The internet is entering a strange phase. For two decades, the main challenge for digital platforms was managing an overwhelming amount of content produced by people. Today the nature of the problem has shifted. What is growing fastest is no longer the number of humans posting online, but the number of machines doing it. Bots write articles, generate videos, compose music, post comments and even simulate debates. In many cases they do this faster, cheaper and at volumes no human team could realistically match. The result is a phenomenon that some researchers have begun to describe as “algorithmic pollution”: platforms saturated with synthetic material that looks authentic but does not come from human experience. Inside technology companies and newsrooms, a question is quietly becoming unavoidable. How do you separate what humans made from what machines generated? Behind the scenes of the internet, the race to answer that question has already started. ### <br>The invisible avalanche The numbers alone reveal how serious the situation has become. Music streaming platforms, for example, are receiving tens of thousands of tracks created with artificial intelligence every single day. In a recent disclosure, Deezer reported that it detects more than **60,000 AI-generated songs daily**, representing roughly **39% of all uploads** on the platform. According to the company, up to **85% of those tracks appear to be fraudulent**, created to manipulate royalty systems or recommendation algorithms. [https://www.theverge.com/news/870186/deezer-ai-music-detection-commercially-available](https://www.theverge.com/news/870186/deezer-ai-music-detection-commercially-available) Music is only one part of the story. A report from the research group AI Forensics identified **hundreds of automated accounts on TikTok producing synthetic content at industrial scale**, generating billions of views every month. Many of these accounts post dozens of videos per day and rarely disclose that the material is artificial. [https://www.theguardian.com/technology/2025/dec/03/anti-immigrant-material-among-ai-generated-content-getting-billions-of-views-on-tiktok](https://www.theguardian.com/technology/2025/dec/03/anti-immigrant-material-among-ai-generated-content-getting-billions-of-views-on-tiktok) For platforms this creates a structural dilemma. Recommendation algorithms were designed to reward volume, engagement and consistency. Artificial intelligence excels at all three. Left unchanged, the systems that decide what goes viral may end up promoting machines over people. ### <br>The first battlefield: algorithmic detection The first response from platforms has been to try to detect AI-generated content using specialized machine learning models. Tools such as GPTZero, Copyleaks and similar detectors analyze statistical patterns in text, images or audio. Two concepts often appear in this type of analysis: perplexity and burstiness. In simple terms, human writing tends to include unpredictable variations in vocabulary and sentence structure, while text produced by language models often follows more statistically predictable patterns. [https://en.wikipedia.org/wiki/GPTZero](https://en.wikipedia.org/wiki/GPTZero) These detectors process massive datasets and estimate the probability that a piece of content was produced by a machine. Educational institutions, publishers and online platforms have already started using similar systems to flag potentially synthetic text. Another approach focuses on semantic comparison. Platforms such as Copyleaks analyze style and meaning across large databases to identify patterns typical of automated generation. [https://en.wikipedia.org/wiki/Copyleaks](https://en.wikipedia.org/wiki/Copyleaks) But there is a major limitation: detection systems are far from perfect. Recent studies show that many of these tools produce substantial error rates, sometimes labeling human writing as artificial while failing to detect actual machine-generated content. In controlled tests, some detectors struggled to reliably identify material that was clearly generated by language models. [https://www.mdpi.com/2078-2489/16/10/904](https://www.mdpi.com/2078-2489/16/10/904) In practice, relying solely on automated detection may not be a stable long-term solution. ### <br>The second strategy: invisible watermarks A more ambitious idea is gaining traction in research labs and large tech companies. Instead of trying to detect AI output after publication, the content itself could carry an invisible signature. The concept is known as watermarking. Google, for example, introduced a system called SynthID that embeds invisible digital markers into images, audio or text generated by its AI tools. Specialized detectors can then identify those markers and confirm the origin of the material. [https://timesofindia.indiatimes.com/technology/tech-news/google-introduces-synthid-detector-to-identify-ai-generated-content-at-google-i/o-2025/articleshow/121337620.cms](https://timesofindia.indiatimes.com/technology/tech-news/google-introduces-synthid-detector-to-identify-ai-generated-content-at-google-i/o-2025/articleshow/121337620.cms) Think of it as a kind of digital DNA. If widely adopted, platforms could instantly verify whether a piece of content originated from an AI generation system. The obstacle is obvious. For this approach to work globally, most major AI developers would need to adopt compatible standards. So far, the industry is far from that level of coordination. ### <br>The third path: tracking identity and authorship Some platforms are experimenting with methods inspired by copyright enforcement systems. YouTube, for instance, has been expanding tools designed to detect when a creator’s face or voice is replicated using artificial intelligence. Creators can register their likeness in the system, allowing algorithms to detect deepfakes or unauthorized simulations. [https://www.axios.com/2025/09/16/youtube-ai-likeness-detection-deepfakes](https://www.axios.com/2025/09/16/youtube-ai-likeness-detection-deepfakes) The model is similar to YouTube’s Content ID system used for copyrighted music. Instead of asking whether something is artificial, the system asks whether a real person is being imitated. While useful for protecting individuals, this approach addresses only a narrow slice of the broader problem. ### <br>Structural limits may become unavoidable Even with increasingly sophisticated algorithms, many platform engineers are beginning to acknowledge an uncomfortable possibility. Technology alone may not be enough to filter out synthetic content. As a result, conversations inside the industry are slowly shifting toward structural limits within the platforms themselves. Some of the proposals circulating include restricting posting frequency per account, requiring identity verification for high-volume publishers, reducing algorithmic reach for accounts that publish at inhuman speeds, and even separating feeds for human-created and AI-generated content. Music platforms have already taken steps in that direction. Some services automatically remove or demonetize tracks identified as AI-generated when they appear to be manipulating royalty systems or recommendation engines. [https://www.theverge.com/news/870186/deezer-ai-music-detection-commercially-available](https://www.theverge.com/news/870186/deezer-ai-music-detection-commercially-available) In some internal discussions, companies are even considering banning synthetic content entirely in specific categories. ### <br>The human factor remains critical Ironically, as artificial intelligence becomes more powerful, many studies suggest that human moderation remains essential. Automated systems are extremely effective at scanning massive volumes of content, but they often struggle with context, cultural nuance and subtle forms of manipulation. Recent research shows that hybrid moderation models combining automated classification with human review significantly improve detection of problematic content patterns. [https://arxiv.org/abs/2512.03553](https://arxiv.org/abs/2512.03553) In other words, even in the age of AI, the infrastructure of online platforms still depends heavily on human judgment. ### <br>The arms race is only beginning There is a deeper irony at the center of this entire issue. As detection systems improve, tools designed specifically to evade them are evolving just as quickly. Software already exists that rewrites AI-generated text in ways intended to bypass automated detectors by introducing linguistic variability that mimics human writing. [https://en.wikipedia.org/wiki/Undetectable.ai](https://en.wikipedia.org/wiki/Undetectable.ai) The dynamic resembles the early days of spam. For years email filters evolved while spammers developed increasingly sophisticated ways to circumvent them. Now the same cat-and-mouse game is beginning to play out across the entire internet. The difference is scale. The web was originally built to amplify human voices. But as machines begin producing virtually unlimited volumes of plausible content, the challenge becomes larger than technology alone. If every platform can be flooded with synthetic articles, videos, music and opinions, the real question may not be how to detect artificial content, but how to preserve spaces where authentic human expression still carries weight. And that leads to a question that engineers, journalists and researchers are starting to confront with growing urgency: when most of what we see online can be generated automatically, are we still participating in a social network… or simply interacting with an ecosystem of algorithms talking to themselves?