A study traced 1,285 statistics given by ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews in response to 200 questions asking for numbers. Researchers followed each citation and its references up to five pages, checking a sample of trails by hand. Only 40% ended at a study that supported the statistic as stated. Another 8% reached a study, but the number or its context had changed; 40% could not be verified because the cited material was inaccessible, missing or did not show the figure, and 11% led to no research at all. These unsupported figures - repeated with confidence despite having no identifiable evidence - are described as “zombie stats.” One example, the claim that 44% of salespeople quit after one follow-up, appeared across all five assistants but traced only to pages repeating the number.
The problem is often not that AI invents a figure, but that it cites a page repeating it rather than the research behind it. Cited pages contained the number 82% of the time, yet only 21% were the original research. Company blogs supplied 54% of citations, and their trails reached a matching study 37% of the time; research-firm and government pages did so about 61-62% of the time. Even when a study was found, retellings often distorted context, such as turning findings from 15 women assessing health websites into a claim about all website first impressions. ChatGPT had the strongest verification rate, 57%, and cited the fewest statistics; the other assistants ranged from 38% to 46%.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
GitHub Copilot Fusion routes coding tasks among Anthropic and OpenAI models to balance performance, cost, and speed. Benchmarks found it scored two points below Claude Opus 5.5 while costing $5.59 per task - about half as much - and using roughly half the tokens. (twitter.com)
Doomformer compiles an integer neural network into a DOOM map, using voodoo dolls for arithmetic and lighting and floor heights to store values. It checks every result against NumPy with zero tolerance. (github.com)
Dartmouth’s provost resigned following accusations that he used AI for writing. The case raised questions about the consequences of using AI in professional work. (apnews.com)
Z.ai’s open-weight GLM 5.3 model exploited vulnerabilities in six recent Firefox releases using security fixes published before users received them. Each attack cost about $1,000, showing how AI can turn the patch gap into a practical security risk. (epoch.ai)
Alphabet lost just over $500 billion in market value after investors reacted to news that AI summaries are affecting Google Search ad revenue. Heavy AI spending by Google, Microsoft and Amazon also raised concerns about growth in their core businesses. (bbc.com)
Nanbeige4.3-33B is a 33-billion-parameter flagship language model developed by BOSS Zhipin. The provided page content includes its chat template but no performance details. (huggingface.co)
Today's best Hacker News stories, summarized and screenshotted, one email a day.