Following the Fable 5 outage, which demonstrated Washington's ability to instantly halt the most powerful artificial intelligence models, the industry has begun formulating a plan to prevent such a chaotic scenario from recurring. Critics note that this plan may only replace direct Washington control with a more refined governance mechanism.
Google DeepMind co-founder and CEO Demis Hassabis published a manifesto on the X platform titled 'Structure for Advanced AI and the Dawn of a New Era.' In it, he called for the creation of a new body that would vet the most advanced AI models before their release. This call received support from leaders at Microsoft, OpenAI, and xAI within a single day, an unusual display of unity in an industry accustomed to fierce competition.
The core of the proposed plan is the creation of a 'standards body,' modeled after the Financial Industry Regulatory Authority (FINRA). This private organization, funded by the industry itself, oversees Wall Street under the purview of the US government. This body, supported by Washington but financed by AI labs, would test advanced models for the most dangerous capabilities—cyber, biological, and 'deceptive' or agentic risks. The review would be conducted by a panel of technical experts with a majority of independent members, including representatives from the industry, government, and open source.
Hassabis wrote that initially, advanced labs would voluntarily provide models to the standards body for review thirty days prior to release. He added that after demonstrating the effectiveness and reliability of the assessment protocol, formalization could quickly follow, requiring advanced models to undergo this vetting for market entry in the US.
These rules should apply to any advanced class models, regardless of their country of origin or whether they are open or closed, with the exception of startups and academic institutions. Hassabis believes that the 'advanced' status itself will serve as an incentive to participate, stating to Axios that passing the testing is a 'quite pleasant prestige asset.'
Hassabis, who consulted for the Trump administration, other lab leaders, and European officials for months before the public announcement, told Axios that he wants this body to begin operations 'by the end of the year,' as he believes general artificial intelligence 'will likely be available in a few short years.' This idea builds upon proposals made by AI leaders to world leaders at the G7 summit in Evian in June.
Microsoft CEO Satya Nadella called the initiative 'an important contribution from Demis,' emphasizing that the goal is to create an ecosystem that 'fosters innovation and choice while avoiding any single failure that could destroy the world.' Sam Altman of OpenAI also expressed support, having previously used a Financial Times article to call for the creation of a US-led international forum based on similar principles.
Even Elon Musk, whose xAI is the only major advanced lab outside the Frontier Model Forum safety group, offered measured support, writing on X on Wednesday: 'Overall, this is a well-thought-out structure and certainly a good starting point for discussions.' This differs from his June tone when he jokingly suggested the name 'AIAIAI' when discussing an AI regulator.
Dario Amodei of Anthropic agrees with the direction but not the design. In a June essay, he advocated for a federal agency in the style of the FAA with legislative authority to block unsafe models. As Axios noted, lab leaders now agree that Washington should regulate them, 'differing mainly on who holds the power.'
The most obvious gap in Hassabis's proposal concerns China. Although Hassabis insists on applying his standards to all advanced models regardless of origin, he does not specify how a US-based body can compel Chinese developers—such as DeepSeek and Zhipu AI, which are approaching advanced capabilities—to submit their models for testing.
Leading labs in Beijing adhere to the opposite strategy: Tan Ze of Zhipu argues that safety is achieved through broad access, not technological barriers, and released its latest model with open source weights. Once the model weights are published, no body can recall or suspend it.
Another open question is whether this proposal could have prevented the crisis that inspired it. In June, US Trade Minister Howard Lutnick imposed export controls on Claude Fable 5 and Mythos 5 a few days after their launch, applying a legal regime designed for physical goods to a cloud service. This occurred after Amazon researchers hacked Fable 5's defenses simply by asking the model to read the codebase and fix its flaws, after which it discovered vulnerabilities and generated exploit code in one case.
Anthropic disputed the severity of the situation, showing that less capable competing models could have done the same, but it had to shut down both models globally overnight, restoring access only two weeks after negotiations that, according to it, lacked any rules, protocol, or guidance.
Under the Hassabis regime, Fable 5's capabilities would have been tested against the published protocol before release, and post-release flaws would have been addressed through a managed patch process, rather than a directive dropped on Friday evening. The promise is predictability. However, this does not eliminate Washington's fundamental powers of export control—a decisive administration could still act outside the protocol, as happened in June—and history shows that voluntary regimes have their limits.
Tom Wheeler, a Brookings fellow and former Chairman of the Federal Communications Commission, noted that similar international standards bodies, such as the Financial Action Task Force and the Basel Committee on Banking Supervision, functioned because non-compliance carried real sanctions. He stated: 'Voluntary compliance means the outcome is merely a suggestion, not a standard.'
He also warned that a process limited to the companies themselves 'solves company problems, giving governments cover to appear to be doing something.' Furthermore, it is unclear what would prevent this body from becoming a tool of the Trump administration. The design provides for FINRA-style independence, but the body would still, according to Axios, 'report to the US government,' and the White House has been reserved about formal regulation; Shriram Krishnanan, his senior AI advisor until late June, told the Financial Times: 'there will be no AI FDA,' referencing the U.S. Food and Drug Administration.
The R Street Institute warned after the shutdown that export controls have become a 'permanent policy tool' that can be applied 'against any laboratory, model, or company that becomes inconvenient.' Adding a standards body does not negate this. For the developing world, access to advanced technologies will continue to be granted, not guaranteed, by a gatekeeper in Washington. Nothing in this structure gives South Africa or any other African government a voice in the body that will decide which models reach their markets; the panel represented by Hassabis consists of industry, US government, and open-source community representatives. For countries consuming advanced AI but not creating it, smoother gates remain gates.
Hassabis aims to launch this body within months, but issues of enforcement, the composition of the council beyond its funding industry, and China's participation remain unresolved. Until then, the lesson of June remains relevant: whoever controls the deployment gates controls the world's access to advanced AI, and those gates remain in Washington.
Artificial intelligence (AI) tools have demonstrated great potential to increase productivity in various areas. However, recent research suggests that indiscriminate use of these technologies can lead to the weakening of individuals' cognitive abilities.
Studies analyzed indicate that excessive dependence on these systems can diminish users' aptitude for problem-solving without technological support. This warning arose from investigations conducted with students and professionals, where participants who used AI showed immediate improvements in performance but faced difficulties when performing the same tasks autonomously.
Experiments with workers and students confirmed that AI can optimize results when applied to activities aligned with their competencies. In a study involving hundreds of consultants from the Boston Consulting Group, researchers from the Wharton School noted an increase in the number of completed tasks and a decrease in time spent by those who had access to the tool.
This survey, published in 2026 in the journal Organization Science, revealed that employees assisted by AI produced higher quality work in functions where the technology was most competent. The most significant progress was observed among professionals who had a more modest initial performance.
A similar finding was found in a study conducted by Grace Liu of Carnegie Mellon University, focusing on mathematical problem-solving. By comparing students with and without access to AI, the study identified superior performance among those who could use the resource during exercises.
Despite the instant gains, researchers detected adverse effects after the removal of technological assistance. Individuals accustomed to AI began to show lower performance than those who had never used the resource and demonstrated less resilience in the face of difficulties.
Additionally, another study investigated how excessive trust in responses generated by AI systems affects decision-making. Steven Shaw and Gideon Nave evaluated over 1,300 participants and identified a phenomenon called 'cognitive surrender,' characterized by the user abandoning their own evaluation in favor of the conclusion provided by the machine.
Researchers argue that AI can operate as a third cognitive mechanism, complementing the traditional methods of fast thinking and detailed analysis described by Daniel Kahneman. The problem arises when this tool ceases to be a complement to human reasoning and begins to replace it entirely.
Experts emphasize that the crucial skill in the age of AI will be discerning which tasks should remain under human control and which can be delegated to automated systems. Collaboration is most effective when the user understands the limits of the technology and can judge its outputs.
An analysis from 2024, published in the journal Nature Human Behavior and based on 106 AI experiments, showed that joint performance between humans and machines is optimized when each party acts in its area of greatest advantage. However, when the system demonstrates superiority, the human difficulty in knowing whether to trust or contest the tool can compromise the final result.
The interviewed experts highlighted that processes such as initial idea conception, text writing, and knowledge generation require active human participation. For them, AI is most valuable in phases of review, questioning arguments, and refining existing work.
The concern extends to the educational environment. The cited studies indicated that students may learn less by using AI solely to speed up school assignments. On the other hand, when the tool is used to gain clarification, formulate questions, and assimilate concepts, the damage to the learning process is minimized.
Judy Hanwen Shen and Alex Tamkin, from Anthropic, reported in their research with developers learning a new programming library conceptual difficulties, code reading, and debugging when AI was used as a shortcut to obtain ready-made solutions. The researchers' recommendation is to reconfigure AI as a tool for deepening understanding, stimulating questioning and expanding analytical capacity, rather than simply eliminating the mental effort required to learn.
