Episode blurb
LinkedIn shipped a button that lets anyone report your post as AI slop, the same day it removed its own AI writing tool. On a network where reach is proximity to a job, that isn’t moderation. It’s a permission structure, and the people most exposed are the ones who already write like everyone else because they were trained to. Also this week: Pew put a number on how much of the web is written by AI, and a federal research office said that same kind of detector can’t stand alone as evidence of anything. The institution with more at stake demanded more evidence.
Longer description
On 30 July, LinkedIn added “Seems like AI slop” to the menu on every post and removed Enhance Post, its own AI writing tool, the same day. The order matters. The platform built the tool, the tool produced the output, the platform took the tool away and handed everyone a button to report each other for it. Nobody planned that, which is exactly what makes it gravity rather than malice.
What it does to a person is the part worth sitting with. LinkedIn isn’t where people go to tell stories; it’s where they go to be employed, so a reach penalty isn’t a content-moderation outcome. It’s an economic one, applied silently, by a stranger, at no cost to them and with no appeal for you. And the people most exposed are the ones taught to write in corporate templates, the non-native speakers whose careful prose reads smooth, the juniors imitating a senior to be taken seriously.
That sits inside a bigger story. Pew Research published a census-scale estimate of how much of the web is AI-written, and the number came out of a detector. Pew did the work carefully and documented it all — the threshold, the agreement statistic, the false-positive warning, the dating limitation. What happens next isn’t Pew’s fault. It’s what always happens: the caveats fall away, and the number travels alone. Meanwhile, the HHS Office of Research Integrity looked at the same class of instrument and said it can’t be the sole basis for a misconduct finding.
And in Where I’m Wrong: I thought training was the easy part of governance. Three studies took that apart, and the one that moved me most wasn’t the one saying training fails.
In this episode
Pew’s census-scale estimate of AI-written pages, and the caveat that decides what the number means
Why only 10 to 15 percent of sampled pages could even be dated, and what that does to the headline
The HHS Office of Research Integrity guidance, and why a federal office demanded more evidence than a publishing platform did
What it means to admit a probabilistic estimator into a fact-producing role without treating it as a change
The structural middleware underneath the whole story: a 490,000-page apparatus built to approximate a fact every generator knew with certainty and discarded
LinkedIn’s slop button, the removal of Enhance Post, and the order those two things happened in
Why a reach penalty on a professional network is an economic outcome, not a moderation one
Who’s actually most exposed: the people trained to write like everyone else
Where I’m Wrong: training, and the difference between a content problem and a conditions problem
Key ideas
Organizational Gravity — organizations accumulate mass faster than they shed it. Here it shows up as a measurement admitted into the operating model with none of the questions a change would have to answer.
Governance is not measurement — the order is issue, root cause, control, then measurement. This story runs it backwards: the detector arrives first and becomes the fact.
Structural middleware — when no provenance layer connects generation to publication, somebody has to build a 490,000-page estimator to recover what was known for free at the moment of writing. That’s incomplete architecture, and the fix is integration.
The permission structure — a report button between peers on a network where reach is proximity to a job isn’t moderation, it’s a mechanism.
Sources
Segment 1 — the news hit
Pew Research Center, “How much of the internet is written with AI?” (20 August 2026) — the primary study. 490,000 English-language pages sampled across 49 Common Crawl snapshots, January 2021 to July 2026. Supports the 10 percent raw July 2026 figure, the “over one third” figure for pages with a detectable post-ChatGPT publication date, and the methodology caveats including the dating limitation. Also carries Pew’s own false-positive warning and stated threshold — worth reading before repeating either number.
HHS Office of Research Integrity, guidance on generative AI in federally funded research (14 August 2026) — the second institution. Supports the central claim that detector output may constitute evidence but that ORI recommends against relying on such tools as the sole basis for a misconduct finding. Also covers disclosure expectations and the penalty framing.
Segment 2 — the Gravity Audit
Tech Times, “LinkedIn pulls its own AI writing feature, crowdsources slop detection” (31 July 2026) — supports both halves of the sequence: the addition of “Seems like AI slop” to the post menu and the removal of Enhance Post on the same day. Also carries the reporting on how flagging feeds LinkedIn’s classifiers and affects recommended-feed distribution.
The 40-plus percent figure for fully AI-generated LinkedIn long-form content is Pangram’s research, not LinkedIn’s — attribute it that way on air and in writing.
Segment 3 — Where I’m Wrong: training
Arthur, Bennett, Edens & Bell (2003), “Effectiveness of Training in Organizations: A Meta-Analysis of Design and Evaluation Features,” Journal of Applied Psychology 88(2), 234–245 — free full-text PDF. 397 findings across roughly forty years and about 33,000 people. Supports the claim that the average trained person outscored about three of four untrained people. Note the measures are largely taken at the end of training, which is the limitation that sets up the next study.
Park, “Does Compliance Training Decrease Corporate Misconduct? Evidence From Field Data” (Harvard DASH) — five years of internal records from one multinational manufacturer, roughly 100,000 employees across about 100 countries, alongside 2,237 misconduct allegations. Supports the no-meaningful-association finding on average, the roughly two-month decay of in-person effects, the null result for video, and small classes outperforming large ones by around ten to one. Say on air that this is a dissertation, not peer-reviewed, one company, and compliance training specifically. Also available as Harvard Law Olin Center fellows paper 89.
Blume, Ford, Baldwin & Huang (2010), “Transfer of Training: A Meta-Analytic Review,” Journal of Management 36(4), 1065–1105 — free full-text PDF from an author’s site; journal version at SAGE. Eighty-nine studies, roughly 12,500 people. Supports the finding that work-environment support — manager backing, a job that rewards the new behavior — predicted transfer better than trainee motivation, which ranked near the bottom. This is the study behind “it’s the room they go back to.”
Read the full version
This week’s newsletter went out on training — all three studies, the long version, and the part I didn’t get to on air: LinkedIn newsletter
Subscribe
I’m writing a book out loud, thinking it through in public, one essay and one episode at a time. Subscribe, and you’ll watch it take shape:
Note
Stories on this show are anonymized. Names and identifying details are always changed.
