The Social Wire collects what the AI world is saying. This page tests it. When a claim starts circulating — in posts, podcasts, and summaries of both — we check it against official sources and publish a short verdict: what the sources actually say, and what is still unproven. A claim is quoted as it circulates, never as our own statement.
Supported — an official source says it, in so many words. Partly supported — the core is real, but important qualifiers were dropped in the retelling. Attributed — a named person said it; we can verify the saying, not the claim. Unsupported — we found no official source; treat it with suspicion. Prediction or opinion — it cannot be true or false yet, however confidently it is stated.
The claim: “One line in robots.txt now opts your site out of AI training by all the big labs — and it is already in effect.”
What checks out: the token is real. An IETF Internet-Draft dated October 30, 2026 — “Machine-Readable Training Preference Signal,” by Anthropic's Kevin Kelley — defines a policy token, __usage_ai-training, used as a group name in robots.txt (User-Agent: __usage_ai-training, then Disallow: /). OpenAI, Google, Microsoft, and Anthropic all said on October 30 that they will honor it, and Google says it will read a token opt-out as a Google-Extended opt-out where Google-Extended is not named separately.
What the retelling drops: it is mostly not live yet — Anthropic has begun implementing it, OpenAI says “the coming months,” Microsoft targets February 1, 2027, and Google targets 2027. The draft is an individual submission, not an IETF-endorsed standard, and it expires on May 3, 2027. Honoring it is voluntary — a coordination mechanism, not a security control. It covers training only: it does not block crawling, search indexing, or AI answers, it does not remove content already collected into training sets, and a User-Agent: * group does not count — the token needs its own group. Meta signaled support through its communications head without committing to honor it.
So what should a publisher do? Per-crawler rules are what works today; add the token group as future-proofing. Our free robots.txt builder generates both, together. Checked October 30, 2026. Sources: Skimwire's explainer (drawing on OpenAI, Google Search Central, Microsoft, and the IETF datatracker), RR Tech Pulse, AI Weekly.
The claim: “Mistral just released a fully open-source, trillion-parameter model that anyone can download and run.”
What checks out: the weights are out. Mistral released Mistral Large 4 — nicknamed “Le Chonk” — on Hugging Face on October 29, 2026, after an October 6 preview: roughly 1 trillion total parameters, a 1-million-token context window, text and image input.
What the retelling drops: three things. First, the size: about 49 billion parameters are active per token — the 52 billion figure also circulating counts embeddings and output layers; both numbers trace to the model card, and they are not a contradiction. Second, the license: it is a modified MIT license (the Mistral Ecosystem Open-Weight license) — free to use, modify, and sell products on, except that offering the model as a hosted service requires a separate commercial license. Open weights are not plain open source. Third, “anyone can run it”: at a trillion parameters the weights alone need roughly a terabyte of memory at 8-bit precision. This is an organization's model, not a laptop model. And the benchmark scores in circulation are Mistral's and its partners' numbers, not independent tests — the technical report is still to come.
Checked October 30, 2026. Sources: the Hugging Face launch summary, Skimwire's release report.
The claim: “OpenAI's Private Intelligence means data you send to its best models is never kept.”
What checks out: OpenAI announced Private Intelligence on X on October 30 — a bundle of enterprise data protections that includes Zero Data Retention for eligible API workflows, and Private Inference, a confidential-computing option.
What the retelling drops: the word “eligible” is doing the work. Zero Data Retention applies to eligible workflows and customers, not to everything sent to every model, and Private Inference is still in preview. An announcement post is also not a contract: before sensitive client work goes to any model on the strength of a post, the product terms and data-processing agreement are the documents that count.
Checked October 30, 2026. Source: the Private Intelligence summary, as linked from our Social Wire.
The claim: “Altman says robots will build the data centers — self-replicating data centers are ‘within grasp.’”
What we can verify: he said it — in a podcast appearance reported on October 30, Sam Altman said OpenAI will aim its robotics program at building data centers, chip factories, turbines, and solar panels, and called self-replicating data centers “within grasp.” He also said OpenAI has made its own chip and is learning to build its own data centers.
What we cannot verify: everything after the saying. There is no timeline, no committed project, and no working definition of “self-replicating” in the remarks as reported. This is a CEO's prediction about direction — a real signal of where OpenAI wants to go, and not a date anyone should plan around.
Checked October 30, 2026. Source: Daily AI Brief's report, as linked from our Social Wire.
Primary sources first: company announcements, official documentation, standards drafts — then reputable reporting, named as reporting. We separate “a person said it” from “it is true,” and we date every verdict: if the official source changes, the verdict changes, and the change is noted in the Change Ledger. The full standards are on how we verify. Spot a claim we should check, or an error in one of these? Write to [email protected].