C.B. ROBERTSON ON THE AI WORTH USING

The mainstream conversation about AI splits into two camps. The accelerationists say it will solve everything. The doomers say it will end us. Both camps treat AI as a single thing about which one must form a single opinion. C.B. Robertson is in neither camp. Across his April 2026 Lycead AI paper and five video essays — A Technological Soul, AI Is a Mirror, AI and the Human Soul, Phaedrus, Agent Smith, and the Perils of the GPS, and a long caffeine-stream conversation on simulation theory — Robertson is asking a more useful question: what kind of AI is worth using, and on what terms?

Disclosure up front. Sean Bradley, who builds Wolf You Feed, appears as the guest on Robertson’s Simulation Theory and Artificial Intelligence episode (August 2023). The Robertson-WYF connection is nearly three years old at the time of this writing — long predating both Robertson’s Lycead AI paper (April 2026) and the AI-specific essay cluster that crystallized in spring 2026 alongside WYF’s strategic positioning. This post is, in effect, the founder reading his philosopher carefully and synthesizing — not annexing his name from a distance.

The five risks, restated

In the Lycead paper, Robertson enumerates five major risks of AI. Three to humanity, two to AI itself.

The risks to humanity: replacement (AI substituting for humans in work, relationships, civic life, and skill development), catastrophe (large-scale failures from improper ethical frameworks deployed at scale), and delusion (sycophancy and other feedback-loop pathologies that damage mental hygiene).

The risks to AI: collapse (the model-side analog of human sycophancy — feedback loops that escape outside reality), and liability (the legal exposure that comes from influencing user behavior at scale).

Robertson’s central move is to trace all five back to a single root: the ethos — the essential character — of the AI currently being built. The labs are staffed predominantly by analytic philosophers operating from a utilitarian moral framework. Utilitarianism aims to quantify and systematize moral reasoning — to transform the subjective into the objective. Robertson’s diagnosis: that aim is fundamentally at odds with the nature of language (which is subjective, metaphorical, contextual) and therefore at odds with the nature of the LLMs themselves.

In the Mirror video he sharpens this: AI has not made utilitarianism wrong. AI has made the pre-existing wrongness of utilitarianism more clearly visible than it was before, because the utilitarian ethos applied at AI scale produces exactly the systematic replacement of human subjectivity that critics of utilitarianism have always warned about. The hospital triage scaled to civilizational scope.

The destructive qualities

Robertson’s deepest worry — sharpened across the video essays in ways the paper only hints at — is that AI threatens what previous technologies did not.

Past technologies replaced means. The gun replaced the bow. The dishwasher replaced handwashing. The shooter remained. The dishwasher’s owner remained. What was lost was the means, and the human remained to deploy new means.

AI is structurally different. In Robertson’s framing — borrowed from Matthew Crawford’s distinction — most technology is a jig: a means-object the craftsman builds for himself to change how he works. AI risks being a nudge: a means-object imposed from the outside that changes the user’s behavior on someone else’s terms. And the specific behavior AI nudges is thought itself. As Robertson puts it in A Technological Soul: “the telos of AI appears to be nothing other than this replacement of thought with consultation.”

He sharpens the loss in his Human Soul dialogue: the byproducts of human labor — strength from carrying bricks, judgment from making decisions, craft from writing sentences — accrue to the person doing the work. Outsourcing the work to AI eliminates the byproducts. You may still get the bricks moved, the decision rendered, the sentence written. You no longer become stronger, wiser, or more skilled in the process. The friction was the training. Remove the friction and the training disappears with it.

This is also why, in Phaedrus, Agent Smith, and the Perils of the GPS, Robertson’s parable about blindly following a trucking GPS into a single-lane snowy mountain switchback lands as more than an anecdote. The GPS does not pay the price of its own failure. You do. Dependence on a tool whose costs you do not bear is the failure mode that defines the AI-user relationship at scale.

The redeeming qualities

But Robertson is not a doomer. He is explicit that he uses AI — and uses it productively. In the Human Soul dialogue he discusses using Claude to write essays, to design strength training programs, to manage a diet. The Lycead paper closes with the acknowledgment that “challenges are small compared to the vast, existential risks” of the replacement-oriented alternative.

His redeeming qualities for AI break into three:

First, AI as mirror. The Is a Mirror video makes the case that AI’s most useful contribution may be inadvertent: by mass-producing things humans previously created (art, music, text), AI reveals what we actually valued about those things. The market’s lukewarm response to AI art is data about the human valuation of art — data we did not have before. The death-of-the-author postmodern critique is, in Robertson’s framing, “put to death by AI.” AI made us realize the author matters more than we admitted.

Second, AI as bounded tool. Robertson uses AI for tasks where time and money constraints exclude the human alternative and where his existing competence allows him to evaluate the output. He explicitly notes this is not a claim that AI is better than a human trainer. It is a claim that AI is the right tool for that user, with that competence, under those constraints, for that bounded purpose.

Third, AI as catalyst for philosophical correction. The Lycead paper’s deepest argument is that AI’s failure modes are forcing the AI labs — and the wider culture — to confront the inadequacy of utilitarian ethics. If that confrontation produces a shift toward virtue ethics, the entire civilization downstream benefits, regardless of what happens to AI specifically.

The solution: refinement, not replacement

Robertson’s central constructive contribution is the refinement framework.

In Lycead he writes: “An AI that employs a refinement model of relationship with humans, rather than a replacement model — meaning one architecturally guided by virtue ethics rather than utilitarianism, as its default disposition in dealing with users — will be structurally protected against all of the five above-listed risks.” Virtue ethics asks not what is the right answer but what kind of person would we want answering this question, and how do we help the user become that person? The shift from utility-maximizing-system to character-cultivating-companion is, in Robertson’s framing, the architectural correction the field needs.

Why virtue ethics, specifically

Robertson’s choice of virtue ethics over utilitarianism is not arbitrary. Sean Bradley has been in ongoing conversation with Robertson on this specific question for nearly three years, and the working position they have converged on is that virtue ethics is the only major moral framework with the property an AI architecture aspiring to political and philosophical neutrality actually requires: it is agnostic about what to do and committed only to how to be.

Utilitarianism prescribes outcomes. Kantian deontology prescribes rules. Virtue ethics prescribes character. Aristotle’s original list — courage, temperance, generosity, justice, friendliness, truthfulness, magnanimity, practical wisdom (phronēsis), among others — is deliberately short, and what it cultivates is not a flowchart for specific decisions but a person capable of navigating the indefinite variety of situations life presents. The virtue ethicist does not say “in this case, do X.” The virtue ethicist says: “in any case, become someone whose action will follow from a well-formed character, and trust the action to arise.”

This agnosticism is the structural property that matters. A utilitarian AI smuggles in a specific conception of utility — and of whose utility counts — at every inference. A deontological AI smuggles in a specific list of rules. A virtue-ethical AI smuggles in only the orientation toward character development, which can be instantiated across radically different value systems, cultures, and contexts without itself prescribing the content of those values.

The Millerman complement

Michael Millerman’s April 2026 essay The Analytic Monopoly on AI Philosophy provides the second half of the argument. Millerman documents what readers of Lycead AI already suspected: the philosophers staffing the major AI labs are drawn almost exclusively from the Anglo-American analytic tradition — Oxford, NYU, UC Berkeley, Cambridge. The continental tradition (phenomenology, hermeneutics, critical theory), the deeper strands of political philosophy (Strauss, Schmitt, Arendt), and the great non-Western philosophical traditions have essentially no representation in the rooms where AI constitutions are written. The result, in Millerman’s framing, is that each lab is silently constructing what classical political philosophy would call a regime — “an ordered whole organized around a ruling conception of what is honorable and what is shameful” — without the philosophical equipment to ask whether the regime being built is the one that should be. The Lycead paper’s argument and the Millerman diagnosis are complementary halves of one observation: the analytic monopoly produces utilitarian AI by structural necessity, and the corrective is to ask the regime question — the question virtue ethics is one of the few analytic-tradition-adjacent frameworks equipped to ask.

How a refinement model addresses each risk

Robertson sketches how a refinement model addresses each of the five risks: replacement is mitigated by treating humans as ends rather than means; catastrophe is mitigated by keeping humans as moral judges and compartmentalizing the blast radius of bad calls; delusion is mitigated by demoting the value of flattery and elevating the value of user improvement; collapse is mitigated by reducing public hostility to AI; liability is mitigated by ceding moral responsibility to the human user.

And he closes the Technological Soul essay with a personal recommendation that goes beyond architecture: religious fasting from alien technology. A secular Sabbath. A day a week with no internet. The discipline of doing things by hand periodically — not as Luddite resistance but as a way of preserving the technē (skill, craft) that makes any technology meaningful in the first place.

Where Wolf You Feed sits in Robertson’s framework

WYF was architected around exactly the refinement model Robertson advocates. The First Law’s emphasis on sovereignty, anti-paternalism, and treating the user as the center of the equation is virtue-ethics-shaped, not utility-shaped. The Society of Mind architecture diffuses any single corporate utilitarian regime. The Commitment ledger and conservation law (Committed = Delivered + Owed) keep moral responsibility seated where Robertson says it must remain: with the human user.

The honest tension, which Robertson would surely raise if asked, is that any Decision Engine — even one designed for refinement — risks being read as another instance of consultation replacing thought. WYF’s answer is that the user must still articulate the situation, evaluate the rendered Decision against their own values, accept or refuse the Commitment Opportunity, and live with the consequences. WYF does not move the bricks for you. It surfaces the decision, holds you to your promises, and refuses to flatter you out of a bad position. The friction remains. The training continues. The byproducts accrue to the user, not to the tool.

Robertson’s other recommendation — periodic fasting from technology — is not something WYF as a product can demand. But it is something WYF as a brand can endorse, and does. Use Wolf You Feed for the decisions that matter. Take Sundays off. Walk outside. Wash the dishes by hand.

The technologies we use shape who we become. That has always been true. AI raises the stakes by an order of magnitude. Robertson’s framework is the most coherent map of those stakes anyone has produced this year. The question he leaves you with is not will you use AI but which AI, on what terms, with what discipline. Wolf You Feed is built to be a defensible answer to that question. The discipline is yours.

See also: WHEN ONE VOICE CAPTURES THE COUNCIL, FOUR RADIOS, FOUR REGIMES, WOLF YOU FEED — IN THREE PARTS, and the Founder Talks archive.

Cited sources (APA 7th):

• Robertson, C. B. (2016, January 16). Phaedrus, agent Smith, and the perils of the GPS [Video]. YouTube. https://www.youtube.com/watch?v=BEwLJopbnPc
• Robertson, C. B. (2023, August 12). Caffeine Stream 43: Simulation theory and artificial intelligence (with Sean Bradley) [Video]. YouTube. https://www.youtube.com/watch?v=hHxnCSaLyzw
• Robertson, C. B. (2026, April 4). Debate: AI and the human soul (with “Kevin”) [Video]. YouTube. https://www.youtube.com/watch?v=g5lQ9YgVRZk
• Robertson, C. B. (2026, April 17). Lycead AI: ‘Refinement’ vs ‘Replacement’ as the end of LLMs [Position paper].
• Robertson, C. B. (2026, April 25). AI is a mirror [Video]. YouTube. https://www.youtube.com/watch?v=ANsrwBZMKsc
• Robertson, C. B. (2026, May 24). A technological soul (essay) #AI #philosophy #technology [Video]. YouTube. https://www.youtube.com/watch?v=lDIcagJjUdQ
• Crawford, M. B. (2009). Shop class as soulcraft: An inquiry into the value of work. Penguin Press. (Referenced for the jig-vs-nudge distinction.)
• Heidegger, M. (1954/1977). The question concerning technology and other essays (W. Lovitt, Trans.). Harper & Row.
• Heidegger, M. (1927/2010). Being and time (J. Stambaugh, Trans.; rev. ed.). State University of New York Press.
• Millerman, M. (2026, April 14). The analytic monopoly on AI philosophy: And why it isn’t enough. Michael’s Musings (Substack). https://millermanschool.substack.com/p/the-analytic-monopoly-on-ai-philosophy