with model quality challenges in production, especially in regulated sectors like finance and healthcare,” noted Forrester’s Charlie Dai.
Permanent Challenge, New Strategies
Experts believe the inevitability of hallucinations calls for a governance shift. “This means stronger human-in-the-loop processes, domain-specific guardrails, and continuous monitoring,” Dai said, adding that existing risk frameworks “underweight epistemic uncertainty.”
Shah drew a comparison to automotive safety regulations. “Just as car components are graded under ASIL standards, AI models should be dynamically graded nationally and internationally based on reliability and risk profile.”
Recommendations include calibrated confidence targets, real-time trust indices for evaluating AI output, and updated benchmarks driven by regulatory pressure and enterprise demand.
A Reality Check for Enterprises
The OpenAI findings echo earlier warnings from academia. A Harvard Kennedy School study found that downstream oversight often fails to catch subtle AI-generated falsehoods due to constraints of cost, scale, and context.
The OpenAI team concluded that the path forward requires industry-wide reform in how systems are tested and trusted, while acknowledging that hallucinations will never fully disappear.
For enterprises, the message is that hallucinations are not an engineering flaw to be patched out, but a mathematical certainty requiring new governance, oversight, and adaptation strategies.
By admitting this publicly, OpenAI not only sets itself apart from rivals who continue to promise engineering solutions but also reframes the debate over AI reliability. Thus, the question is no longer when hallucinations will disappear, but how businesses, regulators, and users adapt to their permanence.