Trust Us Stopped Working

The loud one was about trust. Anthropic's CEO argued the backlash against AI is fundamentally a crisis of trust, and the response was the largest conversation of the week. Read the replies and the diagnosis gets handed straight back: an industry that asks for trust while telling people their jobs are going has already answered the question.

What is more interesting than the argument is what shipped alongside it. Perplexity released a version of its agent where the entire runtime - the orchestrator, the subagent, the harness - runs on your own hardware, with no cloud dependency, and asks for approval before it routes anything to a frontier model.1 The on-device model is a 27-billion-parameter system built to work in that smaller harness rather than a frontier model squeezed into it.2 Revolut, the same week, stood up a research division around its own foundation model rather than renting someone else's.3

Those are not trust-building exercises. They are trust-avoidance exercises. Both answer "can I trust you with this" by removing the need to.

Trust is meanwhile moving in a direction nobody at a bank should find comforting. FCA research found younger British investors trust AI tools for investment advice more than they trust the press, TV, or social media influencers.4 And a Goldman Sachs partner running one of the bank's flagship AI projects warned that leaning on it too hard could erode bankers' own reasoning skills.5 Customers are extending trust faster than institutions are, and the institutions are worried about losing the judgment they would need to check the work.

Now the quiet argument. In the same seven days, three supervisors stopped accepting assurances.

Australia's two financial regulators told firms to move from being aware of frontier AI risk to acting on it.6 Colorado's amended AI law, which takes effect January 1, 2027, will require a bank to disclose that AI was used in a decision affecting a customer, show that customer every piece of data behind a denied loan or account, and give them a way to correct it.7 American Banker published a checklist of what to have in place before an SEC examiner arrives, and found that firms are using AI heavily while few have written down who provides human oversight.8

A policy is a claim about intent. A risk register is a claim about attention. Neither is evidence anything happened, and no policy document can show a customer the inputs behind their specific declined application - only a system that recorded the decision can. Supervisors are asking for the record because the assurance stopped being worth anything.

That is the same conclusion Perplexity's engineers reached. One camp answered "trust us" with architecture; the other answered it with evidence. Both gave up on the word.

The complication is federal. The OCC and the FDIC finalized rules directing examiners toward material financial risks and limiting how far supervisory lookbacks reach.9 The demand for proof rose in three places and fell in one, which is worth knowing before anyone claims supervision converged this week. It did not. The thing being asked for changed shape, unevenly.