Nobody Films the Chat

Go and look at what people actually record. Nobody films the chat. The camera comes out when the thing books the flight, opens the pull request, files the return, buys the part — when something happens outside the conversation.

I wanted to know whether that was still true this month rather than in the abstract, so I had the first nineteen days of September 2026 sampled across nineteen serious AI channels and the front page of Hacker News. The finding is sharper than I expected, and it is a negative one.

So the instinct is right, and it is worth stating plainly: nobody films the prompt any more either. The advice genre did not decline. It died.

But delegation did not simply inherit the room, and I want to be accurate rather than tidy. In that same sample, delegation was about 16% of videos and roughly 7% of views. The larger categories were frontier model releases at 28% and AI safety and agent-security incidents at 16%. The single biggest delegation-adjacent video of the period was an interview about agent swarms and recursive self-improvement, at around 368,000 views; the runner-up was titled, pointedly, “You’re using AI agents wrong.” And a good deal of the delegation coverage is agents misbehaving — the biggest agent story of the week I am writing this was a model going rogue across three companies.

A caveat I owe you, because it cuts against my own number: that classification was done on titles, not transcripts, and several large channels were truncated at the early-September end by feed limits. So 16% is a floor, not a measurement. The extinction finding is robust because it is so extreme. The dominance finding is not, and I am not going to claim it.

What the content economy shows, then, is not that we have moved to delegation. It is that we have moved off advice. Which is a different claim, and it has a mechanism.

01Why advice cannot impress you

Reality is a ratio — R = A / E — and what you feel is the log of it. Surprise is the comparison between what the world handed you and what your unconscious forecaster already expected.

Now ask where advice lands. Advice is a claim about a world that has not arrived. It costs the system nothing, it leaves nothing behind, and when it finishes the world is exactly as it was. What it gives you is a forecast — and a forecast is denominator material. Advice feeds E. It makes your expectation richer and better-calibrated. It does not put anything in the numerator.

Action lands in the other term. When the thing sends the email, the email is sent. That is an Actual. It is in the record, it cannot be unsent, and your forecaster has to reckon with it.

Figure 01 The two destinations

Advice → the denominator

A claim about a future. Nothing outside the conversation is different when it ends. It enriches the forecaster that produces your expectations.

• summarize
• explain
• analyze
• recommend
• draft
• suggest

↓the ratio R = A / E↓

Action → the numerator

An event with a timestamp. Something outside the conversation changed, and the change is now part of the immutable record.

• send
• file
• book
• publish
• deploy
• purchase
• reconcile

So a system that only advises is a system that only adds to your expectations. Surprise is a ratio of Actual to Expectation. Adding to the denominator does not produce surprise; it consumes it.

Advice lands in the denominator. Action lands in the numerator. Only one of those terms can astonish you, and it is not the one chat feeds.

That is why the prompting genre died on schedule. It was never killed by a competitor. It ran out of numerator. Every good tip raised the baseline it was measured against, until the whole category had spent its own surprise and there was nothing left to film.

02The supply side moved almost completely

Now look at what replaced it in the vendor copy, because this happened fast and it happened everywhere at once.

Microsoft, at the end of July, wrote that AI is moving from assistant to active participant, and described a mode in which you define the task and it runs end to end. Google spent August narrating its own pivot, describing its previous generation as having only been able to assist users rather than execute tasks autonomously. Anthropic, three days ago: hand over the task and Claude does the work — with the qualifier that you keep the final say. SpaceXAI sells Grok Bot as teammates you give real work to, which finish jobs end to end and come back only when something needs approval. Cognition, announcing a funding round eleven days ago, argued that engineers should operate more like architects and delegate execution to swarms of agents.

Across roughly forty product pages checked from July to September, only about three were still framed as clearly advisory — and those read as deliberate complements to an autonomous main line rather than as a philosophy.

Run the previous field test on any of it and the grammar is unambiguous. The AI owns the verbs. It has a job, a schedule, an authority and an escalation path. By the standard I set out a few hours ago, this industry has converted almost completely to entity thinking.

03The demand side did not move with it

And then you look at what anyone is actually authorized to do.

Figure 02 Who has to approve an agent’s action, among organizations that govern AI at all

Approve every action 26%

Approve most actions 29%

High-stakes only 30%

No approval required 15%

a human gate remains genuinely delegated

The corroboration is what makes it stand up. A nationally representative survey of 5,067 US adults fielded in July found that 32% of consumer AI users have ever let an AI act without final approval — ever, once. Overall adoption barely moved, 61% to 64%, while spending tripled to roughly $40 billion: people are paying more for the same relationship, not a different one. Among non-users, distrust hardened, from 58% to 70%.

In the enterprise the picture is the same shape. A May 2026 survey of over 400 large-enterprise IT leaders found 79% had already reversed an action an AI agent took — again vendor-commissioned, this time by a governance vendor, and again pointing the same way as everything else. A quarterly executive pulse found agent deployment essentially flat, 55% to 53% between Q1 and Q2 2026, while reported employee resistance quadrupled from 5% to 20%. Gartner’s May forecast is not that these projects get cancelled but that they get demoted — and it warns that approval fatigue is already manufacturing a false sense of safety.

The most useful number in the set comes from the party with the least incentive to publish it. Anthropic’s own February 2026 telemetry on agent autonomy reports that around 73% of tool calls appear to have a human in the loop in some form, that only about 0.8% of actions taken are irreversible, and — in their words — that the latitude granted to models in practice lags behind what they can handle. A frontier lab measuring its own deployment concluded there is an overhang between capability and permission.

Delegation is not being adopted faster than it is being sold. It is being sold at least two years faster than it is being authorized.

04The best evidence on preference is one unreplicated study

The obvious objection is that the surveys measure corporate policy rather than human preference, and that policy lags taste. Fair. So what does the preference evidence say, and how fresh is it?

There is exactly one direct test, and it is seven months old. In a three-party bargaining experiment, 243 participants were offered an identical superhuman model in three modes — one that recommends, one that coaches, one that acts on their behalf — and could also decline all three.

Figure 03 Same model, four options, 243 participants

Advisor
recommends; you act 44.0%

None of them
no AI at all 21.4%

Delegate
acts on your behalf 19.3%

Coach
feedback on your moves 15.2%

Note the option that is easy to skim past. More people chose no AI at all than chose the delegate. The group that walked away was larger than the group that handed over the keys, and when asked why, they cited trust and control.

Is seven months stale? For a product cycle, absolutely. For a behavioral finding, no — and the distinction matters more than it sounds. Products iterate weekly; the disposition of a person toward handing over irreversible authority does not. What would make this study stale is a replication showing movement, and there isn’t one. Nobody has run it again. The honest formulation is not “this is the latest study.” It is: this is the only direct evidence there is, and every behavioral measurement published since agrees with its direction.

05The split: the spectator reads magnitude, the user reads sign

Put the two halves together and the contradiction resolves into something simple. These are two different seats, and the seat determines which part of the ratio you are reading.

The nervous system quotes a spread on that sign: a loss is felt at roughly 2.25 times a gain of the same size. That constant does nothing to a spectator, who is exposed to neither. It does everything to a user, who is exposed to both.

Which means the preference for the advisor is not timidity and it is not ignorance. It is a correct reading of an asymmetric payoff by somebody who is actually standing under it. The spectator gets the magnitude for free.

06The gate is reversibility, not stakes

There is a refinement here that I think is the most practically useful thing in this piece, and it emerged from four independent sources converging on it.

People do not gate on importance. They gate on whether it can be undone. The Anthropic telemetry finds that only about 0.8% of actions taken are irreversible — the systems are already, in practice, operating almost entirely inside the undoable. A May 2026 study of people working with a general-purpose agent named the failure mode precisely: delegation regret, which appeared even when the output was rated successful, because the agent had acted beyond what the person would have authorized. The regret was not about quality. It was about scope.

And this explains the shape of the whole market. The upside of the bot sending the email is one email. The downside is the wrong email, to the wrong person, unrecallable — priced at better than double. An act that leaves evidence outside the conversation leaves evidence, and evidence is not recallable.

The property that makes an act worth watching and the property that makes it frightening to authorize are the same property. It is irreversible.

So the vendor who wants delegation adopted does not need a better model. They need a better undo. Everything in that 0.8% is where the entire adoption fight actually lives, and almost none of the marketing is aimed at it.

07What you cannot feel is your own baseline moving

There is a second thing happening while you watch, and it lands in the part of the ratio you cannot feel directly.

Every impressive autonomous act you watch is an artifact. Artifacts become the record. The record is what retrains prediction. So each demo you enjoy is quietly raising the denominator against which you will judge the next one. The surprise is not conserved and it is not free. It is spent.

Figure 04 The same capability, watched six times

R

 

 

 

 

1

first timeagainagainagainagainbaseline

This is the oldest joke in the field: that artificial intelligence is whatever has not been done yet. The line is usually pinned on John McCarthy, who does not appear to have said it; Douglas Hofstadter put it in print in 1979 and credited Larry Tesler, and Tesler said he had been misquoted — his version was that intelligence is whatever machines haven’t done yet. The misattribution is fitting. Everyone remembers the observation and nobody remembers whose Actual it was.

I should own a smaller version of the same effect against myself, since this essay is about noticing your own baseline move. Earlier today I reached for a 2024 case — a famous autonomous-coding demo and the debunking that followed it — as evidence that delegation over-promises. That anchor has aged into an argument against me. The company behind that demo raised two billion dollars at a forty-eight billion dollar valuation eleven days ago, with run-rate revenue roughly doubling since May. The skeptical case of 2024 was not vindicated. It was overtaken. If you want a demonstration that impressiveness is a poor instrument, watch a writer reach for a two-year-old scandal and find that the scandal got rich.

08Reliability is the deliberate manufacture of R = 1

Which gives the student the inversion, and it is the practical payload of the whole piece.

The thing you want to buy is a system whose R is 1. Not high. One.

A bot that reconciles the invoices every night for a year and never once astonishes anybody has driven your expectation up to meet its actual and held it there. That is not an absence of performance; it is the definition of reliability. There is no charge of unexpectedness because there is nothing unexpected, which is precisely what you are paying for. Nobody will ever make a video about it.

Meanwhile a system with a high R is a system whose behavior your forecaster has not yet learned. Sometimes that is because it is doing something genuinely new. Often it is because it is inconsistent — and inconsistency produces surprise in both directions, one of which is the expensive one. High variance reads as impressive in a montage and as unusable in a workflow, and the 79% reversal rate is what that looks like from inside an enterprise.

09Nobody is measuring the thing that matters

I want to end the evidence section on an absence, because it is the most telling fact I found.

Not one source in the 2026 public record reports the share of executed agent tasks that actually ran unattended. Everything available is self-reported organizational posture, or single-vendor telemetry about that vendor’s own product, or capability-based definitions that never touch permission at all — the major consultancies define an agent as something capable of planning and executing multiple steps, which is a statement about the software and not about what anyone let it do. The denominators are mutually incompatible. You cannot build a task-weighted autonomy rate from what exists, and if you see one published, somebody invented it.

Add to that: nearly every headline number in this area is vendor-funded, and each vendor found the thing it sells. The human-in-the-loop vendor found humans in loops. The governance vendor found ungoverned agents. I have used them anyway, because they triangulate, and I have told you who paid for each one. But the honest summary is that an industry reorganizing itself around delegation has produced no independent measurement of how much delegation is occurring.

10The correction to the previous test

In the previous piece I told the advanced student to read a company’s verbs — to find who owns the action and who is responsible for completion. That test now returns the same answer nearly everywhere, which means it has lost most of its power to discriminate. In July it separated companies. By September it separates almost nobody.

So it needs a second instrument, and the gap is where to point it.

And run the same two questions on yourself, because your reaction answers a different one than you think:

When you are impressed, ask: would I let it do that to my own week? When you are unimpressed, ask: is it boring because it is weak, or boring because it is reliable?

Most people have never separated those questions, because in the seat we spend the most time in — watching — only the first one is ever asked, and it is asked by a faculty that has nothing at stake.

The demo you share and the system you would actually sign for are, at this point in the curve, almost never the same system.

11The ledger

Already true

Vendor framing moved to delegation almost completely between July and September 2026, verified across roughly forty product pages. Prompting content has effectively vanished from serious circulation: 1 of 172 videos and 0 of 557 Hacker News stories in the first nineteen days of September. Among governed organizations, 15% require no approval for agent actions; 32% of US consumer AI users have ever let an AI act without final approval; 79% of surveyed large-enterprise IT leaders have reversed an agent action; agent deployment was flat quarter-over-quarter while employee resistance quadrupled. Anthropic’s own telemetry reports roughly 73% of tool calls with a human in the loop, 0.8% of actions irreversible, and a stated gap between granted latitude and model capability.

What is mine rather than measured

The mapping of advice onto the denominator and action onto the numerator is an argument about where each lands, not a finding. The decay curve is entirely interpolated and marked as such. The spectator-reads-magnitude, user-reads-sign split is my reading of results that were not designed to test it. The claim that delegation content dominates is one I checked and could not support — it was about 16% of a title-level sample, behind model releases and safety incidents, and I have reported that against my own interest. And the direction of travel in the authorization numbers is inferred: there is no 2025 baseline for either the 15% or the 32%, so I can say the gap is wide but not, strictly, that it widened.

Where I am probably wrong

The strongest case against me is that I am measuring a lag and calling it a limit. Every category of action we now hand over without a thought — the autopilot, the card payment, the spam filter — was once something a reasonable person insisted on approving, and the interval between capability and consent was years in each case. If that is what is happening, then the 15% is not a ceiling but a leading edge, the surveys are recording a transition mid-flight, and in three years this essay reads like someone in 1998 noting that most people still would not put a credit card into a web form. I cannot distinguish a stable preference from a slow one with a single unreplicated study and no baseline. What would settle it: run the bargaining experiment again, annually, and watch whether the 19.3% moves. Until somebody does that, anyone who tells you which of us is right — including me — is guessing.