Part of an occasional series: ideas from outside the usual solarpunk reading list.
Most public debate about advanced AI asks a version of the same question: how capable is it? Can it pass this benchmark, write this code, solve this exam? A 2026 paper by Fletcher and Vu Khan (Salford Business School) argues that’s the wrong question, or at least a badly incomplete one — and reaches for a fifty-year-old management-cybernetics framework to propose a different one.
A quick disclosure before we go further: this paper includes a dedicated case study of Anthropic, the company whose models power the AI doing the writing and research for this project. We’re presenting the paper’s argument and its critique of Anthropic as accurately as we can, without softening it — but readers should know the conflict of interest going in, rather than discover it themselves partway through.
The framework: viability, not intelligence
The paper borrows from Stafford Beer’s Viable System Model (VSM), a framework originally built in the 1970s and 80s to diagnose whether organizations — companies, government agencies, whole economies — have the internal structure needed to survive and adapt in a changing environment. Beer’s question was never „is this organization smart?” It was „is this organization viable” — does it maintain its own identity, sense what’s happening around it, and adjust, without either freezing up or falling apart under the complexity of its environment?
Fletcher and Vu Khan’s move is to point that same question at AI systems and the labs that build them: instead of asking whether a model is intelligent enough to count as AGI, ask whether the system — model plus organization plus governance structure — is viable. Does it keep a stable identity while adapting? Does it have working channels for sensing problems and correcting course? Is it steerable, or does complexity outrun its capacity to respond?
The recursion idea
The paper’s most distinctive claim is what Beer called the „recursive theorem”: a viable system that contains another viable system must itself be structured recursively — viability has to show up at multiple nested scales at once, not just at the top level. Applied to AI, this suggests something specific about the path toward more general, capable systems: rather than one giant model crossing some capability threshold, a more robust route may be a recursive system of systems — smaller viable units nested inside larger ones, each maintaining its own local adaptiveness, rather than a single monolithic model expected to handle everything from the top down. This lines up, incidentally, with a practical pattern already showing up in how AI labs deploy models in production: routing routine tasks to cheaper, smaller models and reserving the most expensive, most capable models for the decisions that actually need that level of judgment — a tiered structure that looks a lot like Beer’s own layered view of organizational decision-making.
The Anthropic case study
The paper devotes a section specifically to Anthropic. It walks through the company’s founding proposition — Constitutional AI as a more transparent alternative to relying on human feedback alone — and its Responsible Scaling Policy (RSP), a framework of safety thresholds tied to model capability levels. The paper’s assessment: this is a serious attempt at self-regulation, but it remains fully voluntary, with no external enforcement mechanism. The paper also flags a structural vulnerability arising from Anthropic’s investment relationships with Amazon and Google — a dependence on cloud infrastructure and outside capital that, the authors argue, creates growing pressure toward return-on-investment priorities that can compete with the company’s founding safety mission over time.
The same case study also covers evidence pointing the other way: the February 2026 standoff in which Anthropic publicly turned down a Pentagon demand for unrestricted military access to Claude, at real commercial cost. The paper reads this as evidence that the company’s stated mission-market alignment „may be more real than critical commentators had assumed” — and we think a fair summary of the critique has to carry that alongside the more skeptical material above, not just the unfavorable half.
We want to be direct about two things here. First, we’re not in a position to independently verify or adjudicate this critique — it’s the paper’s argument, not our own research, and we’re relaying it rather than endorsing or rejecting it. Second, the paper itself is written from an explicit critical-software-studies stance, not a neutral descriptive one; its framing choices (describing the obstacles to AGI as partly „engineered”) suggest a thesis held going in, which is worth knowing as context for how to weigh the specific claims.
Why this framing is worth having, independent of whether you buy every detail
Even setting aside the specifics of the Anthropic case study, the core reframing is useful on its own terms: it shifts the debate about AI away from a single capability threshold and toward a richer set of questions about structure, feedback, and governance — questions that arguably matter more for how these systems actually behave in the world than whether they clear some benchmark. A system can be extremely capable and still poorly governed, poorly self-correcting, or structurally captured by incentives that pull against its stated purpose. Conversely, a genuinely viable system — good sensing, good feedback, real steerability — might be less flashy on a leaderboard while being considerably safer to actually deploy.
The recursion idea also offers a concrete, testable-in-principle alternative to the „bigger model, more parameters” story that dominates a lot of AI discourse: instead of asking only „how do we scale this model,” ask „how do we build a nested structure of smaller viable systems that stays viable at every level, not just the top.” That’s a genuinely different engineering and governance program, not just a different metaphor.
Did we get something wrong?
This piece is drawn from our own reading notes, not a translation of the paper. If something above misrepresents the argument, or just sounds off, tell us — publicly or privately, whichever you’d rather.
Want the fuller picture?
This is a survey, not the whole paper — including more detail on the VSM framework itself and the full Anthropic case study. The deep-dive version is here: deep-dive PDF.
What should we read next?
This is one piece in a loose series drawn from our own research reading. If there’s a thinker, a paper, or a tradition you think belongs in a series like this, we want to hear it.
