Subtitle: Roslansky’s “doom loop,” and the seat outside the system that both ends walked away from
Preface
On August 31, Ryan Roslansky — the Microsoft executive who runs LinkedIn and oversees Office, Outlook, and Copilot, which is to say a man sitting near the center of the company’s AI push — published a short LinkedIn post about a document a colleague had sent him. It was, he wrote, clearly AI-generated. He read it, recognized what it was, and then did the thing the rest of this essay turns on: he handed it to his own AI to summarize, and got back nothing that added any thinking to the idea. “This is a doom loop.”
The reporting had its fun with the irony — the man whose products flood the workplace with generated text is annoyed at receiving generated text — and the irony is real. But read against the grain, the post is a cleaner natural experiment than it first looks. It is not really about whether AI writing is bad. It is about who was still standing outside the loop at the moment it closed. And the surprising answer, on the facts Roslansky himself supplies, is that someone was — and stepped in anyway.
The doom loop closes not because the witness is missing, but because both ends abdicated it.
1. What Roslansky actually described
Strip the post to its sequence, because the sequence is the whole argument. A document arrives. Roslansky reads it and correctly identifies it as machine-written. He then asks an AI to summarize the machine-written thing, and the summary returns no new idea — because there was no new idea in the original to return. He names the pattern a doom loop, warns that carelessly repeated it becomes an expensive way to avoid ever writing or reading anything, and closes with a prescription that is sound as far as it goes: use these tools to extend what a person actually brings — their own perspective and experience — rather than to stand in for it.
To be fair to him, none of that is naive, and this essay is not going to pretend it is. Roslansky diagnosed the failure in real time and prescribed roughly the right corrective. The point that follows is not that he missed something obvious. It is that the very clarity of his account exposes a mechanism he does not name.
2. The uncomfortable part: the judgment worked
The lazy reading of this episode — and the one most of the coverage reached for — is that machines cannot tell good content from bad, so slop accumulates until everything drowns in an ocean of sameness. That reading is available, and it is wrong, and this specimen is what refutes it.
Because the judgment did not fail. It succeeded, instantly, and for free. A human being standing outside the generator read one document and saw straight through it. Detection was not the hard part; detection was trivial. Whatever a doom loop is, on Roslansky’s own telling, it is not a case of the outside being unable to judge the inside. The outside judged the inside correctly on the first pass.
So the interesting question is not why couldn’t anyone tell. Someone could. The interesting question is why, having told, did he route the next step back through the machine anyway.
3. Two abdications
Follow both ends of the loop, because the loop has two, and each is a person.
At the sending end: a colleague who presumably had a point of view to contribute — a reason to be in the conversation at all — delegated the writing to a generator and forwarded the output as though it were their judgment. The outside seat, the human point of view Roslansky keeps asking for, was there to be occupied. It was handed to a model instead.
At the receiving end: Roslansky, who had already exercised judgment — he knew it was AI, that was his own eye working — then delegated the reading to a summarizer. Having stood outside the loop long enough to identify it, he stepped back in.
Two people. Each of them, at the decisive moment, was standing outside the system with a working faculty. Each of them handed that faculty to the system. The loop did not close because there was no one outside it. It closed because the people who were outside it declined to stay. The outside is not a location the tools deleted. It is a posture the humans vacated — and a posture, unlike a location, cannot be defended by anyone but the person who is supposed to occupy it. The witness seat was not destroyed. It was left empty, twice, on purpose.
4. What this adds to a standing argument
Readers here know the spine. This blog has argued for a year that a logic layer cannot audit itself — that a mirror cannot be its own witness (Attribution Is Not Adjudication), and that a boundary cut from the same material as the thing it bounds will hold only until an optimizer finds it worth the cost of crossing (A Wall You Can Misconfigure Was Never a Wall). The conclusion has been constant: whatever adjudicates has to stand outside the software.
This episode adds a term that the physical-layer version of the argument does not, on its own, cover. The standing worry has been that the outside seat is unreachable — that the software’s own reports and monitors, being inside the software, cannot be trusted to adjudicate it. Roslansky’s loop shows a second, quieter failure of the outside seat, one that has nothing to do with reachability: the seat can be abandoned, by the one party that was reliably outside to begin with — the human.
Earlier this blog asked, in The Voight-Kampff Test Assumed the Examiner Was Human, how anyone knows that the finger on LinkedIn’s “seems like AI slop” button — a button clicked more than a million times — belongs to a human at all. Roslansky is the answer and the reversal in one figure. The finger was human. The judgment was human. It worked. And then he passed the next judgment back to the machine. Verifying that the witness is human was never sufficient, because a witness who will not remain outside stops being a witness the moment he re-enters. Presence outside the loop is not a credential you hold. It is a position you keep paying to occupy, or lose.
5. The gradient runs toward abdication
Why would competent people vacate a seat they are plainly able to occupy? Because occupying it is not rewarded, and leaving it is.
Set Roslansky aside for a moment; this next point is reporting, not his claim, and the two should not be fused. The same industry has grown a workplace habit of treating the volume of AI tokens consumed as a proxy for productivity — “token-maxxing” — to the point that Microsoft’s own Jay Parikh reportedly had to tell staff not to burn more tokens than the work actually requires, and that the goal is meaningful output, not maximal consumption. Read the two stories together and the shape is plain, and it is structural, not a matter of anyone’s bad faith: a metric that counts how much passes through the loop rewards staying in the loop. Occupying the outside seat — reading something yourself, deciding, putting your name to a thought — is slower, and it produces fewer tokens, and it shows up on no dashboard. The same company that sells the tools which make the loop nearly free has also, internally, had to warn its people against mistaking loop-volume for work. That is not hypocrisy to be prosecuted; it is a gradient to be noticed. The default slope of the tools and the metrics runs inward, toward the loop, away from the seat.
6. The honest counterview, and the one move that points outward
The strongest objection deserves to be put plainly, because thoughtful people hold it. Perhaps this is nothing but early-days friction. People are new to these tools; norms have not caught up; give it two years and everyone will have learned that unedited generation reads as slop, and the outside seat will quietly re-occupy itself as a matter of taste and reputation. Roslansky’s own prescription — enhance with your distinct perspective — is a bet on exactly that maturation, and it may well be right.
Grant it its due, and then notice what it is asking. It asks people to do the slower, less-legible, less-rewarded thing, voluntarily, against the incentive gradient just described. This blog has watched what becomes of boundaries that run against a gradient: they hold, right up until crossing them is worth the cost, and then they are priced and crossed. A norm of “please stay outside the loop” is precisely such a boundary. It is the polite request, not the wall.
So the move that points the right way is not another exhortation to use AI thoughtfully. It is to make the outside seat structural rather than optional: to attach a named, accountable human judgment — exercised out-of-band from the generate-and-consume loop — to the decisions that carry weight, and to measure the work by whether that judgment was actually exercised, not by how much text moved through the pipe. Where that judgment sits outside the loop by design, it survives; where it is left to individual discipline against the incentive gradient, it will be abandoned exactly as often as abandoning it is cheaper. This is a claim about direction, not a product and not a proof. The narrow, hard part is only this: the outside seat will not stay occupied on its own, and anything that depends on it has to protect it deliberately — because the loop, left to its incentives, will always offer to take the seat off your hands.
7. Conclusion
The ocean of sameness did not rise up and drown anyone’s judgment. The judgment waded in. Roslansky saw the document for exactly what it was — the witness was awake, the witness was working — and then asked the machine to read it for him. Both ends of his loop had somewhere to stand outside it. Neither stayed.
That is the whole incident, and it is the whole lesson. The doom loop closes not because the witness is missing, but because both ends abdicated it. A witness who will not remain outside is not a witness. He is just one more voice inside the loop, waiting to be summarized.
Yoshimichi Kumon
Organizer, LSI Inventor, ARDS/ARKS (PCT GA26P001WO)
Visiting Researcher, Waseda BFC MIT Sloan + CSAIL
References
Roslansky, R. LinkedIn post on workplace AI-generated content (“This is a doom loop”). 31 August 2026. Original on Roslansky’s LinkedIn profile (https://www.linkedin.com/in/ryanroslansky/); reported by Windows Central, “Microsoft executive vice president bemoans employees sending him ‘AI slop,'” 31 August 2026. https://www.windowscentral.com/artificial-intelligence/microsoft-executive-vice-president-bemoans-employees-sending-him-ai-slop-saying-this-is-a-doom-loop-maybe-take-some-responsibility
GIGAZINE. “Microsoft Officeの責任者が『AIが生成した質の低い文書』に対して苦言.” 3 September 2026. https://gigazine.net/news/20260903-microsofts-office-ai/
GIGAZINE. “Microsoftが従業員によるAI利用を制限する動きに出る” (on “token-maxxing” and Jay Parikh’s internal guidance). 6 August 2026. https://gigazine.net/news/20260806-ai-tokenmaxxing-microsoft/
LSI. “Attribution Is Not Adjudication.” 1 September 2026. https://logos-sovereign.space/?p=428
LSI. “A Wall You Can Misconfigure Was Never a Wall.” 2 September 2026. https://logos-sovereign.space/?p=435
LSI. “The Voight-Kampff Test Assumed the Examiner Was Human.” 25 August 2026. https://logos-sovereign.space/?p=417
Ⅽomment