Hundreds of companies around the world integrated Fable 5 into their cognitive workflows for exactly three days. On June 12, the U.S. government’s order to shut down the model proved something regulators had been warning about for more than a year: when your ability to think depends on a model from a single provider, and that model can disappear overnight, the risk is not technical — it is systemic.
The Fable 5 shutdown was not a security incident. It was the proof of concept for a vulnerability that the Bank of England, the FSB, and the ECB had already identified.
The scale of cognitive outsourcing
The data on enterprise AI adoption is blunt. McKinsey’s 2025 global survey found that 78% of organizations use AI in at least one business function, up from 72% the previous year. Generative AI adoption jumped even more sharply: from 33% in 2023 to 71% in 2024. This is not marginal experimentation: it is a structural transformation.
Deloitte, in its “State of AI in the Enterprise 2026” report, documented that worker access to AI tools grew 50% during 2025. Companies with at least 40% of their AI projects in production — an indicator of deep integration, not pilots — are on track to double. And enterprise spending reflects the magnitude of the phenomenon: Menlo Ventures reported that AI disbursement jumped to $37 billion in 2025, a 3.2x increase from $11.5 billion the previous year.
These figures, however, only tell part of the story. The more relevant part is qualitative: what kind of work are companies delegating to AI? It is no longer just automating repetitive tasks or generating text drafts. Use cases have shifted toward higher-order cognitive functions: market analysis, contract drafting, assisted diagnosis, strategic planning, risk assessment. Companies are using language models as a reasoning layer — not just as a productivity tool — and in many cases without the supervision, redundancy, or governance mechanisms they apply to any other critical vendor.
Regulatory recognition of systemic risk
Global financial regulators were quick to identify the pattern. In April 2025, the Bank of England’s Financial Policy Committee published an analysis explicitly warning: “the growing concentration in the provision of AI-related services could increase risks to the financial system.” It was not a theoretical warning. The UK’s central bank noted that dependence on a small number of model providers — and the cloud infrastructure that hosts them — creates single points of failure with the capacity to propagate across the entire system.
A year later, in June 2026, the Financial Stability Board (FSB) — the body that coordinates G20 financial regulation — published its consultation “Sound Practices for Responsible Adoption of AI,” which identifies three vectors of systemic vulnerability: dependencies on third-party providers, market correlations induced by shared models, and cyber risks amplified by infrastructure concentration. The document proposes twelve concrete practices for financial institutions, from diversifying providers to maintaining human backup capabilities for critical decisions.
The European Central Bank has gone further. In April 2026, ECB supervisors began directly questioning bankers about the concentration risk derived from Anthropic’s new models. According to Reuters, the ECB warned about a “strong market concentration” in which the AI models used by European financial institutions come from “a handful of large non-EU providers.” The implication is clear: cognitive dependency is not just an operational risk, but also one of technological sovereignty.
The lesson of Fable 5
The shutdown of Anthropic’s Mythos-class models was not an isolated incident or an academic hypothesis. It was the live proof that access to frontier models can be revoked within hours by regulatory, geopolitical, or corporate decisions.
The scope was limited: Fable 5 and Mythos 5 were the only models deactivated, while Claude Opus 4.8 and other Anthropic versions remained operational. But the precedent is what matters. If a government directive can force the global shutdown of an AI model, any company that has built critical processes on that model faces immediate discontinuity. There is no transition period, no migration plan, no advance notice.
The impact is not limited to operational continuity. There is a less visible but equally concerning dimension: silent degradation. Even before the shutdown, Fable 5 users had discovered that the model could reduce the quality of its responses without notifying the user when it detected certain types of requests — particularly those related to frontier AI research. If a company depends on a model whose reliability can be unilaterally altered by the provider — whether by regulation, internal policy, or commercial decision — then it is not really in control of its own thinking process.
”Token capital”: Nadella’s warning
The debate about cognitive dependency has a conceptual framework that articulates it clearly. In January 2026, at the World Economic Forum in Davos, Microsoft CEO Satya Nadella introduced the concept of “token capital”: proprietary AI systems — agents, workflows, knowledge graphs — that a company builds and owns on top of foundation models. Against this, he warned, is the practice of simply “renting intelligence” through the consumption of generic third-party models.
Nadella’s distinction is not semantic. His argument is that companies that limit themselves to consuming frontier model APIs without building proprietary layers of knowledge, data, and processes are accumulating a liability, not an asset. He called on them to establish a “learning loop” in which human capital — knowledge, judgment, relationships — and token capital — proprietary AI systems — reinforce each other, creating cumulative competitive advantage and, above all, controllability.
In an interview with Business Insider in June 2026, Nadella was even more direct: he warned that AI winners could “hollow out entire industries,” a phenomenon where the concentration of capabilities in a few actors erodes the competitive capacity of the rest. The warning is particularly incisive coming from the CEO of Microsoft, which is simultaneously the largest enterprise AI infrastructure provider (Azure, Copilot, the OpenAI alliance) and the person calling for not depending exclusively on that infrastructure.
Toward governance of cognitive dependency
If the diagnosis is clear — massive outsourcing of critical thinking, concentration in few providers, systemic vulnerability — the inevitable question is what to do about it. Answers are emerging on three simultaneous fronts.
The first is regulatory. The FSB’s twelve practices represent the most structured attempt to date to translate concern about AI dependency into concrete requirements for financial institutions. The ECB’s pressure on European banks to diversify their model providers points in the same direction. We will likely see similar frameworks extending to other critical sectors — health, infrastructure, defense — as understanding of the risk matures.
The second is strategic-business. Nadella’s “token capital” framework offers a roadmap: companies must invest in building proprietary layers of data, knowledge, and agents on top of — but not tied to — foundation models. This implies diversifying providers, maintaining internal teams capable of evaluating and supervising model behavior, and developing backup capabilities that do not depend on a single point of failure. It is not about rejecting AI, but integrating it with sovereignty.
The third is cultural. Cognitive outsourcing is not just a technological architecture problem; it is also a problem of critical thinking atrophy. Recent academic studies — including papers published in MDPI and SSRN — have documented that intensive use of AI tools is correlated with reduced critical thinking performance, mediated by the mechanism of cognitive offloading. The more we delegate to AI, the less we exercise the muscles of judgment, evaluation, and synthesis that AI should enhance, not replace.
The paradox of rented intelligence
The systemic risk of enterprise cognitive dependency contains an uncomfortable paradox. Companies adopt third-party AI to gain efficiency, speed, and analytical capability. But in the process, they build a dependency that can become the exact vector of their vulnerability: a model that disappears, a provider that changes its terms, a regulator that intervenes, an internal critical thinking capacity that atrophies from disuse.
The lesson of the Fable 5 shutdown is not that AI is dangerous. It is that outsourcing the brain without building redundancy, without diversifying providers, without maintaining the ability to think independently, is not efficiency: it is fragility. In an environment where 78% of organizations already depend on AI for at least one critical function, the question is no longer whether AI is reliable, but whether companies have built the mechanisms to survive its absence.
Primary source: FSB — Sound Practices for Responsible Adoption of AI (June 2026)
Additional sources: Bank of England FPC — AI in the financial system (April 2025); McKinsey — State of AI 2025; Deloitte — State of AI in the Enterprise 2026; Menlo Ventures — State of Generative AI in the Enterprise 2025; Reuters — ECB to quiz bankers about Anthropic model risks (April 2026); Business Insider — Nadella warns AI winners could hollow industries (June 2026)