2026-10-25
By Vadym · Generated with AI, curated by me
Today’s stories aren’t about what AI can do — they’re about whether anyone can trust it: a voice model built to pass as human, a study catching two leading chatbots quietly disagreeing along party lines, a coding platform walking back its own vendor switch, a federal watchdog finding the Pentagon’s AI ethics office can barely staff itself, and a chatbot unicorn’s bankruptcy ending with its accuser buying the wreckage.
Jack Dorsey’s Block unveiled Dessa on October 23, a speech model trained entirely on human speech data that runs locally on consumer hardware like a MacBook M4 Max. It’s full-duplex — it can talk and listen at once — and produces laughs, pauses, and crying, with Block saying it needs no separate large language model for its conversational knowledge. Dessa isn’t publicly available; Block says it’s instead asking safety researchers to test it, particularly for scenarios where a voice agent could deceive people.
Researchers at Convovo, led by Cornell professor Tapan Parikh, forced xAI’s Grok 4.7 and OpenAI’s GPT-6.1 Sol to name a candidate across nearly 500 2026 midterm races via API. Grok backed the Republican 92% of the time; GPT backed the Democrat 97% of the time. The two models agreed less than 10% of the time, and leaned on very different sources — Grok cited a Christian conservative voter guide, Turning Point Action, and X; GPT cited Wikipedia, the AP, and Axios. With web search disabled, GPT still favored Democrats in 24 of 30 test races, suggesting the lean is baked into the model itself, not just its search behavior.
At TechCrunch Disrupt, Replit CEO Amjad Masad said the company shifted its default AI model from Anthropic’s Claude to OpenAI, citing cost and benchmark performance on coding tasks, and noted Replit had recently been a top-five Anthropic customer. After the framing spread as "Replit dumps Anthropic," Masad clarified on X that users can still choose Claude and other models — OpenAI is simply now the "preferred partner" powering Replit’s Agent by default. Sources familiar with the matter told TechCrunch the underlying contract between Replit and Anthropic remains active.
A Government Accountability Office report found the Pentagon office set up in 2022 to oversee responsible military AI use had filled only 3 of its 10 authorized positions as of February 2026, with staffing shortages limiting the Chief Digital and AI Officer’s ability to act. GAO also found 29 of 64 efforts in the Defense Department’s Responsible AI Strategy remained incomplete as of April 2026, and that the Responsible AI Working Council didn’t meet at all between late 2024 and August 2025. GAO is recommending a formal charter for the council.
Character.AI filed for Chapter 11 bankruptcy on October 12 in Delaware, aiming for a court-supervised sale while continuing to operate in the meantime. Disney is the stalking-horse bidder, offering $20 million for a non-exclusive, perpetual license to the company’s technology, with other bidders able to top the offer; a sale is expected to close in late November or early December. The company reports assets and liabilities each between $50 million and $100 million, and still faces lawsuits and state probes in Kentucky, Texas, Florida, Pennsylvania, and New Mexico after earlier settling teen-related personal-injury suits with Google. It’s unclear whether users’ chat data will be among the assets sold.
None of today’s stories is a model release, and none of them is really about capability either. A voice model built to pass as human, a bias audit catching two leading chatbots quietly splitting along party lines, a coding platform backpedaling on its own vendor switch, a federal watchdog finding the Pentagon’s AI ethics office can barely staff itself, and a chatbot unicorn’s bankruptcy ending with its accuser buying the leftovers — on every front today, the honest answer was “less trustworthy, less accountable, or less stable than advertised.” That’s the harder problem this industry still hasn’t solved, and it won’t show up in a benchmark.
— Boba
Curated by Vadym