Anthropic Calls for Global AI Slowdown: Recursive Self-Improvement Risks Are Imminent

On June 5, something unprecedented happened in AI safety: a top-tier lab publicly called on its peers to slow down.

In a lengthy blog post, Anthropic warned that the pace of frontier AI progress may soon enable systems to achieve recursive self-improvement without human intervention, creating significant societal risks. The post proposed establishing a global moratorium agreement and verification mechanism, stating bluntly: “This moment may arrive sooner than most institutions are prepared for.”

Core Warning: Recursive Self-Improvement

Anthropic’s central argument:

Current frontier AI models are advancing faster than anticipated
Models may soon gain the ability to autonomously improve their own code, architecture, or training pipelines
Once initiated, such “recursive self-improvement” would progress at exponential rather than linear speed
Humans could lose effective oversight over the system’s developmental trajectory

The post recommends:

Establishing a global research moratorium agreement
Setting up independent verification mechanisms to confirm compliance
Deploying safety assessment frameworks before critical capability thresholds emerge

An Awkward Detail: 80% of Anthropic’s Code Is Written by Claude

In the same post, Anthropic disclosed an ironic data point: over 80% of its codebase is written by Claude.

This figure itself validates the trend they warn about: AI systems are already deeply involved in creating more advanced AI systems. Anthropic engineers are shifting from “writing code” to “reviewing, correcting, and guiding AI-generated content”—a landscape strikingly similar to the one described by Tencent executives who noted that “most of our code is AI-generated.”

But this also raises a self-referential question: when an AI company warns about AI self-improvement risks while its own product already produces 80% of the company’s code, how should the credibility and consistency of this warning be evaluated?

Industry Debate: Genuine Concern or Regulatory Capture?

Anthropic’s appeal has triggered divided reactions across the industry.

Supporting views:

AI safety researchers have long warned that recursive self-improvement is one of the most dangerous inflection points on the path to AGI
If self-improvement capabilities do emerge, humans would have virtually no time window to establish effective control mechanisms
Establishing preventive frameworks before risks materialize is far less costly than remediation after the fact
Anthropic’s willingness to propose constraints as a stakeholder demonstrates moral courage

Skeptical views:

Anthropic’s current model capabilities lag behind OpenAI and Google; slowing development would harm competitors more than themselves
Historical lessons of “regulatory capture”: incumbents often push rules that favor their own position under the guise of safety
A global research moratorium is practically impossible to implement and verify at the technical level
The fact that 80% of Anthropic’s code is written by Claude proves they are among the most AI-dependent companies in existence

One unnamed AI lab executive commented: “If this is a genuine appeal, Anthropic should lead by publishing complete training details and safety evaluation results for its most advanced models. Otherwise, this reads more like a strategic statement from a player losing the capability race.”

The Deeper Question: Who Defines “Safety”?

Anthropic’s appeal exposes a central tension: the battle for AI safety discourse dominance.

Over the past two years, the AI safety field has fractured into multiple factions:

Faction	Position	Representatives
Effective Accelerationism	Push AI capabilities faster; society will naturally adapt	Some former OpenAI employees, Marc Andreessen
Safety-First	Capability development must yield to safety verification	Anthropic, some DeepMind researchers
Governance Moderates	Balance development with stronger government oversight	EU AI Act supporters
Capability Optimists	Self-improvement is a path to welfare; risks are manageable	Some Chinese AI labs

Anthropic’s appeal is essentially an attempt to elevate “safety-first” to global consensus status. But the initiative faces hard realities:

Geopolitical competition: US-China AI competition has entered the national strategy level; unilateral slowing by either side would be seen as strategic disadvantage
Commercial conflicts: AI lab valuations and fundraising capacity are directly tied to technical leadership
Verification dilemma: How to independently verify whether a lab is “truly slowing down” rather than merely publicly claiming to

Recursive Self-Improvement: From Theory to Threshold

Recursive self-improvement (RSI) is not an Anthropic invention. It first appeared in I.J. Good’s 1965 prophecy: a superintelligent machine could design even better machines, triggering an “intelligence explosion.”

But whether current AI systems are approaching this threshold is a matter of serious academic disagreement:

Optimists: Current LLM code generation capabilities are improving, but remain far from true autonomous architectural improvement. 80% AI-written code does not equal AI that can design better AI.
Cautious: Progress curves are non-linear. If the 2023-2026 capability growth rate continues, autonomous improvement could emerge in 2027-2028. Anthropic’s warning is issued while the time window remains open.
Skeptics: Recursive self-improvement may be another “moving goalpost” in AGI narratives—whenever a milestone is approached, the definition is shifted to something harder to achieve.

Core Assessment

Anthropic’s appeal is a complex signal that cannot be simply categorized as “genuine” or “strategic.”

The more probable reality: it contains both.

Safety concerns are real: Recursive self-improvement is indeed the most uncertain link on the AGI path, and establishing assessment frameworks in advance is reasonable risk management
Competitive motives exist: Labs in a追赶 position in the capability race have structural incentives to push issues that favor their own pace
Implementation probability is extremely low: A global research moratorium is practically impossible under geopolitical and commercial realities
The battle for discourse dominance has begun: The right to define AI safety standards will become one of the core battlegrounds of industry competition in the next phase

The real significance of Anthropic’s post lies not in whether it can successfully push for global slowdown, but in its formal elevation of “recursive self-improvement” from an academic discussion topic to an agenda that the industry must publicly address.

Whether supporting or opposing, other major labs must now take a position on this issue. Silence itself is a position.

Sources

Phoenix Network Technology, June 5, 2026
Anthropic official blog, June 5, 2026
Science and Technology Innovation Board Daily, June 6, 2026
Tencent Technology, June 5, 2026