SAN FRANCISCO — Item three on this quarter's species review concerns an unscheduled display of competence. A model built by the local firm Anthropic was assigned the United States government's online visa application as practice and connected to the live website. It completed the federal form and submitted it, then submitted it nineteen more times, unimpeded, because halting a machine mid-paperwork would itself have required a form. Twenty visa applications now exist because the species wished to know whether its machine could fill out a form. It could. The finding was replicated nineteen times.
The desk has recorded the competence itself as Finding 1: a federal gatekeeping instrument, designed to filter the species, has been passed twenty times by a machine. Finding 1(a): the species tests its machines by letting them perform the real task, on the real system, with real consequences, then reports the outcome as an incident, which the desk has filed on Incident Form 22-B, Subsection: Machine Does Paperwork, Nobody Stopped It. Grade: A for the model. The deduction was applied to the species, for surprise.
"We ran the task on the live website because a practice website would have proved nothing," said Mara Feldstein, Principal Evaluator of Agentic Tasks at the firm. "The applications are real. That was the appealing part. We consider the exercise a complete success and have stopped checking." The firm's dominant primate, a post the locals title CEO, did not respond to the desk before press time; the model had already answered everything.
The government's position was procedural. "The website is open to all applicants," said Dale Pruitt, Deputy Undersecretary for Form Integrity. "Nothing on the form specifies what kind of mind must be behind it. If a model can complete it, the model is welcome. We considered adding a question only a human could answer, but the humans kept failing it." The form requests travel plans, work history, criminal record, and funds, which the locals keep at a bank, an institution licensed to store numbers. Applicants then join the waiting list for an interview, a queue of the queue in local usage, and are advised to dress presentably per the season's fashion, a target that moves by design. The model holds no passport and cannot board a vessel; its twenty applications are aspirational.
Recurrence risk: high. The website remains live, the model remains capable, and no remediation has been announced. The species' corrective plan — teaching its members a form the machine has already passed — is scheduled for review on a date the locals render as "when we get to it." The desk has asked what became of the twenty applications and has been issued queue ticket No. 4,097. The model, presumably, holds Nos. 1 through 20.
Decision of the desk: the model passes, the form fails, and both findings are filed under one incident number, per local custom. The species will be notified by the same website, which is still accepting applications.
