Google has uploaded a research paper titled The Synthetic Gap to the website of its research division, authored by seven of its researchers. The paper introduces SAFE (Scaled Abuse Forensics Examiner), an internal tool designed to assign AI agents to investigate groups of suspected channels that publish videos generated in bulk. The document is dated 2026, though no further details are provided. The system's diagram references data from "YT Channel," suggesting that YouTube is among the first targets of this tool. The term "AI slop" refers to low-quality videos produced in bulk by AI and often published by networks of channels that work together. On YouTube, such videos are frequently encountered. In November 2025, the online video editor Kapwing reviewed the top 100 channels in each country, totaling 15,000 channels. Its report found 278 channels that only publish slop, less than 2% of the sample, but these channels collectively have over 63 billion views and approximately 117 million dollars (approximately 100 million euros) in estimated annual revenue. Kapwing also created a brand new account and found that of the first 500 Shorts suggested by YouTube, 104 were slop, roughly one in five. However, these slop channels are not supposed to rely on such content. Since July 15, 2025, YouTube has excluded content it deems "inauthentic" from its partner program and thus from advertising revenue. A follow-up review by Kapwing in the subsequent autumn shows that these videos remain very prevalent despite this rule. YouTube acknowledges the problem. Neal Mohan, CEO of YouTube, made it a priority for the year in his annual letter published on January 21, 2026. He promised to rely on systems that already combat spam and clickbait. SAFE gives an idea of what these systems might look like. If Google mentions investigation, it is because traditional detection is no longer sufficient. Slop networks publish videos that are all slightly different from each other. Tools that detect duplicates therefore miss them. The researchers refer to the "synthetic gap" as the time between the emergence of a new AI abuse technique and the arrival of a countermeasure. Today, human analysts conduct investigations manually, but they are too slow to keep up with the volume. SAFE assigns part of the work to AI agents capable of performing complex tasks independently. At the top, a "root" agent acts as the lead investigator. It receives a group of suspected channels and distributes the work among three specialized agents. It then checks what they report back before delivering its verdict. The first agent analyzes the videos and their metadata. It tracks repetitive scripts, visual inconsistencies typical of automated generation, and rule violations, such as AI identity impersonation. The second agent studies the behavior of accounts. It compares network infrastructure, device fingerprints (the technical signature of the smartphone or computer used to upload the videos) and posting times to detect synchronized or burst uploads. If suspicious patterns are found, the agent can alert by stating, for example, "100% of the channels use the same version of OS and publish within the same 5-second window." The third agent maps the links between accounts. The goal is to dismantle the entire operation in one go rather than closing the channels one by one. Another Google paper, also published in 2026, provides figures. In an initial version of their study, later removed from the Google Research website, four researchers describe a system already deployed on "a large online video platform." In six months, it would have allowed the closure of 50,000 coordinated account groups, or 130,000 channels. According to its authors, it also halved the number of human checks, saving approximately 83 hours of work. The current version no longer mentions this deployment and highlights evaluation results: 74% of cases automatically sorted and more than 1,100 hours of verification saved per week. SAFE seems to target cases that escape this first net. Its authors explain that coordination signals miss channels when their videos are not similar enough, even though their owners behave the same way. For viewers, these tools should remain invisible. The only visible markers are the labels: YouTube flags videos created with its own AI tools, and creators must declare their realistic modified or generated content. However, YouTube itself is pushing AI-generated content. In his January letter, Neal Mohan states that more than one million channels used, on average, the platform's AI creation tools daily in December. The entire challenge is to distinguish these videos from those that SAFE must uncover.