Writing generated by artificial intelligence has spread across the main social platforms, but it has not done so in a homogeneous way. Linkedin is, according to a Pangram analysis, the network with the highest proportion of long-form content identified as generated by AI, ahead of Medium, X/Twitter Articles and Substack.
The study, prepared from anonymized data from users of the Pangram Chrome extension, has analyzed 1,002,627 publications on Linkedin, Medium, Substack, X/Twitter and Reddit. The tool only scanned content of more than 50 words found by users during their browsing, and used the Pangram 3.3 model, which the company claims has a false positive rate of 0.01%.

The research concludes that the problem especially affects long content. Among pieces longer than 250 words, one in four was marked as entirely AI-generated. On four of the five platforms analyzed, long texts were more likely to be synthetic than short posts.

In this context, Linkedin appears as the platform most saturated by content generated with artificial intelligence. In the case of long posts, 40.5% of posts of more than 250 words were identified as written entirely by AI. Another 4% were classified as mixed content, while 55% were considered human.

The professional platform represents approximately a third of the content scanned by Pangram, but accounts for 62% of all content marked as generated by AI. For the company, this indicates that users are using AI even in environments associated with real identity, professional reputation and personal brand building.

Pangram also notes differences between top posts and comments. On Linkedin, top posts are 1.35 times more likely to have been generated by AI than comments. Still, when comparing Linkedin comments with top posts on other platforms, the former are also slightly more likely to be written with AI.

The data is especially significant due to the very nature of Linkedin. The network has integrated writing assistance tools, such as the “Enhance post” button, which makes it easier for users to improve or generate texts with the help of AI. At the same time, the company has announced that it will detect and reduce the visibility of content apparently generated by AI when it considers it to be low effort or lack its own perspective. Thus, the network has recognized the deterioration caused in its feed by the so-called AI slop: publications generated with artificial intelligence that sound correct, but are repetitive, automated or lacking their own criteria. The platform claims to have developed systems capable of identifying this type of content with an accuracy of 94%, although without detailing the rate of false positives or the technical operation of the model. Its intention is to reduce the visibility of posts and comments that do not provide perspective, experience or original thinking.

Although Linkedin leads in content entirely generated by AI, X/Twitter Articles stands out for the high proportion of assisted or mixed texts. According to Pangram, 24% of long-form articles on X/Twitter were classified as AI-generated and 23% as mixed. This means that almost half of the articles analyzed on the platform contained writing generated or assisted by artificial intelligence. The remaining 53% were identified as humans.

Medium also shows a significant presence of synthetic writing in long-form content. 31% of posts longer than 250 words were identified as AI-generated, 10% as mixed, and 60% as human.

Substack appears as the platform with the lowest proportion of long texts generated entirely by AI. 10% of the publications analyzed were classified as AI, 12% as mixed and 78% as human. However, even on the best-performing platform, more than a fifth of long-form content appears to be AI-generated or assisted.

Reddit, for its part, was part of Pangram’s overall analysis, although it does not appear in the comparable long posts graph. The platform concentrated the highest volume of scanned content, with 37% of the total, but registered one of the lowest combined proportions of content generated or assisted by AI.
The explanation lies in the composition of the sample. 72% of the Reddit content analyzed were responses, and these were mostly identified as human. Specifically, Pangram notes that 98% of the responses on Reddit were classified as written by people.

However, the top Reddit posts show a different reality. 12% of these publications were identified as generated by AI, a figure closer to that observed in main publications on other platforms.
This suggests that anti-automated content strategies may be more effective at curbing mass responses or low-effort spam than at detecting AI-generated top posts, which have lower volume but greater impact within the conversation.

Long content, preferred terrain for AI

The research indicates that AI is especially affecting formats that require more textual development. The longer a piece of content is, the greater the incentive to delegate some of your writing to generative tools.
On Linkedin, this behavior translates into seemingly elaborate professional posts, but which may lack your own experience, nuances, or personal point of view. On Medium and X/Twitter Articles, AI appears to intervene in longer pieces of opinion, analysis or explanation. In Substack, although the proportion is smaller, a relevant presence of assisted texts is also observed.

Pangram uses the term “AI slop” to refer to content generated with AI that sounds correct on the surface, but that provides little value, reproduces generic ideas or does not incorporate a truly human perspective. The problem is not only the existence of AI in writing, but also its massive, undeclared and low-quality use in spaces where users expect judgment or experience.

The Pangram study does not allow us to conclude that all AI-assisted content is necessarily poor, irrelevant or misleading. Artificial intelligence can be used to structure ideas, revise style or improve the clarity of a text. But the issue becomes more complex when the tool completely replaces the experience, opinion or voice of the signer. And that nuance is especially important on platforms like Linkedin, where publishing content has become a professional positioning tool. If a significant part of long texts is generated by AI, the network runs the risk of losing part of what gives it value: the circulation of real learning, business perspectives and conversations between professionals.

Linkedin itself has recognized this situation. In a recent post, Laura Lorenzetti, Global Editorial Executive Editor of the platform, pointed out that AI can help refine language, but that excessive use of automated, low-effort content dilutes the valuable learnings that can arise from human conversations.

The conclusion of the report is that AI-generated content is no longer a peripheral phenomenon or limited to low-quality websites. It is present in social spaces where users consume information, build reputations, debate and decide what they pay attention to. The question for platforms won’t just be how much AI-generated content they allow. It remains to be seen what kind of value they want to preserve in their environments: publication volume, appearance of activity, or actual human conversation.