AI: Giants Pay Publishers, Fear a "Doom Loop"

AI-generated content is flooding the web and threatening the sources on which models depend. Internal notes from Microsoft refer to a "doom loop," while studies describe "model collapse" when AI trains on AI. Google and Microsoft are testing compensation schemes for publishers to secure human data.

Google and Microsoft Launch Payments for Content Used by Their AIs

In mid-September 2026, Google began compensating certain publishers whose content is used in responses from Gemini, AI Overviews, and AI Mode. This initiative, called the "AI contribution pilot," operates by invitation and involves nearly a hundred publishers, with compensation determined by the proportion of their content present in each response. As early as February 2026, Microsoft had established the Publisher Content Marketplace, a platform allowing publishers to set their terms and receive compensation based on the use of their content by AIs.

The Bet on Synthetic Data Proposed by Sam Altman in 2023

In 2023, Sam Altman, co-founder of OpenAI, discussed the possibility that AIs could one day generate their own high-quality training data. According to him, if a model reaches a level where it produces good synthetic data, that could be sufficient to continue progress without relying on human content.

Pressure on Publishers: Declining Clicks and Rising Automated Production

Google has long claimed that the introduction of AI Overviews and AI Mode would not negatively impact site traffic. However, publishers are noticing a decline, particularly for evergreen content like guides and comparisons. By the end of July, Bérangère Farcot, a journalist at Le Temps, reported that the click-through rate on these formats had been halved and noted that the situation is becoming more complicated for media specializing in practical information when the Overview provides a deemed sufficient answer. In response to this trend, newsrooms are altering their organization: by the end of March 2026, Prisma Media announced the elimination of 261 positions, representing nearly 40% of its workforce, while training 400 journalists in AI and already using this technology for part of its content. Voici generates about 20% of its online content using AI from news wires, and according to a Next survey, AI could account for up to 40% of the group's articles.

Documented Technical Risk: Towards a "Model Collapse" in Case of Recursive Training

In 2024, a study published in Nature by researchers from Oxford, Cambridge, Toronto, and Imperial College showed that training models on content generated by other AIs leads to irreversible flaws, a phenomenon called "model collapse." Rare information disappears first, and generation after generation, the model diverges from reality. In one experiment, after nine iterations, a model that started with a text on medieval architecture produced a list of hares with tails of different colors. Researchers observe that models gradually generate more probable sequences derived from the original data while introducing their own errors. They estimate that this phenomenon could become widespread across the web, making it more difficult to train new models without access to human data or corpora predating the rise of synthetic content.

Internal Recognition of a "Doom Loop" and Legal Tensions

The issue is also present on the legal front. At the end of 2023, the New York Times sued OpenAI and Microsoft in New York for the alleged unauthorized use of its articles, an action joined by other publishers. By mid-September 2026, a 92-page memorandum was filed to seek a ruling in their favor, citing internal notes where OpenAI and Microsoft acknowledge the impact of their technologies on the web. Microsoft indicated in 2024 that its products create a vicious cycle and threaten the economic foundations of their essential suppliers, describing the aspiration for content to train AIs as "the greatest theft of labor in human history." Nick Turley, head of ChatGPT, speaks of an existential threat to publishers, as AIs are largely substitutive to news sites. Microsoft refers to a "doom loop" that could harm the performance of its models and the web. Furthermore, language models have absorbed billions of pages written by humans, and services like ChatGPT, Perplexity, and AI Overviews are gradually replacing sites to respond to users.