AI Writing Now Comes With a Paper Trail

There is a new option in LinkedIn's report menu. It reads "Seems Like AI Slop," and a post flagged with it disappears from the flagger's feed while the signal routes back into LinkedIn's detection systems.

The button is the smaller half of what LinkedIn did this month. The larger half is a subtraction: the platform is sunsetting the AI assistant that rewrote members' drafts, a feature it shipped when every platform was bolting generation onto every text box. Dan Roth, LinkedIn's editor in chief and VP of content, cited low demand and the risk of deepening the platform's AI-content problem, per Inc. What replaces the assistant proofs grammar and leaves the writing alone.

A platform that sold the shovel has concluded the digging was the problem. That is one verdict. The same week produced two more, from different directions, and the three together are the story.

An employer and a vendor start keeping receipts

The second verdict came from inside a company. Clay, the GTM data platform, made an internal AI writing policy official across the whole company, per co-founder Varun Anand. The policy was written by an engineer, Sophie Alpert, and it was scoped to engineering before other teams asked for it. Its first principle: you stand behind every idea and every sentence in a document that carries your name. Nothing in it bans a tool. All of it regulates output.

What matters as much as the content here is the direction of travel. Most corporate AI policies are compliance documents, written by legal, about what employees may not paste into a chatbot. Clay's came up from the people doing the work, and it governs what you may put your name on. It is a quality bar, not a firewall.

The third verdict came from the vendor side, and it is the least escapable. Anthropic began adding invisible watermarks to text generated by new Claude models on August 2, per Fortune. The mark works by statistically biasing word choices according to a key Anthropic holds. No single sentence looks unusual, but across enough text the pattern becomes detectable. It survives copy and paste, and Anthropic expects some marks to survive editing. The company says it will publish technical detail so institutions can check content for the signal. The move tracks a transparency requirement in Article 50 of the EU AI Act, and it applies globally, through the API and the cloud platforms where Claude is offered.

The reaction is its own data point. Forbes reported the backlash: users upset that workplaces and classes will now be able to catch undisclosed AI writing. The objection rewards a second reading. The mark doesn't misidentify human work; it correctly identifies AI work whose value depended on nobody knowing.

Anthropic is careful about what the mark proves, and the caveat is fair. Claude processes plenty of human material; a translated or summarized document carries the mark without being machine-authored. The watermark proves processing, not authorship. But provenance questions don't need courtroom certainty to change behavior. They need a plausible check, asked at an awkward moment.

Undisclosed volume stopped being free

Here is the distance between what the market says and what it does. Publicly, nearly every GTM organization credits AI with making its content operation faster, and by volume that is true. Privately, the same organizations are drowning in drafts nobody will admit to skimming and thought leadership no thought preceded. The arrangement rested on one assumption: that the reader, the employer, and the tooling couldn't tell.

In one week, all three said they intend to tell. The platform built a reporting mechanism for readers. Accountability got written into the employer's policy. And the vendor put a signature inside the text itself.

None of these is a ban, and that is what makes them effective. LinkedIn still hosts AI-assisted posts. Clay still uses AI everywhere. Anthropic still sells the model. What changed is that undisclosed volume stopped being free. Content produced faster than anyone could stand behind it now carries a discoverable cost, whether discovery arrives as a slop flag, a policy conversation, or a watermark check.

Detection is an arms race. Paraphrasing tools will chase watermarks, and a report button will be abused by people who simply dislike a post. All true, and none of it changes the economics. The volume play was profitable because detection cost the audience effort, and every one of these moves cuts the cost of asking. An arbitrage doesn't need perfect policing to close. It needs to stop being reliably cheap.

"Who wrote this" is now an answerable question

Accountability now runs through the platform, the employer and the model vendor at once, and none of the three banned a tool to get there. LinkedIn still hosts AI-assisted posts after retiring its writing assistant, Clay still uses AI everywhere, and Anthropic still sells Claude.

What each added is a way to ask who wrote something. LinkedIn routes "Seems Like AI Slop" flags into its detection systems, Clay's rule makes every author answer for each sentence in review, and Anthropic says it will publish the technical detail institutions need to check text for its mark.

By Business of GTM