Summary
- In June 2026, Anthropic published “When AI Builds Itself,” revealing that Claude already writes more than 80% of the code merged into its own production codebase, a direct signal pointing toward recursive self-improvement.
- OpenAI is targeting full automation of AI research by March 2028, while more than 1,100 employees from OpenAI, Anthropic, Google and Meta are calling for an international slowdown mechanism in the July 2026 letter “Pacing the Frontier.”
- A loop that feeds on its own output loses diversity: a study published in Nature in 2024 describes model collapse, and a 2026 study estimates that around 35% of new websites created in mid-2025 are AI-generated or AI-assisted.
- Online content moderation has demonstrated for several years what the warnings from AI giants now confirm: combining automation with human expertise produces more reliable results than AI left to make decisions on its own.
- These warnings divide experts: some see a genuine emergency, while others suspect a strategy aimed at securing coordination rules that would favor incumbent American players.
Since spring 2026, the heads of the largest artificial intelligence labs have been issuing, one after another, warnings that would have seemed unthinkable from them just three years earlier. Dario Amodei, co-founder of Anthropic, set things in motion on September 12, 2026, with an essay calling for a slowdown in the development of frontier model capabilities. Sam Altman responded two days later on X, stating that the industry “could lose control of the future to AI” and that such a scenario remains “unacceptable.” These statements are rooted in specific facts: a report published by Anthropic in June 2026, a letter signed by more than 1,100 employees of the leading labs, and a documented security incident involving an OpenAI model.
This new context rests on a specific phenomenon, recursive self-improvement: an AI that takes part in designing its own successors.
AI that builds AI
Anthropic’s “When AI Builds Itself” report and the 80% self-generated code threshold
On June 4, 2026, Anthropic published a report titled “When AI Builds Itself,” in which the company states that Claude now writes more than 80% of the code merged into its own production codebase, up from a marginal share before the launch of Claude Code in early 2025. Anthropic’s engineers now ship roughly eight times more code per quarter than they did a few years ago, an acceleration the company directly attributes to using its own models to develop the next ones. The report sets out a simple line of reasoning: taken to its conclusion, and given enough computing power, this dynamic logically leads to a system capable of fully designing and developing its successor, without any decisive human involvement. Marina Favaro, from Anthropic’s research institute, sums up the ambivalence of this finding, explaining that the dynamic carries the potential for considerable benefits to the world, while also creating real risks of loss of control. The finding carries particular weight coming from a company that built its reputation on AI safety, and that is documenting here, with its own internal metrics, the very trajectory it is simultaneously recommending be slowed.
OpenAI and the openly declared race toward fully automated AI research
While Anthropic documents its own acceleration, OpenAI takes a different stance on the same ground: that of an openly embraced strategic goal rather than a risk to be contained. Jakub Pachocki, OpenAI’s chief scientist, describes the automation of AI research as the company’s “most important goal,” with an explicit roadmap: a “virtual intern” capable of carrying out research tasks autonomously by September 2026, followed by full automation of AI research targeted for March 2028. This difference in messaging between the two labs illustrates a defining tension in the sector, documented by several specialist media outlets: the companies publicly warning about the risks of self-improvement are, at the same time, investing heavily to be the first to achieve it. Daniel Kokotajlo, a former OpenAI researcher now involved in forecasting work on AI trajectories, notes that some observers even consider his own timelines too conservative, a sign that the debate over timing is far from settled among the experts themselves.
The “Pacing the Frontier” letter: when 1,100 employees call for a collective brake
In July 2026, more than 1,100 employees of OpenAI, Anthropic, Google and Meta signed an internal letter, later made public, titled “Pacing the Frontier.” The text asks the US government to support an international effort to build the technical and governance tools needed to deliberately slow the frontier of automated AI development, should system capabilities advance faster than society’s ability to oversee them. It is a request for coordination infrastructure, not a call for an immediate halt, an important nuance that several media outlets initially glossed over. Notably, the companies employing the signatories endorsed the initiative within a day of its publication, setting it apart from open letters by independent researchers in the past. Critics, quoted in the specialist press, nonetheless point out that such a US-led mechanism would mechanically favor incumbent American players, while having no leverage over models developed outside US jurisdiction.
Beyond the sensational “loss of control,” a very real loss of diversity
Noah Giansiracusa, a researcher at Bentley University, considers Anthropic’s messaging to be more of a communications strategy than a genuine call for a slowdown, with the company actually seeking to keep developing at full speed. Mark Riedl of Georgia Tech, for his part, speaks of media hype fueled more by a desire to draw regulatory attention toward competitors than to document a real risk. These criticisms take on particular significance given the financial context of the period: Anthropic’s warnings came shortly after a confidential IPO filing and a funding round valuing the company at nearly $965 billion.
These spectacular announcements call for a critical eye: an alarming statement can reflect a real risk just as easily as it can serve a corporate strategy, and every claim is best checked against independent sources. It is also worth noting that the labs setting the pace for the most advanced models today are American. Europe is speaking up, but it is playing catch-up. Ursula von der Leyen states that European rules put the Union in a position to shape global efforts, while Mistral argues that some established players are using the moment to consolidate their position, Clément Delangue (Hugging Face) calls for accelerating, and French Finance Minister Roland Lescure sees the leaders’ self-interest at work.
Whatever the case, one thing is certain: an AI that learns from text written by other AIs becomes impoverished. Researchers led by Ilia Shumailov demonstrated this in Nature on July 24, 2024. They trained AI models in a chain: each generation learned exclusively from text produced by the previous one. The models “forgot improbable events”: rare cases disappeared first, resulting in increasingly uniform output.
In recent months, the web has filled up with generated text. A study by researchers from Imperial College London, the Internet Archive and Stanford analyzed websites created between August 2022 and May 2025: around 35% of new websites from mid-2025 are classified as AI-generated or AI-assisted (“The Impact of AI-Generated Text on the Internet,” April 2026). These sites resemble one another more than other sites do (+33%) and adopt a markedly more positive tone (+107%).
Faced with an increasingly uniform web, the researchers behind Ilia Shumailov’s study concluded that data drawn from genuine human interactions will become more valuable as AI-generated content proliferates. Human expertise, more attuned to the codes and subtleties of the world it operates in, therefore has a real role to play in this new order of information and communication.
The winning combination of AI that sorts and experts who correct: a focus on moderation
AI: fast and useful, but imperfect
In moderation, AI brings speed but also reveals its limitations. A study published on arXiv in August 2025 compared professional reviewers with multimodal AI models (GPT-4o, Gemini, Llama) on 1,500 videos from TikTok, YouTube and Meta, in 12 languages, to identify content that poses a risk to brands. On a reliability score from 0 to 1 (the F1 score, which combines detecting risky content with avoiding false alarms), humans scored 0.98 and the three AI models between 0.86 and 0.91. The AI errors stemmed from misleading word associations (for example, a “grass dealer” classified as drug-related), weaker performance on non-English text, and ambiguous situations.
Online content moderation offers a textbook case that maps directly onto the debate on AI self-improvement, since for several years it has faced the same question on a smaller scale: how far can automation alone make reliable decisions? Automated moderation systems are generally calibrated to prioritize limiting false negatives, meaning problematic content that slips through the net, which mechanically increases false positives, meaning perfectly legitimate content wrongly removed. Jean Vanderdonckt, an artificial intelligence researcher at UCLouvain, sums up the structural limitation of these systems, pointing out that AI reliability is inherently imperfect, with algorithms struggling in particular to grasp the subtleties of language: irony, cultural context, fast-evolving slang, or deliberate attempts at circumvention. This brings us back to the idea of a lack of diversity when AI alone is in control.
Toward hybrid governance
Organizations that handle large volumes of online content and interactions, whether brands, platforms or communities, therefore have a clear interest in building hybrid governance now. According to current knowledge documented by METR, the EU AI Act and lessons learned from content moderation, human oversight remains the only proven mechanism for detecting drift, correcting bias and keeping decisions explainable.
Netino supports brands and organizations in managing their content, moderation and customer interactions at scale, placing this very alliance between artificial intelligence and human expertise at the heart of its operating model.
As a Marketing Process Outsourcer, Netino combines AI technologies to handle volume with the rigor of human experts trained to arbitrate ambiguous cases, ensure regulatory compliance and protect its clients’ reputation, across five pillars: governance, operations, technology, performance and resilience. This hybrid approach, already proven across hundreds of French and international brand accounts, is a concrete illustration of what the warnings from AI giants are calling for: powerful AI serving informed human decisions rather than replacing them.
To find out how to build this hybrid governance for your own content and interactions, the Netino team is available at netino.com/fr.
Sources
- CNBC, “OpenAI boss Sam Altman spells out how and why the AI industry wants to slow down,” September 14, 2026
- Yahoo Finance, “OpenAI CEO Sam Altman on AI: We could lose control”
- Scientific American, “Anthropic warns AI may soon begin recursive self-improvement”
- Tom’s Hardware, “Anthropic’s warning over AI self-improvement has a hidden message”
- AI Magazine, “Anthropic Warns Autonomous AI Risks Loss of Human Control”
- Time, “What Happens When AI Starts Building AI? Inside Recursive Self-Improvement,” August 7, 2026
- TechTimes, “Over 1,100 AI Employees Petition for US-Backed Pacing Mechanism After OpenAI’s Sandbox Escape,” July 28, 2026
- METR, “Measuring AI Ability to Complete Long Software Tasks,” March 2025
- Shumailov et al., “AI models collapse when trained on recursively generated data,” Nature, July 24, 2024
- “A Note on Shumailov et al. (2024),” arXiv
- Dolezal, Alam, Graham, Bohacek, “The Impact of AI-Generated Text on the Internet,” April 2026
- “AI vs. Human Moderators: A Comparative Evaluation of Multimodal LLMs in Content Moderation for Brand Safety,” arXiv, August 2025
- RTBF, “Modération en ligne : l’intelligence artificielle en renfort, mais l’intervention humaine reste primordiale”
- Netino
