LinkedIn’s Feed Is 41% Fake and the Cleanup Starts Now

More than 40% of LinkedIn posts longer than 250 words are fully generated by artificial intelligence, according to AI detection firm Pangram. The company analyzed more than 1 million posts across LinkedIn, Medium, Reddit, Substack and X. It found LinkedIn had the highest rate of long-form AI content on any platform.

Substack posted one of the lower rates, with roughly 20% of its long-form content flagged as AI-generated or AI-assisted, Pangram said. Reddit’s overall rate was just 4.4%. That figure is skewed by replies, which were 98.1% human-written and made up most of what Pangram scanned. Reddit’s top-level posts – the ones comparable to a LinkedIn post or a Substack piece – were AI-generated at a rate similar to X.

Platforms are now racing to fight back. Substack launched an AI detection feature on July 21 built with Pangram. Readers can scan any post, note, comment or reply of more than 100 words published on or after that date. The tool shows an estimated split between human-written and AI-generated text, according to a July 21 company blog post written by Substack CEO Chris Best.

LinkedIn followed Thursday. Chief Product Officer Hari Srinivasan announced a “Seems like AI slop” reporting button, accessible from the three-dot menu on a post. “We are ramping up a series of new and improved classifiers that identify if a post is AI-slop or generally low-quality content,” Srinivasan said in a Thursday LinkedIn post. “This will reduce the amount of AI slop you might see in suggested content and content from outside your network.”

LinkedIn is also retiring its own “Enhance your post” AI writing tool for a more conservative proofreading feature that preserves a user’s original voice, Srinivasan said. The company will privately notify users when their content reads as inauthentic due to heavy AI use. LinkedIn blocks hundreds of thousands of automated comment attempts each day.

Every report becomes training data for classifiers, reducing how often flagged content appears in posts suggested to people outside a poster’s existing network.

Substack took a different approach. It put detection in the reader’s hands rather than routing everything through moderation. Readers can scan a post and see a percentage breakdown. Creators can add a “How I make this” disclosure, disable detection on individual posts, or dispute a result, according to Substack’s support documentation.

In the July 21 blog post, CEO Best framed the launch around a specific fear: that Substack could drift toward what LinkedIn had become. “When content made by no one takes over parts of the internet that are supposed to be human, it pollutes the commons and makes it hard to discover and hear human voices,” he said.

Neither company claimed certainty. Pangram’s scores are estimates based on stylistic patterns, not proof of how text was produced. Independent researchers have found that some AI-generated text can read as human to detectors.

The business logic is more straightforward than the detection technology. A social platform’s product is attention; a newsletter platform’s product is subscriber trust. Both erode when readers cannot tell a human voice from a mass-produced one.

Online news aggregator Digg shut down its Reddit-competitor product in March, citing an inability to keep pace with bot-driven content, The Verge reported March 13. Cloudflare has said automated bot traffic on the web now exceeds human traffic, a threshold it reached earlier than forecast, NBC News reported in June.

LinkedIn and Substack are betting that the platforms that survive the next phase of generative AI will be the ones that can prove a real person is still on the other side of the post.