OpenAI Fixes ‘Unintentional Chart Crime’ After GPT-5 Livestream

While most people were glued to OpenAI’s newest AI model, GPT-5, during a demo Thursday, some couldn’t help notice something had gone awry in the background.

Several charts included in OpenAI’s GPT-5 livestream on Thursday had some clear mistakes.

One chart compared the models GPT-5 with thinking and OpenAI o3 on a metric called “coding deception.” The former model has a deception rate of 50%, but this bar was less than half the size of the bar for o3, which has a smaller deception rate of 47.4%.

Another chart comparing models on a different metric depicted 69.1% and 30.8% with the same size bar. A bar for 52.8% was also larger than both of these, though that’s smaller than 69.1%.

OpenAI CEO Sam Altman tweeted, “wow a mega chart screwup from us earlier.”

In a Reddit AMA on Friday, he said, “The numbers here were accurate but we screwed up the bar charts in the livestream overnight; on another slide we screwed up numbers.”

He added, “People were working late and were very tired, and human error got in the way. A lot comes together for a livestream in the last hours.”

A marketing employee for OpenAI tweeted that the issue had been fixed in the blog post of the announcement, apologizing for the “unintentional chart crime.”

Tech demos aren’t immune to the occasional flub.

When Microsoft unveiled Bing’s AI chatbot in 2023, it made several mistakes during its demo. When asked for pros and cons of a particular vacuum, it listed noisiness and a short cord as cons, but the vacuum is cordless. In an ad that same month, Google’s AI chatbot Bard, which has since been renamed Gemini, gave an incorrect answer to a question about the James Webb Space Telescope.

It’s unclear if AI was used to make the charts, and OpenAI didn’t immediately respond to a request for comment.