OpenAI Faces Backlash Over Vibe-Graphing Technology

During its major GPT-5 livestream on Thursday, OpenAI presented several charts that appeared to showcase the model’s impressive capabilities—but a closer look reveals some inconsistencies in the data visualization.
Ironically, in one chart illustrating GPT-5’s performance in "deception evals across models," the scale seems misaligned. For instance, the onstage graphic claimed that GPT-5 with "thinking" achieved a 50.0 percent deception rate. Yet, OpenAI's smaller o3 model—with a reported 47.4 percent—appeared as a larger bar. The data published in OpenAI’s official GPT-5 blog post, however, shows GPT-5’s actual deception rate at 16.5 percent.
This visual mismatch meant that onstage, GPT-5’s bar was shown as taller than o3’s, even though its reported score was lower. Elsewhere in the same chart, o3 and GPT-4o received different scores but were represented by bars of identical size. The error was significant enough for CEO Sam Altman to publicly acknowledge it, labeling the incident a "mega chart screwup," while clarifying that the blog contained the correct version.
An OpenAI marketing staff member also apologized, stating, "We fixed the chart in the blog, everyone—sorry for the unintentional chart crime."
On Friday, responding to a Reddit user’s query about the graphs, Altman explained that “the numbers here were accurate, but we messed up the bar charts during the livestream. On another slide, we got the numbers wrong.” He confirmed that the blog post and system card data were accurate, adding that “team members were working late and exhausted, which led to human error. It all comes together last-minute for a livestream.”
Still, the errors presented an unfortunate look for the company on such an important launch day—especially as it promotes the new model's “significant advances in reducing hallucinations.”
Updated August 8th: Added Altman's Reddit comment.
Related article
Sequoia-backed Empirik raises $21M to forecast outages before they strike
Before joining Sequoia Capital in 2020 as chief digital and information officer, Avon Puri spent over a decade running infrastructure at Rubrik and VMware. Three years ago, as large language models began showing their true potential, Puri, alongside
OpenAI launches GPT-5.5 Instant as ChatGPT’s new default model
OpenAI launched GPT-5.5 Instant on Monday, replacing GPT-5.3 Instant as the default ChatGPT model. The company states that this update minimizes hallucinations in high-stakes fields like law, medicine, and finance, while preserving the fast response
Nvidia to Derive 25% of Next Year’s Revenue from Funded Labs
Nvidia has deployed nearly US$50 billion into AI labs purchasing its chips, securing commitments for over $500 billion in total.On August 26, CFO Colette Kress informed analysts that these labs, backed by Nvidia’s balance sheet, will account for roug
Related Special Topic Recommendations
Comments (1)
0/500

During its major GPT-5 livestream on Thursday, OpenAI presented several charts that appeared to showcase the model’s impressive capabilities—but a closer look reveals some inconsistencies in the data visualization.
Ironically, in one chart illustrating GPT-5’s performance in "deception evals across models," the scale seems misaligned. For instance, the onstage graphic claimed that GPT-5 with "thinking" achieved a 50.0 percent deception rate. Yet, OpenAI's smaller o3 model—with a reported 47.4 percent—appeared as a larger bar. The data published in OpenAI’s official GPT-5 blog post, however, shows GPT-5’s actual deception rate at 16.5 percent.
This visual mismatch meant that onstage, GPT-5’s bar was shown as taller than o3’s, even though its reported score was lower. Elsewhere in the same chart, o3 and GPT-4o received different scores but were represented by bars of identical size. The error was significant enough for CEO Sam Altman to publicly acknowledge it, labeling the incident a "mega chart screwup," while clarifying that the blog contained the correct version.
An OpenAI marketing staff member also apologized, stating, "We fixed the chart in the blog, everyone—sorry for the unintentional chart crime."
On Friday, responding to a Reddit user’s query about the graphs, Altman explained that “the numbers here were accurate, but we messed up the bar charts during the livestream. On another slide, we got the numbers wrong.” He confirmed that the blog post and system card data were accurate, adding that “team members were working late and exhausted, which led to human error. It all comes together last-minute for a livestream.”
Still, the errors presented an unfortunate look for the company on such an important launch day—especially as it promotes the new model's “significant advances in reducing hallucinations.”
Updated August 8th: Added Altman's Reddit comment.
Sequoia-backed Empirik raises $21M to forecast outages before they strike
Before joining Sequoia Capital in 2020 as chief digital and information officer, Avon Puri spent over a decade running infrastructure at Rubrik and VMware. Three years ago, as large language models began showing their true potential, Puri, alongside
OpenAI launches GPT-5.5 Instant as ChatGPT’s new default model
OpenAI launched GPT-5.5 Instant on Monday, replacing GPT-5.3 Instant as the default ChatGPT model. The company states that this update minimizes hallucinations in high-stakes fields like law, medicine, and finance, while preserving the fast response
Nvidia to Derive 25% of Next Year’s Revenue from Funded Labs
Nvidia has deployed nearly US$50 billion into AI labs purchasing its chips, securing commitments for over $500 billion in total.On August 26, CFO Colette Kress informed analysts that these labs, backed by Nvidia’s balance sheet, will account for roug





Home






