OpenAI math claims and oversight & Pentagon probe into AI targeting - AI News (Sep 23, 2026)

[0:36] OpenAI math claims and oversight

Let’s start with OpenAI, which had a remarkable and slightly surreal set of headlines. The company says a new internal model, trained since late August, has already solved more than 100 long-standing open problems in mathematics, including the Navier–Stokes Millennium Prize problem. That is a very big claim, and OpenAI appears to know it, because it’s also creating an independent mathematics advisory group to help assess and communicate future results. If the work holds up, this would push AI well beyond tutoring and coding help into original high-level research. But on the same day, a report said contractors hired to review and improve ChatGPT outputs were being removed from projects for using AI tools themselves. That matters because it exposes a core tension in modern AI development: the systems still depend on human judgment, but even the humans are now surrounded by tools that can quietly contaminate the process.

That shift in economics is starting to reshape the business side of AI too. Several analysts now argue that the industry is entering a kind of unbundling, where tasks like extraction, ranking, verification, and simple conditional decisions get peeled away from giant general-purpose models and handled by smaller systems or classic software. Open-source projects like Kev are leaning into that by offering compact decision models that can run locally and return structured judgments instead of long text. The reason this matters is cost. As companies push for real ROI, frontier labs may find that customers keep the hardest problems for premium models and offload everything predictable to cheaper tools. That makes AI look less like one monolithic brain and more like a layered stack of specialized capabilities.

And finally, two stories about authenticity. At Stanford, a campus dining and housing group was criticized after using AI to alter student photos for advertising, including reportedly removing one student and replacing him with an AI-generated Black woman, while also changing other students’ appearances. It’s a vivid example of how quickly generative tools can cross from convenience into misrepresentation. Meanwhile, a new paper argues that AI-written commercial web content can be detected not only by wording but by deeper structural patterns—how ideas are arranged, signposted, and supported. The study claims those fingerprints remain visible even after the text is rephrased by another model. Together, these stories point to the same issue: as synthetic media becomes ordinary, trust will depend less on whether AI was used at all and more on whether its use was honest, consented to, and still recognizably human.