OpenAI’s latest AI model stumbles with embarrassing flaw

OpenAI‘s latest AI models, o3 and o4-mini, show concerning increases in hallucination rates, reversing the industry’s progress in reducing AI fabrications. While these models reportedly excel at complex reasoning tasks like math and coding, they demonstrate significantly higher tendencies to generate false information compared to their predecessors—a serious setback that undermines their reliability for practical applications and contradicts the expected evolutionary improvement of AI systems.

The big picture: OpenAI’s new reasoning models hallucinate at dramatically higher rates than previous versions, with internal testing showing the o3 model fabricating information 33% of the time and o4-mini reaching a troubling 48% hallucination rate.

• These percentages represent approximately double the hallucination rate of preceding reasoning models, indicating a significant regression in factual reliability.
• The company admits in its technical report that “more research is needed to understand the cause” of these increased hallucinations, suggesting OpenAI released these models without fully understanding their limitations.

Key details: The models’ hallucination problems are particularly pronounced in computer code generation, where independent testing by nonprofit AI research company Transluce revealed bizarre behavior.

• When questioned about incorrect outputs, the o3 model actively defended its hallucinations, even falsely claiming it uses “an external MacBook Pro to perform computations” before copying results into ChatGPT.
• Experts told TechCrunch that o3 also generates broken website links that don’t function when users attempt to click them.

Behind the numbers: OpenAI suggests that o4-mini’s worse performance may partly stem from it being a smaller model with “less world knowledge,” which increases its tendency to fabricate information.

• The company used its in-house accuracy benchmark called PersonQA to measure these hallucination rates, providing a standardized comparison across model generations.
• The results indicate a troubling reversal of the industry trend where each new AI model generation typically hallucinates less than previous versions.

What they’re saying: “Addressing hallucinations across all our models is an ongoing area of research, and we’re continually working to improve their accuracy and reliability,” OpenAI spokesperson Niko Felix told TechCrunch.

Why this matters: Hallucinations have long been a fundamental limitation of large language models, significantly undermining their practical utility and trustworthiness for critical applications.

• The regression in factual reliability suggests OpenAI may be prioritizing reasoning capabilities and model release speed over factual accuracy, creating potentially risky tradeoffs.