allworld.mediaNews from all over the world
LinkedIn reports 40% reduction in low-quality AI-generated posts after launching reporting button
Read more
Tecnoblog
tecnoblog.net

LinkedIn reports 40% reduction in low-quality AI-generated posts after launching reporting button

LinkedIn announced that it has managed to decrease by 40% the number of views on low-quality posts produced by artificial intelligence. This significant reduction occurred just a few weeks after the social network made available a specific reporting mechanism for this type of content. The purpose of this initiative is to purify the feed, penalize materials that do not add value, and subsequently notify those responsible for the posts that are flagged by the public.

Hari Srinivasan, LinkedIn's product director, confirmed that users are already observing a considerable decrease in automated texts. The new moderation tool, informally called the 'AI junk button' or 'AI slop' in technological jargon, has been rapidly adopted, being triggered in over one million instances since its launch.

Despite this initial success, the company cautions regarding the method of processing user reports to downgrade what it classifies as 'low-quality content.' To prevent abuse or undue persecution, systems have been implemented that prevent an excessive volume of negative criticism from being used to unfairly attack or silence other members.

This implies that simply clicking the report button does not guarantee that a post will be hidden if the internal system does not confirm the suspicion of improper AI use. Furthermore, the platform has made algorithmic adjustments, beginning to identify and limit the reach of posts that exploit textual clichés already recognized as hallmarks of generative AI systems, such as those that organize arguments in predictable patterns, exemplified by the format 'it's not about X, it's about Y.'

The Microsoft network has also integrated a transparency system. Users will receive a notification whenever one of their posts is reported by third parties as 'possible generic artificial intelligence text.' This alert will not be made public, remaining restricted to the 'post analysis' panel within the application.

A recent study conducted by the AI detection company Pangram indicated that more than 40% of extensive posts on LinkedIn were likely created by artificial intelligence. This unprecedented volume was not accidental; the application itself previously encouraged the use of the technology, even offering a button in the creation interface that suggested 'enhancing' posts using writing algorithms.

However, this ease ended up harming the platform's identity. Recognizing the negative impact on user experience, the company decided to completely remove this feature once it launched the current reporting tool.

Similar stories

Research indicates that one-third of websites created after ChatGPT show signs of artificial intelligence
Read more
tecnoblog.net

Research indicates that one-third of websites created after ChatGPT show signs of artificial intelligence

The launch of ChatGPT in 2022 marked a turning point for artificial intelligences, which have become increasingly prominent in various types of content on the internet. An investigation conducted by the Pew Research Center provided data to support this claim, revealing that since the introduction of the OpenAI model, one in every three registered websites contains elements of AI to some degree, especially in published texts.

This survey analyzed recent links and identified a significant amount of content exhibiting linguistic characteristics very common in responses generated by Large Language Models (LLMs). Among these characteristics are the excessive use of em dashes (–), the use of the Oxford comma, considered unnecessary for textual clarity, and the insertion of vocabulary excessively ornate for the specific context.

As reported by TechCrunch, which covered the study, these findings emerge shortly after another research conducted by Cloudflare, which pointed out that bot traffic on the internet surpassed that of human users, an event that occurred earlier than predicted by the platform. This suggests an increase in the number of AIs conducting online searches and consuming materials written by artificial intelligence.

The Pew Research Center's research detailed the use of AI on the web in recent years. Of the ten thousand domains monitored currently, at least 10% were produced with some contribution from AI. It is important to note that this data covers the total universe, including pages prior to the release of ChatGPT in 2022. However, when focusing on a more recent period, it was found that 35% of publications incorporate artificial intelligence into their creation.

Other data presented by Pew indicate a 10% higher prevalence of AI in domains with the ".com" extension. However, texts with LLM markers were also detected in ".edu" and ".gov" domains, which are generally associated with educational institutions or US government agencies. ".org" domains were also examined, registering 4.6% of productions made with the aid of artificial intelligence.

It should be noted that the analysis was restricted to English-language content. The links were collected through the Common Crawl platform, and AI detection was performed using the Open Pangram tool. Although TechCrunch mentioned that the tool can make detection errors, this does not prevent a macro-scale assessment, given the vast quantity of pages analyzed.

AI-generated publications have peculiarities that signal the use of the technology. The texts frequently present common flaws in AI response generation, such as the exaggeration in the use of em dashes (–), superfluous commas, and superficial conclusions, exemplified by phrases like "that's it, therefore, it's not that".

It is relevant to mention that some of these patterns are common in certain types of production, such as in journalistic articles – where the author expresses a preference for using em dashes to structure ideas. Similarly, scientific articles and academic publications may also contain such flaws, in addition to tending to detail explanations extensively to provide maximum information.

The distinction lies in the contexts where these flaws appear, as they are usually used unnecessarily. The Pew study points to the growth in the frequency of these linguistic peculiarities as proof: there was a 63% increase in the use of optional commas in texts, double the common words in AI responses, and three times more negative parallelism.

Cases of AI Slop have become more frequent on the internet, and not just in this scenario. TB recently reported on the challenges faced by Legislative Consultants in the United States Congress, who spend a large part of their time correcting bills with typical AI flaws. Simultaneously, social networks like LinkedIn are seeking to mitigate this situation by reducing the proportion of AI-generated posts.

Artificial intelligence companies themselves are aware of the massive presence of content produced by their models on the web and use printed books released before 2022 as training material for their AIs in increasing volumes. To this is added the Cloudflare research, which highlights bot traffic surpassing human searches.

Popular