On May 31, 2023, OpenAI published research on a way to reduce mistakes in AI reasoning. The work compared process supervision with outcome supervision when training models to solve math problems.
At the time, GPT-4 had just launched, and ChatGPT was already known for answers that sounded sure but were sometimes false. These false answers became known as AI hallucinations. They can include fake facts, people, events, sources, or legal cases.
OpenAI’s early tests found that process supervision worked better than judging only the final answer. The method gives feedback on each step of a model’s reasoning, instead of checking only whether the final result is right. OpenAI also released its training data to support more research.
The problem became very real in 2023 when lawyer Steven Schwartz used ChatGPT while working on a case against Avianca. The chatbot gave him fake legal cases, and a U.S. judge later sanctioned the lawyers involved. The case showed why fluent AI text still needs human checks.
Why AI can still sound sure when it is wrong
In September 2025, OpenAI said standard training and tests can reward a model for guessing instead of admitting that it does not know. A model may score better by guessing, even when a careful answer would say “I’m not sure.”
OpenAI said GPT-5 had fewer hallucinations than earlier models, especially when it used deeper reasoning. But the company also said hallucinations remained a basic problem for large language models.
This matters for crypto users because AI tools are often used to check token data, smart contracts, legal rules, market news, and wallet activity. A wrong answer can lead to a bad trade or a wrong claim in a report. Users still need to check important facts against trusted sources.
The same lesson applies to developers. Better model scores do not always mean a model will know when a claim is false. A system that can say “I don’t know” can be safer than one that gives a confident but wrong answer.
Also Read: Microsoft Warns of New USB Malware Aimed at Crypto Investors
Newer models show progress, but the problem is not gone
OpenAI released GPT-5.6 in July 2026, with Sol, Terra, and Luna models. Its safety tests included checks for factual errors and user-reported hallucinations. OpenAI said GPT-5.6 Sol made slightly fewer factual errors than GPT-5.5 and was less likely to repeat errors that users had already flagged.
The bigger step came in September 2026 with GPT-6 Astra. OpenAI said Astra made substantially fewer factual errors than GPT-5.6 Sol in tests built from past user-reported mistakes. The company also said Astra was much less likely to repeat those reported hallucinations. OpenAI warns that these tests were designed to find weak spots, so their results do not show the normal error rate for all ChatGPT use.
Enjoyed this? Bookmark DeFi Planet, explore related topics, and follow us on Twitter, LinkedIn, Facebook, Instagram, Threads, and CoinMarketCap Community for seamless access to high-quality industry insights.
Take control of your crypto portfolio with DEFI PLANET PRO, DeFi Planet’s suite of analytics tools.
