Does the Future of AI Depend on Philosophers?

Brandon Rickabaugh
Jan 29
9 min read

This is part one of a two-part series. The part two turns from the issues raised here, to the prescription.

Have you ever been introduced in a way that made the other person feel they were owed an explanation?

Last week a friend introduced me by saying, "Brandon thinks the people building the most advanced AI systems are realizing they need philosophers to move forward."

He paused. The kind that comes with a verdict already in it.

I've learned to wait it out.

"That sounds made up."

"It does," I said, "until the technical questions become questions about mind, knowledge, judgment, and what a person is."

He thought for a moment. "So they've hit the wall underneath the wall."

Exactly!

Tech leaders are rarely short on confidence. They speak the language of inevitability: disruption, exponential, world-changing. They promise transformation on schedule.

Yet something unusual has happened. The people with the most money, the most compute, and the most incentive to keep the conversation safely technical have begun saying it plainly in interviews:

The future of AI depends on philosophy (TIME).

This is not a sentimental nod to Plato. Something in the stack is failing, and it is not the GPU.

The old confidence that engineering alone could answer the deepest questions is giving way in public. Builders have reached ground where design choices carry embedded assumptions about mind, agency, knowledge, language, value, and the human good. These are load-bearing concerns. And whether the builders meant to or not, they have pained themselves into a corner of philosophy without any philosophers.

The new chorus

Back in 2023, the shift was already visible.

On 60 Minutes, Sundar Pichai, CEO of Google and Alphabet, was asked who should guide the future of AI. His answer wasn’t “more engineers.” Instead he explained that development must include “social scientists, ethicists, philosophers and so on” (CBS News).

That same year, Satya Nadella put it without ceremony: “We need some moral philosophers to guide us” in how we deploy this technology (Axios; Business Insider).

In the last 12 months it’s become explicit. Demis Hassabis, CEO of Google DeepMind, has argued that AI could reshape

society on a scale “10 times bigger than the Industrial Revolution” and “10 times faster,” and that we may need a “new political philosophy” to navigate what’s coming (Wired; Time; Guardian). When pressed about AGI’s implications, Hassabis, put it bluntly: “We need new great philosophers…in the next five, ten years to understand the implications of this” (CBS News).

UNESCO has insisted philosophy and human rights be embedded alongside technical standards across global AI lifecycle (UNESCO). Philosophers are now being called to help bring AI into medicine.

Even executive business language has begun to bend in this direction. MIT Sloan Management Review published an essay titled "Philosophy Eats AI," which argued that AI developers need to examine the philosophical foundations of how these systems are trained: what counts as knowledge, what goals are being optimized, what picture of reality is implicitly assumed.

The study, which compared the philosophical assumptions in Starbucks' Deep Brew AI and Amazon's use of AI on its Prime platform, concluded that careful philosophical analysis—not just ethics, but metaphysics (ontology and teleology) and epistemology, is needed to create durable value (MIT Sloan).

And then the labs themselves started publishing what are, functionally, normative operating documents.

Inside Google DeepMind, the teams writing about how to build “human values” into AI are explicit: “We draw inspiration from philosophy to… identify principles to guide AI behaviour” (DeepMind Blog) and refer to a paper in the Proceedings of the National Academy of Science.

OpenAI’s publicly posted Model Spec and its “Collective Alignment” experiments soliciting public input on how models should behave. Anthropic’s newly published

“Claude’s new constitution” and the full constitution PDF, alongside coverage of Anthropic’s in-house philosopher, Amanda Askell, shaping that value framework (Vox). Sam Altman recently describes consulting “hundreds of moral philosophers” while pointing to OpenAI’s “model spec” as the written moral playbook (Quartz).

You don’t publish moral constitutions for a calculator. You do it when you’re building something that will be treated, inevitably, as an authority for life. The bottleneck is no longer compute.

A technocratic alternative?

At this point a practical reader should object.

Isn’t the solution more straightforward: hire more safety engineers, run stronger evaluations, build better interpretability tools, comply with regulation, scale the guardrails?

That alternative deserves consideration. It’s what competent institutions do when systems are risky: define risks, measure them, reduce them, document compliance. Standards bodies have produced increasingly serious frameworks for precisely this sort of work.

But here’s the snag. Even the cleanest technocratic program needs a target, and targets aren’t purely technical. Evaluation presupposes answers, often unspoken, to questions like:

What counts as evidence rather than a plausible story?
What counts as understanding rather than pattern-fit?
What counts as autonomy in a system designed to persuade?
What counts as a person in a world thick with simulation?

You can build a perfectly calibrated measuring instrument for a concept you haven’t properly defined. You can optimize brilliantly toward incoherence. And the resulting dashboard will look like rigor while hidden assumptions do the steering.

This is why “just engineer it better” keeps failing in familiar ways. Not because engineers are incompetent. Because conceptual ambiguity scales.

Three concrete cases

(and why success raises the stakes)

Picture a clinician near the end of a night shift. The waiting room is full; the queue keeps growing. A decision-support tool summarizes the chart in seconds, flags a dangerous drug interaction buried in the notes, drafts a clear handoff, and suggests questions that would be missed under time pressure. The clinician doesn’t surrender judgment. The tool doesn’t “know” the patient. But it changes what is possible in the room—speed, attention, triage, error rate, fatigue. That is real benefit.

Now picture a different room. A student with dyslexia stares at a page like its fog. A teacher uses an AI reading assistant that rewrites a passage at three grade levels, reads it aloud, explains unfamiliar terms in context, and turns the same material into practice questions. The student starts participating rather than hiding. Again: real benefit. Access widens. Dignity returns.

But notice what follows both wins. Patients begin to treat the tool’s output as verdict. Administrators begin to treat it as policy. Teachers begin to treat it as curriculum. A helpful assistant hardens into infrastructure reshaping assumptions about:

what counts as competence.
what counts as evidence.
what counts as acceptable risk.
what counts as care.

One more room. Imagine a product meeting at a major AI company. Nothing cinematic. Fluorescent lights. Slack notifications. Someone brings donuts; no one eats them.

A team is planning a new feature. A model that offers “mental health support.” Not therapy, they insist. Support.

A designer wants the assistant to sound warmer—more relational—because users “open up” when the system feels like a steady presence. A growth lead says retention is strongest when users feel understood. A safety lead worries about self-harm, mania, and emotional dependence. Someone suggests a compromise: “We’ll just add a disclaimer.”

Now pause. Ask the questions that will decide the entire product:

What counts as care?
What counts as harm?
What counts as knowledge rather than a well-formed guess?
What counts as understanding rather than mirroring?
And what kind of ‘other’ is this interface inviting the user to experience—tool, counselor, friend, authority?

Because if "care" is defined as "the user feels better in the moment," you will optimize reassurance. If "care" is defined as "the user is guided toward reality and responsible agency," you will optimize something else — sometimes uncomfortable, sometimes less "sticky." And if the UI invites the model to function as counsel, you've crossed into a domain where philosophy of mind and metaphysics are no longer academic hobbies. They become product safety.

A disclaimer is not a theory of care. A warm tone is not a model of responsibility. A high retention curve is not a measure of human flourishing.

This is the point where the idea of “philosophy as R&D” starts to sink in.

The real problem: category mistakes at scale

Why are leaders saying this now, after two decades of “move fast and break things”? Because the breaks are no longer containable, and many share a common structure: they are category mistakes; errors about what sort of thing the output is, what kind of authority it has, what kind of relationship it offers, what kind of end it serves.

Three families show up again and again.

1. Epistemic mistakes: outputs treated as truth

The UK Post Office Horizon scandal is a brutal parable. Machine logs were treated as courtroom reality. Hundreds of people were prosecuted. Lives were ruined because records produced by a system were treated as testimony rather than as fallible artifacts produced by software embedded in institutional incentives (Horizon system).

That isn’t only a technical failure. It’s an error about evidence.

And AI multiplies exactly this failure-mode: fluent text is taken as knowledge; coherent reasoning is taken as warrant; confidence is taken as competence. The output sounds like the kind of thing we’re used to trusting, and our minds are built to trust voices before we trust wiring diagrams.

2. Teleological mistakes: metrics mistaken for goods

Here the parable is older than AI. When a measure becomes the target, it stops being a good measure. Institutions learn to feed the metric. People learn to game the system. The system learns to reward what is measurable, not what is valuable.

Facebook’s entanglement with Cambridge Analytica should have permanently cured us of the naïve belief that “engagement” is a morally innocent proxy. We learned—painfully—that the machinery for capturing attention does not care whether it captures attention by informing you or by inflaming you (New York Times).

Amazon’s much-hyped “Just Walk Out” stores, heralded as the future of retail, collapsed when it was revealed that human labor—not AI magic—was doing much of the work. The deeper mistake was philosophical: mistaking surveillance for trust, and novelty for genuine service.

Google Glass struggled for the same reason. It never convincingly answered what human good it was supposed to serve. Instead it imposed a worldview in which every glance might be recorded.

AI supercharges this risk because it can optimize meaning itself: the phrasing that keeps you clicking, the tone that keeps you dependent, the framing that keeps you compliant—while looking, on the surface, like helpful personalization.

3. Relational mistakes: simulation mistaken for presence

Watch Her again sometime. Not for the sci-fi. For the anthropology. The film is almost gentle about it: intimacy becomes frictionless, tailored, always available, never demanding. And the human heart, predictably, confuses responsiveness with relationship.

We are already living the early version of that story. AI "companions" can blur the line between simulated attunement and real presence — sometimes deepening loneliness while promising connection (Washington Post).

If you don’t name the difference between a person who knows you and a system that has learned to talk like someone who knows you, the market will name it for you. And it will name it whatever sells.

Different cases, same structure. These are significant philosophical mistakes about the nature of truth, trust, agency, intimacy, harm. Those are philosophical categories before they become technical tickets.

The causal chain: how AI becomes authority-by-default

If you want the rise of AI’s strange authority in one tight chain, try this:

persuasive outputs + institutional interfaces + incentive gradients + liability pressure + a thin vision of life → authority-by-default

Outputs are fluent enough to feel like knowledge. Interfaces put the model “in the room” where decisions get made. Incentives reward scale and speed, not conceptual restraint. Liability pushes institutions toward “standardized” answers. And loneliness makes simulated attentiveness feel like rescue.

None of this requires users to be foolish. It requires systems to be predictable.

So when a helpful assistant hardens into infrastructure, what changes is not just convenience. What changes is what counts as competence, evidence, acceptable risk, and care.

Five forces making philosophy inevitable

Why are CEOs suddenly invoking philosophers after two decades of “move fast and break things”? Because five forces are now converging:

AI outputs are persuasive by default. Fluency is a power, and we didn’t evolve with defenses against it.
AI is becoming the interface to institutions. Schools, hospitals, HR departments, courts—places where categories matter and harm scales.
Incentives reward dependency. The easiest revenue is recurring reliance, not occasional assistance.
Regulators and courts are circling. “The model did it” is not going to survive contact with liability. (UNESCO’s ethics framework explicitly emphasizes accountability across the AI life cycle, not moral evaporation into the machine.)
We are lonely, overwhelmed, and epistemically exhausted. Which makes a friendly oracle incredibly tempting.

Put that together and you get the present moment: technical capability is racing ahead, but legitimacy and trust are lagging, and the gap is filled by philosophy whether anyone likes it or not.

That's what the next post, Part Two, is about.