All In
FR EN
Back to blog
Algorithme & stratégie LinkedIn 17 min read

LinkedIn AI Content vs. Human Writing: The Signals the Algorithm Tracked Before the Button

| By Patrick de Carvalho

Contents

On July 30, 2026, LinkedIn added a button that lets any member report a post as "Seems like AI slop." I covered why that specific mechanism is broken in a companion piece built from primary sources alone: untrained readers score close to a coin flip at spotting generated text, and a leading commercial detector got six answers out of ten wrong when a well-known algorithm researcher tested it on his own posts. This piece asks a different question. What has LinkedIn's ranking system actually been penalizing, and does a reporting button even touch it? I have watched this platform's ranking machinery change for more than two decades, and the honest answer is that most of what gets suppressed on LinkedIn has never depended on how a sentence sounds.

In short: LinkedIn's detection models were already running before July 30, targeting suggested and out-of-network content (LinkedIn, via TechCrunch and Social Media Today). A separate, older system blocks hundreds of thousands of automated comment attempts a day and penalizes engagement pods by behavior, not prose. Meanwhile, every published test of style-based AI detectors, from OpenAI's own tool to Stanford's bias study to a Substack detector, shows they are unreliable enough that LinkedIn has good reason not to build its main defense on them.

Two systems, one label

Facts first, kept short because the full record is in the companion piece.

Hari Srinivasan, LinkedIn's Chief Product Officer, announced the "Seems like AI slop" option on July 30, 2026: a report option in the three-dot menu of posts and comments. His stated purpose, verbatim: "Slop is hard to define and the definition changes; this lets us tune our models and make better feeds." Member reports become training data for in-house detection models. Srinivasan called AI slop "a top priority for all of us."

Read that sentence again: "tune our models." Not build them. LinkedIn is not describing a new detection system launched on July 30. It is describing an existing one, fed by a new input. That distinction is the entire subject of this piece.

The detection models were already running

The button's own announcement says where the tuning applies first: suggested and out-of-network content, the layer LinkedIn controls most directly through recommendation rather than through a member's own following graph. That targeting choice only makes sense if a detection layer already existed to receive the new signal.

It did. On the same day it introduced the report button, LinkedIn removed its own AI writing feature, "enhance your post," a tool it had designed and promoted. Its replacement is described in one sentence: a feature that "proofreads your words, but does not change your voice." Building, shipping, and then retiring an AI writing assistant is not something a company does the week it starts thinking about detection. It is something a company does after years of watching what that assistant produced.

The volume context supports the same reading. LinkedIn states it blocks hundreds of thousands of automated comment attempts every day, and billions of automation attempts over recent months. Those numbers describe a mature enforcement pipeline, not a pilot. A platform does not discover, on a Tuesday in July, that it needs to stop billions of automated actions. It has been running that fight for a long time, and the AI slop button plugs a new, noisy data source into machinery that predates it.

The system that actually catches manipulation: engagement pods

The clearest evidence of a working, behavior-based detection system on LinkedIn has nothing to do with AI writing. It is the fight against engagement pods.

LinkedIn defines pods publicly as groups of people, often real members, who coordinate to trade likes and comments and artificially inflate how long their content stays visible. In March 2026, months before the AI slop button, LinkedIn escalated its response with wider detection and reach limits on accounts caught coordinating, adding internal labels on content whose visibility looked artificially amplified. Gyanda Sachdeva, VP of Product Management, named the actual target: any third-party tool, browser extension, or plugin that automates engagement manipulation.

Notice what that system measures. Not the words in a comment. The pattern across accounts: who engages with whom, how fast, how often, in what sequence, using what tooling. A pod can be built entirely from typed-by-hand, perfectly human comments, and LinkedIn's detection still catches it, because the tell is coordination, not composition. That is a fundamentally different kind of signal than "does this sentence sound machine-written," and it is one platform engineers can verify with far more confidence: account graphs and timing patterns leave a trail; prose style does not.

This is the detection layer that has been quietly doing the heavy lifting since before generative AI became a talking point on the platform. It is also the layer the AI slop button does not touch. Reporting a single post as "seems like AI slop" tells LinkedIn nothing about whether the accounts behind it are coordinating. It tells LinkedIn what one reader felt in three seconds.

Why style is the wrong foundation to build a defense on

There is a documented reason LinkedIn would rather expand pod-style behavioral detection than bet its ranking system on prose analysis: every independently tested style-based detector has a public failure record.

OpenAI shut down its own AI-text detection tool on July 20, 2023, citing a low rate of accuracy: 26% of AI-written text correctly identified, 9% of human-written text falsely accused. The company that trains the most widely used language models on earth gave up on detecting their own output from style alone. Turnitin, which processes on the order of 100 million student submissions a year, claims under 1% false positives at the document level but concedes 4% error at the sentence level, and Vanderbilt University turned its AI detection off entirely rather than live with that margin.

The bias runs deeper than a raw error rate. Weixin Liang's team at Stanford ran student essays through seven commercial GPT detectors in 2023 and found that essays by non-native English writers were flagged as AI-generated 61% of the time, with all seven detectors wrong in unison on roughly one essay in five. The mechanism is structural: careful, disciplined writing in a second language produces the same narrow lexical variety detectors learned to associate with machines. A platform whose working language is English but whose members write it as a second, third, or fourth language in most of the world's 200 countries cannot safely automate a penalty on that signal.

Even the vendor whose figures dominate the current conversation says so itself. Pangram Labs, which published the widely cited estimate that 40.5% to 41% of LinkedIn's long-form posts are fully AI-generated (a figure collected through an opt-in Chrome extension, meaning it measures the feeds of people who volunteered to install an AI-detection tool, not a random sample of the platform), publishes its own list of blind spots. The company advises against running its detector on bulleted lists, short posts, and templated writing, three formats that describe a large share of everyday LinkedIn content. When Richard van der Blom, who publishes the reference annual study on the LinkedIn algorithm, tested a comparable detector on his own writing, he got six wrong answers out of ten, including a 95%-human post flagged as 100% AI.

Put those numbers side by side and the engineering conclusion is obvious. A detection strategy built on how text reads will misfire on your most careful writers and your shortest, most scannable posts alike. A detection strategy built on how accounts behave, coordinated timing and automation tooling above all, does not carry that specific failure mode. That is very likely why LinkedIn's oldest and most defensible enforcement system targets pods and automated comments, not prose.

The criterion LinkedIn actually wrote down

LinkedIn did put a criterion in writing, and it is not about machines. It is in the ten words describing the tool that replaced "enhance your post": it "proofreads your words, but does not change your voice."

That sentence names authorship, not authorship's tools. A proofreader that corrects grammar without replacing an author's own phrasing and judgment is doing what a competent editor has always done. What the old "enhance your post" feature did, and what LinkedIn apparently decided was the actual problem, was produce a polished, generic paraphrase that erased whatever made the original identifiable as one person's writing.

That is a style criterion, but it is not the same style criterion the AI slop button measures. "Does this still sound like you" is answerable by the one person who can check it: the author. "Does this look machine-made to a stranger scrolling past" is not, for the reasons in the section above. LinkedIn wrote the checkable criterion into a proofreading tool and shipped the uncheckable one as a public reporting button, on the same day.

What gets flagged, and what does not: a signals table

Laid out side by side, the two kinds of signal LinkedIn is working with point in different directions.

Signal What it measures Reliability, as documented Who or what enforces it
Coordinated engagement (pods) Account-graph patterns, timing, third-party automation tools High confidence; behavioral trail is verifiable LinkedIn's detection models, reach limits, internal labeling, escalated March 2026
Automated comments Volume and pattern of scripted posting High confidence; hundreds of thousands blocked daily LinkedIn's automation-blocking pipeline
Prose style ("sounds like AI") Word choice, sentence rhythm, lexical variety Documented as unreliable: 61% false-flag rate on non-native writers (Stanford), 26%/9% accuracy on OpenAI's own retired tool, 6 wrong answers out of 10 in van der Blom's test Member reports since July 30, feeding detection models still being tuned
Voice preservation Whether a proofreading pass changes an author's own phrasing Self-assessed by the author LinkedIn's new proofreading-only writing tool

The pattern is not subtle. Where LinkedIn can verify behavior, it enforces with reach limits and account-level labeling. Where it is asking about style, it is still collecting human opinions to train a model, eight days after launch, with no published false-positive rate, no reporting threshold, and no appeal process.

What regulators check that the button doesn't

Governments that have actually legislated on AI-generated content picked the same kind of criterion LinkedIn's pod detection uses: behavior and responsibility, not the impression a sentence gives a reader.

Article 50 of the EU AI Act, in force since August 2, 2026, requires marking synthetic content, but it carves out an explicit exemption where substantial human editorial review applies, meaning a competent person can approve, modify, or reject the content before it is published. That is a review-process test, checkable by looking at who signed off on a piece of writing, not a stylistic guess about how it reads. California's AI Transparency Act, which took effect the same day, goes further in the opposite direction: it explicitly excludes AI-generated text from its labeling requirements altogether, covering only image, video, and audio. The United Kingdom has no statutory labeling obligation for AI text as of this writing, but its own government consultation drew an explicit line between wholly AI-generated content and AI-assisted work, the exact distinction a one-click "seems like AI slop" report cannot draw. The full three-jurisdiction breakdown, including penalty amounts, lives in the companion piece; the point that matters here is narrower: everywhere lawmakers looked closely at this question, they landed on process and review, the same category of signal LinkedIn already enforces through pod detection, not on how a sentence sounds.

Building content that holds up, regardless of the button

None of what follows is a way to avoid detection, and I want to be direct about that distinction because it is the whole argument. Calibrating a post against a detector, any detector, still leaves you with calibrated content; only its shape changes, and the moment a model updates, the calibration stops working. What holds up is built for a different reason: it survives a coordination check, and it survives a human reader, because there was never anything generic to catch.

Own a number nobody else can cite: how many candidates actually replied to your last hiring post, what your churn looked like the quarter you changed pricing, how long a specific deal actually took to close. A pod cannot manufacture that kind of fact, because a pod trades in reactions, not in operational detail, and a generic paraphrase strips it out the moment it smooths the sentence.

Comment like a person who read the post, not like an account running a schedule. LinkedIn's pod detection watches timing and pattern across accounts. The single best way to never resemble that pattern is to actually read what you are responding to and write a reply specific to it, which also happens to be the only kind of comment a reader finds worth reading.

Keep your own writing tools honest about what they touch. If a tool corrects grammar and keeps your phrasing, that is what LinkedIn's own replacement for "enhance your post" is built to do, and it matches the criterion the platform wrote down. If a tool rewrites your point of view into a smoother, more generic one, you have reproduced the exact failure LinkedIn retired its own feature for causing, whether or not any detector ever flags it.

Disclose your process if you have one. It will not protect a post from a bad-faith report; nothing does, and no reporting threshold has been published. It puts you on the side of the criterion every jurisdiction that has legislated on this question actually chose: a named, accountable author, not an anonymous guess about how a sentence was produced.

What I don't know

I don't know how much overlap exists between LinkedIn's pod-detection signals and the new AI-slop training data, LinkedIn has not said, and no independent audit of that internal system exists. I don't know the false-positive rate of the detection models being tuned by member reports, because LinkedIn has not published one, for this feature or for its older automation-blocking pipeline. I don't know whether behavioral detection will keep outperforming style detection as tooling changes on both sides; that is a reasonable bet based on the record above, not a certainty.

Judge the behavior, not the prose

LinkedIn already runs a detection system that works, catching engagement pods and automated comments by behavior, at a scale of hundreds of thousands of actions a day, with reach limits and account labeling to show for it. It also just launched a second system, built on member impressions of how writing sounds, in a category where every independently published test says humans and machines both perform close to guesswork.

The platform wrote the working criterion into a proofreading tool nobody argues about: keep the author's own voice. It wrote the unreliable one onto a button any member can click on impulse, with no threshold and no appeal in sight. Build for the first criterion. It was already there before July 30, and it will still be there after this button's next redesign.

FAQ

Does LinkedIn's algorithm automatically detect AI-written text?

LinkedIn has detection models in place, targeted first at suggested and out-of-network content, and it states it is using member reports from the "Seems like AI slop" button to tune them. No accuracy or false-positive rate for that specific system has been published as of this writing.

What are engagement pods, and how does LinkedIn treat them differently from AI slop?

LinkedIn defines pods as groups of members who coordinate to trade likes and comments and artificially inflate visibility. Detection there is based on account-graph and timing patterns rather than the content's wording, and LinkedIn escalated enforcement, with reach limits and internal labeling, in March 2026, months before the AI slop button existed.

Is AI-assisted writing penalized on LinkedIn?

LinkedIn's own stated criterion, described when it replaced its "enhance your post" feature, is whether a tool changes the author's voice, not whether AI was involved at all. A proofreading pass that keeps an author's phrasing matches that criterion; a rewrite into generic, polished prose does not, independent of any detector's verdict.

Can AI detectors reliably tell human writing from machine writing?

Published results say no. OpenAI's own retired detector correctly identified only 26% of AI text while falsely flagging 9% of human text. A 2023 Stanford study found seven commercial detectors flagged non-native English writing as AI-generated 61% of the time. A tester who publishes the leading annual LinkedIn algorithm research got six wrong answers out of ten from a comparable tool.

What actually gets a LinkedIn post's reach reduced, if not the words themselves?

The best-documented, longest-running penalty on LinkedIn targets coordinated behavior: engagement pods and automated commenting, both caught through account-level patterns rather than prose analysis. The newer AI slop reporting system is style-based and, as of this writing, has no published trigger threshold or accuracy figure.

Should executives stop using AI tools to draft LinkedIn posts?

Nothing published says AI-assisted drafting itself is penalized. What is documented as unreliable is judging AI involvement from how a sentence reads. A safer target than avoiding AI tools is avoiding generic output: keep a data point, a specific case, and a position that only you could have written, regardless of which tools helped write the sentence.

Sources

  1. TechCrunch, 07/30/2026: announcement, placement, blocked automation volumes, "enhance your post" removal.
  2. Forbes, Cody Luongo, 07/30/2026: Srinivasan verbatims, "AI slop is a top priority for all of us."
  3. Fortune, 07/31/2026: blocked automation attempt volumes.
  4. Social Media Today: engagement pod definition, Gyanda Sachdeva verbatim, March 2026 escalation, reach limits and internal labeling.
  5. Richard van der Blom, LinkedIn post, late July 2026: personal test of a Substack-integrated AI detector, six wrong answers out of ten, non-native-speaker context.
  6. Pangram Labs report, 07/09/2026, corroborated by The Register and Fast Company: the 40.5-41% figure, 1,002,627-post sample, opt-in Chrome extension collection method, and Pangram's own documented blind spots on bulleted lists, short text, and templated writing.
  7. OpenAI, detector shutdown, 07/20/2023: 26% true positives, 9% false positives.
  8. Liang et al., Stanford, 2023: seven-detector bias study, 61% false-flag rate on non-native English writing.
  9. Turnitin (official blog): under 1% document-level false positives, 4% sentence-level error, roughly 100 million submissions screened per year.
  10. EU Regulation 2024/1689 on artificial intelligence: article 50 transparency obligation and its substantial-human-editorial-review exemption, in force since August 2, 2026.
  11. Morgan Lewis, California AI Transparency Act guidance, August 2026: explicit exclusion of AI-generated text from CAITA's labeling requirements.
  12. House of Commons Library, research briefing CBP-10467, 2026: absence of a UK statutory labeling obligation, and the consultation's distinction between wholly AI-generated and AI-assisted content.
  13. Merriam-Webster, 2025 word of the year: definition of "slop" as low-quality content produced in volume by AI.

All In: build on the signal that was already checkable

Coordination leaves a trail. A sentence's rhythm, mostly, does not. The distinction is worth knowing before you spend a week trying to sound more human instead of building the one thing no detector, human or machine, can take away from you.

All In is the B2B media that decodes LinkedIn, expert blog, weekly podcast and newsletter for SME leaders and sales directors who want to turn LinkedIn into measurable growth. An original creation by Patrick de Carvalho, on LinkedIn since 2004. Motto: "I Never Lose."

Discover All In and get what matters on LinkedIn, every week.