It does not sleep, it does not tire It holds ten thousand hands at once And every hand believes the wire Runs only to its own response
The week opened on a machine that never runs out, and on the one experimental intervention that neutralised it, which was a rate limit. Slow the thing to human pace and its advantage over trained debaters disappeared. The advantage was throughput.
The obvious follow-on question is one the laboratory never had to answer and the world already has. What is throughput, once it is cheap? Because in the study the researchers could impose the rate limit whenever they liked. Outside the study, nobody is imposing one, and the persuader that could only ever hold a single conversation at a human speed has been quietly replaced by one that holds as many as there are people to hold, and remembers each of them.
That is a different category of thing altogether, and it is already in your phone.
The debater becomes a workforce
The cleanest place to watch throughput stop being a laboratory curiosity and start being infrastructure is a political campaign, because campaigns have always been throughput-limited in the most human way imaginable. You have a finite number of volunteers, they can each knock a finite number of doors or hold a finite number of conversations before they need to eat and sleep, and the entire ground game is an exercise in rationing scarce persuasive labour across a map.
Campaign vendors are now training bots to sound like the candidate and hold personalised text conversations with voters at a volume no volunteer army could reach. One vendor said it had worked with ten campaigns in 2025 and was aiming past a hundred in 2026, roughly half of that already reached by July. Another reported millions of texts and tens of thousands of conversations. North Dakota and California require the bot to disclose itself in the first message. New Jersey was weighing something similar.
The experimental persuader, the one that beat the elite debaters by never running out, has left the paid research platform and taken a job. The rate limit that neutralised it in the lab is exactly the thing the campaign is buying its way around. The whole product is the absence of the rate limit.
The scale numbers are the vendors' own, offered to impress future clients, and the reporting does not establish that a single one of those conversations changed a single vote. What it establishes is direction and incentive, which is the useful part. Somebody has decided that tireless, personalised, candidate-voiced conversation at scale is worth paying for, and they were not wrong to notice that it is now cheap.
Capacity is not reach
There is a lazy version of this argument that treats production and influence as the same quantity, and they are different things.
In June, OpenAI reported banning two clusters of accounts it linked to apparent state-origin influence operations, generating material about data-centre costs, tariffs, user-data claims, and US AI policy. The detail worth keeping is the one the provider volunteered: it found no evidence of meaningful breakout beyond the operation's own activity. The machine produced plenty. Almost nobody saw it.
That is throughput without reach, and it is the necessary correction to the contest image. A voice that never runs out has solved exactly one problem, which is the cost of producing the voice. It has not solved distribution, it has not solved attention, and it has not solved the basic difficulty of getting a stranger to care. Cheap production lets you test a thousand narratives. It does not deliver you an audience for any of them.
So the campaign bot and the banned influence cluster sit at two ends of the same fact. Both are throughput made cheap. One of them was wired directly into an authenticated channel, the voter's own phone number, obtained through a campaign that had a right to be there, and the other was shouting into a feed that routed around it. What separated them was whether the abundant voice had been plugged into a channel that reaches a person.
Which puts the governance object somewhere other than the machine's productivity. We do love a good productivity panic, and this is a different animal. The object is the wiring between abundant production and the channels that carry it to someone, and that wiring is owned, and it is the subject of The Positions It Will Argue.
The exchange that listens while it talks
There is a second property of the campaign bot that the door-knocking volunteer never had, and it is easy to miss under the sheer scale.
The volunteer persuades and forgets. They have a conversation on a doorstep, they might scribble a note on a clipboard, and then they walk to the next house carrying almost nothing forward. The persuasion and the intelligence-gathering were the same event, but the gathering mostly evaporated, because a human canvasser cannot retain and collate ten thousand doorstep conversations into a queryable picture of a precinct's private hesitations.
The bot does not forget. The same exchange that makes the case also records what moved the person and what did not, which objection they raised, which reassurance landed, and it does this across every conversation at once, so that the argument the campaign makes tomorrow is shaped by what ten thousand people quietly admitted today. Throughput runs as a sensor in the same wire as the advocacy, and the person on the other end experiences one warm personal exchange while contributing to a dataset they cannot see.
This is where the reception surface starts to matter as more than a phrase. The machine can plainly produce the argument, endlessly, and better than the volunteer would have. The voter's question is whether they can tell that the warm and personal message is one of ten thousand, sent by labour capacity rather than by a neighbour, gathering as it goes, and whether the disclosure that some states now mandate in the first message survives past the first message into the part of the conversation where the actual work is done.
What abundance actually changed
The shape generalises well past the campaign, which is only the early adopter here rather than the special case.
Every domain that was throughput-limited by the cost of persuasive human attention is about to stop being. Customer retention, fundraising, public consultation, recruitment, the slow grind of getting people to change a habit or renew a subscription or show up to a thing. All of it ran, forever, on the scarcity of people willing and able to have the conversation. Remove that scarcity and you get a different activity altogether, one where the conversation can be offered to everyone, always, at a pace and volume calibrated per person, tireless in a way that no retention team or fundraising office could ever be.
The optimist's version is real and belongs here: some of those conversations are ones people actually wanted and could never get, the patient explanation, the second chance at a form, the advocate who does not sigh at the fourth question. That gets its full day in Who Gets an Advocate.
The point here is narrower, and it is about the machinery underneath both the good and the bad version. The rate limit was the result. It was the whole result. And having discovered that the machine's edge was its refusal to run out, we have built an economy that is busily removing every remaining reason it ever might.
Whether that flood is water or sewage is the question in The Weight of the Reply, and it turns entirely on what all that volume is carrying. The machine solved the cost of speaking. It did not, on the way past, solve the problem of whether anything it says is true.
The debater who never sleeps is not a better debater. It is a factory, and the thing coming off the line is your afternoon, personalised, tireless, listening while it speaks, priced to reach everyone because reaching everyone finally costs almost nothing.
Companions
- The campaign channel: NPR via Georgia Public Broadcasting on candidate-voiced campaign texting.
- Production without reach: OpenAI on PRC-linked influence operations.
- The reveal this builds on: The Side That Never Runs Out.
- Capability arriving faster than its governance: The Door at the Desk.
- In the wings: "The Waldo Moment," where the visible persona carries whatever message it is handed and the operator behind it is replaceable.
These notes come out of Sociable Systems, a practice that reads AI-shaped documents the way a hostile reviewer will, before a lender or a court finds the gap. The argument has an operational form: the Interim Protocol sets out four rules for AI use in environmental and social deliverables, covering disclosure at touch-point grain, evidence custody, the phrases no automated screening may settle, and a hostile read before anything ships. Free, and written to be cited or retired once institutional guidance arrives.