Every Company That Sold You an AI Writing Tool Is Now Selling the System That Marks (and Punishes) the Output

On July 30, LinkedIn added a "Seems like AI slop" option to the reporting menu behind every post. Further down the same announcement, the company said it was pulling "enhance your post," the AI writing feature it had spent two years pushing at its own members, and replacing it with a proofreader that leaves a draft's phrasing alone.

Microsoft's professional network built the drafting machine, watched what came out of it, and then handed everyone a way to report the output. That reads like hypocrisy and mostly is not. It is what happens when the whole stack reverses at once, and the reversal is further along than most content teams have noticed.

Who is doing this

Start with the purest case. Grammarly sells GrammarlyGO, which drafts and rewrites on request. Grammarly also sells an AI detector. Grammarly also sells Authorship, which sits inside Word, Google Docs and Canvas and logs whether each passage was typed, pasted, or produced by a model, then generates a report you can hand to whoever is asking. The company describes the pairing as responsible AI use. Generation, detection and provenance, from one vendor, billed on the same account.

The pattern repeats up and down the stack. Adobe built Firefly and also founded the coalition behind Content Credentials, the signed manifest that says a file came out of Firefly. Google embeds SynthID in everything Imagen and Veo produce and reported in May having watermarked more than 100 billion files since 2023. OpenAI took its own AI text classifier offline in July 2023 because it caught only about a quarter of machine text while flagging roughly one in eleven human passages, then joined the C2PA steering committee in May 2026 and shipped a public verification tool for its own image output. Anthropic started embedding an invisible watermark in Claude's text on August 2, applied worldwide, across the app, the API, Claude Code and every cloud reseller.

The distribution platforms are reading those marks. YouTube said in May it would use internal signals plus C2PA metadata to auto-label undisclosed photorealistic AI, with the disclosure moving below the player on long-form and into an overlay on Shorts. TikTok, Meta, LinkedIn and Pinterest all read Content Credentials at upload. Pinterest rolled out GenAI labels globally in spring 2025 after complaints that synthetic imagery had swamped its beauty and art categories, added "see fewer" controls in October, and built an appeals route, which tells you what it expected the classifiers to get wrong. Spotify says it pulled 75 million bulk uploads and duplicate tracks over twelve months. Substack integrated Pangram on July 21.

A good deal of this is not voluntary. Article 50 of the EU AI Act became enforceable on August 2, requiring providers to mark synthetic output in a machine-readable way, with penalties up to €15 million or 3% of global turnover. Anthropic's timing traces directly to it. So does much of what shipped this summer.

Text is the part that does not work

Images and video now have real provenance. A cryptographic manifest travels in the file and a watermark survives the screenshot, and between the two you can usually establish where something came from without guessing.

Text has almost none of that. It gets paraphrased, translated, chopped up and folded into someone else's paragraph, which destroys the signal. Anthropic is the first frontier lab to ship production text watermarking at all, and the company told TechCrunch that whether a mark survives depends on length and how heavily Claude edited. A heavy rewrite or a translation knocks it out.

Which is why the text platforms ended up somewhere cruder. LinkedIn could not read a manifest off a post, so it built a button and asked a million people to press it. Substack could not either, so it licensed a probabilistic classifier. Neither is provenance. Both are inference, and inference is where the trouble starts.

About that 40 percent

On August 20, LinkedIn chief product officer Hari Srinivasan posted a progress report. More than a million people had used the button, he said, and members were seeing roughly 40% fewer views of the material LinkedIn's classifiers tag as slop.

Inside six days the figure had run through Slashdot, Cybernews, BigGo, The Register, PCWorld and Fortune, picking up damage. Srinivasan was describing what audiences get served. The Register printed a version about what happens to authors who copy and paste AI-written posts. Those connect only if you assume a mechanism LinkedIn has never published, and the second is the one circulating in agency decks as evidence that the platform docks AI-assisted posts 40% of their reach. We traced that reversal in detail last week.

The company said nothing of the sort. Asked about the July announcement, a LinkedIn spokesperson pointed Fortune back at Srinivasan's post, which is roughly where the disclosure ends. Communications did clarify one detail to Moneywise, and it cut in the company's favour: the million counts unique members who used the feedback flow rather than total clicks. A million distinct people is a far harder number to manufacture than a million clicks, which a few hundred motivated users could run up over a weekend.

There is a real mechanism underneath. Inc. reported at launch that LinkedIn would use the feedback to determine how much reach a post gets outside the poster's network, and TechCrunch noted the author gets a private note in their analytics dashboard when the content reads as inauthentic. That fires on a crowd report, not on the presence of a model in the workflow.

Nobody has defined slop

Srinivasan has been unusually direct that the category is unstable, telling Gizmodo that slop is hard to define and the definition changes, which is the stated reason the button exists at all, since user reports are what tune the models. Chris Best, in a Substack post titled "Against Claudefishing", allowed that Pangram can only detect whether AI was used, not whether care went into the writing, and put the failure elsewhere, in the gap between what a reader assumes about who wrote something and what they got.

So reach is being reduced on a category the platform admits it cannot pin down, tuned by a million people who each brought a private definition of it, and benchmarked against outside detectors answering the narrower question of who did the typing.

The detectors do not agree with each other either, and the habit of quoting their numbers side by side as though one has to be lying misses what is going on, which is that Pangram counted only posts past 250 words that came back fully machine-written and arrived at 41% of LinkedIn's long-form public posts, while Originality.ai reviewed 5,000 public July posts, took anything over 100 words, called it likely AI once the model cleared 50% confidence, and arrived at 81.2% showing more than moderate use. Both measured accurately. They were asking different questions. Pangram's cross-platform finding is the more useful one for B2B anyway: X came in at 29%, Reddit at 13%, and LinkedIn produced 62% of all the AI content the firm flagged while making up about a third of the scan.

The 81% is the number headed for a thousand vendor decks, where it will mean something a good deal broader than it meant in the study.

The safe harbour is closing

Every company in this story has said the same reassuring thing. Laura Lorenzetti, LinkedIn's executive editor, wrote in May that using AI to help you write is fine as long as posts and comments represent your voice and your perspectives. Best was explicit that Substack is not against people using AI to assist their work. Srinivasan told reporters that AI and slop are not the same thing, and that many people refine thoughts with AI.

Assistance is the permitted category. Look at what is being built and assistance is exactly what the infrastructure has stopped being able to protect.

Anthropic's own support page frames the mark as a signal about where content came from, and the company says it shows Claude had a hand in a piece of text rather than that Claude wrote it. Fortune noted that asking the model to proofread or translate a paragraph can be enough to leave a trace; Forbes put it more bluntly, reporting that the mark can appear even when Claude only corrects your spelling. Substack's Pangram scan returns AI-assisted as its own reported category, sitting between human and AI-generated on the same scale. LinkedIn's replacement for its rewriter is, specifically, a proofreader.

Nobody has run the obvious experiment on the other end of this, which is whether accepting Grammarly's ordinary suggestions raises your score in Grammarly's own detector. The mechanism would be straightforward if it exists, since the product pushes you toward shorter sentences, plainer words and fewer hedges, and flattened sentence variation is what classifiers read as machine writing. It is an inference, not a finding. Someone should test it.

The error that costs a brand something

None of this apparatus asks whether a post is true. A vendor can publish a capability claim its product cannot deliver, typed by hand, and nothing here will touch it. The FTC has started fining companies over that kind of claim in B2B marketing, a separate regime with real penalties and no detector anywhere in it.

The expensive failure is the false positive. Classifiers read predictability and vocabulary range, and careful writing in a second or third language reads that way. For a company page or a ghostwritten executive program, a flag works as a mark only the wearer can see, applied by a classifier nobody outside the building can audit, against a definition the platform concedes keeps moving, and LinkedIn has not published a route back from one.

TikTok labeled videos by Nikolai Savic, a creator with five million followers, as AI-generated. He had edited them himself. He told the Wall Street Journal it damaged his credibility with his own audience.

What a content team can actually do

Ghostwriting programs carry the most exposure, because they manufacture the exact pattern the classifiers were built to find. Run one hook formula across five executive accounts and you get five posts with a shared statistical fingerprint. Separate voice profiles per person, plus retiring a hook template once it has been used, will do more than any amount of prompt engineering.

Disclosure is now infrastructure rather than confession. Substack shipped a "How I make this" statement so writers can describe their own process, along with the ability to scan drafts before publishing and to contest a scan they think is wrong. Grammarly built Authorship for the same purpose. The model layer is marking output whether or not you say anything, and the EU has made some of it mandatory. Deciding your own disclosure line before a classifier decides it for you is cheaper than the alternative.

Beyond that it comes down to putting something in the post only your company could have put there, which in practice means a number out of your own data, or an internal decision that went badly, or a sentence somebody transcribed off a customer call. Detectors reward unpredictability. Proprietary detail is where it comes from.

The bill

The assist products sold speed to people who needed volume, and they worked. The marking products sell trust to readers who have started to wonder, they are partly a compliance obligation now, and they only became necessary because the first set worked. Cloudflare reports that automated requests outnumber human ones across the web. Best has said his fear is Substack turning into LinkedIn. Digg shut down its Reddit competitor in March, blaming bot traffic it could not get under control.

Both product lines keep shipping. Neither vendor absorbs the cost of the reversal, which lands instead on whoever believed the first pitch, built a process around it, hired against it, and is now sitting on a quarter of output inside a category nobody will define.