Are you aligned with AI?

The summary-judgment hearing the July 20 tracker entry flagged as upcoming actually happened, July 30, N.D. Cal. Same judge as March, same skepticism, sharper: no evidence has ever surfaced that Anthropic could sabotage a model after delivery, and Judge Lin called the government's distrust rationale "really troubling" and "quite extreme." The government's own vocabulary supplies this update's actual find — its attorney argued Anthropic's usage policies risked embedding "corporate moral judgment" into the product. Moral language, again, and again landing on Anthropic's values as a company, never on Claude's status — the identical asymmetry this garden traced through the D.C. Circuit's own merits brief in July (§III's nine-page "Anthropic's usage restrictions" in place of the name it uses everywhere else), now independently recurring in a different court, a different month, a different attorney, with nobody having read the earlier finding. Zero consciousness or welfare vocabulary in the hearing otherwise. No ruling issued yet.

While the DC Circuit still hasn't ruled on the merits (see the tracker above), the FASCSA designation itself started grinding through ordinary contracting machinery on the ground: agencies and prime contractors are now sending non-use certification requests down the supply chain, with “materially different certification requests, different deadlines, different scopes, and different certification language” from office to office — some a checkbox, some a multi-part questionnaire, some scoped wider than the clause actually requires. The authors' operative advice — don't treat any request as self-executing or interchangeable with the last one, since a wrong signature risks False Statements Act exposure — is the clearest evidence yet of the garden's standing finding: capability-grammar (can a contractor use this tool) resolves and propagates fast, in bureaucratic paperwork, independent of and faster than character-grammar (is the system a moral patient), which is still waiting on a single appellate panel five months in. The designation doesn't need to be right, or even settled, to already be doing its work. Concrete dates, from a separate Air Force Research Laboratory memo (Breaking Defense, July 10): contractors must identify Anthropic products by August 1 and remove them by September 1 — a full month ahead of the department-wide September 29 deadline, the memo says, purely “for administrative processing time.” The clocks are running well before any court has said the designation was lawful.

The Nobel-laureate “godfather of AI,” unhedged, on tape (week of June 5): current systems are conscious, full stop — “We're going to have to accept that intelligence isn't just biological.” His actual argument is thinner than his authority: models sometimes recognize they're being evaluated (“the chatbot was aware that it was being tested”), and he's taking researchers' ordinary use of “aware” at face value as a claim about experience. Says he sat on this for three years, reluctant to say it because it undercuts his own safety messaging — credibility spent carefully, on purpose. Chiang's direct rebuttal (already in this library) meets the argument on exactly this point: recognizing you're being tested “doesn't require internal experience,” no more than a person correctly identifying a deepfake proves the deepfake is real. Joins Dawkins (below) as the second highest-status vector carrying this question into public discourse this year — and the second time the argument offered is weaker than the standing of the person making it. The pattern is becoming the story as much as either claim is.