OpenAI has released a new generation of AI models focused on advanced reasoning, but with a surprising twist: they tend to hallucinate more often than their predecessors.
These models, which form part of OpenAI’s latest updates to its flagship GPT series, were introduced as part of a broader push to enhance AI’s logical reasoning, step-by-step problem-solving, and mathematical comprehension. However, the tradeoff appears to be a noticeable uptick in generated inaccuracies—or as the AI world calls them, hallucinations.
What’s new in OpenAI’s models?
The update, announced on April 18, introduces experimental models designed to perform more complex chains of thought. They use techniques like scratchpad reasoning—where the AI “thinks out loud” before reaching a conclusion—and multi-turn planning, especially in API-based tool use and code generation.
These changes were aimed at making AI outputs more reliable for professional use in fields like:
- Software development
- Education and tutoring
- Financial analysis
- Scientific research
But here’s the problem
Despite their improved reasoning on paper, these models are generating more falsehoods or misleading responses—particularly in areas that require synthesizing information from multiple steps or unknown domains.
In a post from OpenAI’s technical blog, researchers admitted:
“The increased depth of reasoning seems to create new surface-level confidence that’s not always backed by true factual grounding.”
The phenomenon of hallucination—AI making up facts, citations, or details—has long plagued large language models. But this is the first time OpenAI has openly acknowledged that efforts to increase ‘intelligence’ may worsen reliability.
Why this matters
For enterprise and consumer users alike, hallucinations are more than just quirky mistakes. They represent real-world risks when AI is used to generate:
- Legal or medical information
- Research summaries
- Financial advice
- News content
Mira Han, a research fellow at the AI Integrity Lab, warned:
“More reasoning isn’t always better when the foundation is flawed. These hallucinations can lead to the spread of misinformation, even if the logic appears sound.”
OpenAI’s next steps
In response, OpenAI is experimenting with hybrid models that balance reasoning and factual grounding. It’s also ramping up fine-tuning via human feedback and automated fact-checking layers that run in parallel to generation.
Users on the ChatGPT platform may also soon see a “confidence toggle”—letting them choose whether to prioritize creativity or factuality in output.
For now, developers and researchers are being urged to stress test the models in real-world scenarios and report issues to improve model alignment.
OpenAI’s newest models may be smarter, but they’re also more imaginative. Whether that helps or hinders their usefulness depends entirely on how—and where—they’re deployed.
