Wired: the wave of AI sites isn’t poisoning the internet — it’s making it falsely happy

Researchers have for the first time systematically measured what exactly the wave of AI-generated websites does to the internet as an information environment — and discovered something completely counterintuitive. AI content does not so much flood the network with disinformation or toxicity as it makes it artificially happy. The term "AI slop" — literally "AI garbage" — has long been established in professional communities as a designation for the stream of low-quality, faceless content that language models generate at industrial scale.

Automated blogs, SEO factories, pseudo-news aggregators without editorial oversight, affiliated product reviews — all of this has already become a familiar part of daily online experience. By the estimates of several analysts, by 2026 a significant share of new content in search results is created in exactly this way. The problem has become so widespread that search engines have begun introducing their own filters to detect such sites.

But what exactly happens to the tone and character of the network under the influence of this stream — until now remained an open question. A new study, reviewed by Wired, attempted for the first time to answer this systematically, examining a large sample of sites with a high proportion of AI-generated content in comparison with traditional editorial platforms. The main conclusion turned out to be unexpected: instead of increased aggression, manipulation, or outright disinformation, researchers detected a steady tonal shift toward monotonous, almost obsessive optimism.

Materials created by language models systematically avoid conflict, criticism, ambiguity, and uncomfortable questions. The result is a distorted, excessively smoothed picture of reality — not factually false, but emotionally dishonest. This phenomenon of "artificial happiness" is explained by several mechanisms.

Language models are trained using human feedback that systematically rewards polite, positive, and conflict-free responses — the model perceives aggressive or anxious tone as an error. Moreover, commercial customers of such content — content farms, affiliated sites, traffic aggregators — are interested in retaining readers, and conflictual material converts worse. At the same time, production costs are plummeting: some sites publish hundreds of materials per day for minimal expenses, making the scale of the problem practically unlimited.

The study also recorded an important pattern: automated sites do not displace major editorial outlets directly, but rather fill information niches that quality journalism abandoned long ago. Narrowly specialized topics, local news from small towns, reviews of products from niche categories — it is precisely here that AI content has taken a dominant position. And it is precisely here that "false positivity" is felt most acutely: materials sound like advertising press releases, not an independent perspective.

The authors emphasize that the surprising aspect of the results is not that AI content is bad, but rather how exactly it is bad. Public discussion has focused on disinformation, political bias, and direct manipulation. The real threat turned out to be far more subtle: gradual hollowing out of critical perspective, emotional monotony, and declining ability of the audience to distinguish nuances.

The internet is not becoming more dishonest — it is becoming kinder and duller than the real world. For the media industry, this reframes the competitive challenge. An original voice, willingness to ask uncomfortable questions, editorial conflict, and critical analysis — what was previously considered simply professional values — now becomes a direct competitive advantage against the AI stream, which by its nature cannot be inconvenient.