This is Part 2 of a three-part series. Part 1: The Forgotten Book established that moral responsibility is the foundation, not the add-on. Part 3: The Question of Time asks what we do with the hours machines give back.
In Part 1, I argued that “Tech for Good” is a philosophical error — that Adam Smith’s Theory of Moral Sentiments established contribution as the default, not a department. The impartial spectator cannot be outsourced. Medicine understood this. The early internet understood this. Somewhere along the way, technology forgot.
The question now is: what does remembering look like?
Architecture, Not Afterthought
There is a difference between a building designed with structural integrity and a building with a safety inspector bolted on after construction. The first is architecture. The second is liability management. Both may keep the building standing, but only one was designed to.
This is the distinction that matters for AI. The question is not whether we should add guardrails, review boards, and ethics committees to systems that were designed without them. The question is whether we can design systems where the guardrail is the structure itself.
Consider how we build bridges. No civil engineer designs a bridge and then convenes a committee to decide whether it should hold weight. Load-bearing capacity is not a feature request. It is the reason the bridge exists. The constraint — that it must not collapse — is not a limitation on the engineer’s creativity. It is the source of the engineer’s creativity. Every elegant bridge in history is a solution to the problem of gravity, not a protest against it.
Primum non nocere works the same way. It is not a restriction on medical practice. It is the foundation that makes medical practice possible. A surgeon who does not first commit to doing no harm is not a surgeon at all. The constraint is generative, not restrictive.
And yet, in technology, we have somehow arrived at the opposite assumption: that constraints are obstacles to innovation, that ethics is a drag coefficient on progress, that responsibility is what you add after the product ships. “Move fast and break things” is not just a slogan. It is an architectural philosophy — one that treats harm as an externality to be managed rather than a force to be designed against.
The Defaults We Choose
Every system has defaults. The question is whether we choose them deliberately or inherit them by neglect.
When a social media platform defaults to maximizing engagement, it has made a choice — even if no one in the room thought of it as a choice. YouTube’s recommendation algorithm, designed to maximise watch time, was found to systematically guide viewers from mainstream content toward increasingly extreme material — not because anyone intended radicalisation, but because outrage is engaging. The default produced the outcome. When a recommendation algorithm defaults to showing content that provokes the strongest emotional response, it has made a moral decision — even if it was framed as an optimization problem. When a language model is trained on the entirety of the internet without deliberation about what “the entirety of the internet” contains, a default has been set.
The defaults are the architecture. Everything else is renovation.
This is why the “AI safety” framing, while necessary, is insufficient. Safety is what you add to a system to prevent failure. It is defensive. It asks: how do we stop this from going wrong? But the more fundamental question is generative: what does it look like when it goes right?
Consider the difference:
A safety-first approach to AI asks: How do we prevent this model from generating harmful content? A contribution-first approach asks: How do we design a model whose default output advances understanding, respects dignity, and enriches the person using it?
A safety-first approach to a recommendation algorithm asks: How do we filter out misinformation? A contribution-first approach asks: How do we build a system whose natural tendency is to surface what is true, useful, and nourishing?
These are not the same question. The first is a filter. The second is a foundation. One catches what falls through. The other determines what gets built in the first place.
The Hippocratic Compiler
What would a primum non nocere for technology actually look like?
Not a compliance framework. Not a checklist. Not an ethics board that meets quarterly while the engineering team ships daily. It would look like a set of constraints so deeply embedded in the practice of building technology that violating them would feel as unnatural as a doctor deliberately harming a patient.
Some of this is beginning to emerge, though we rarely name it for what it is.
When Anthropic publishes its model card detailing a system’s known limitations before release, that is a form of informed consent — the same principle that requires a surgeon to explain the risks of a procedure before operating. When open-source developers release their models with clear documentation of training data provenance, that is a chain of custody — the same principle that tracks a pharmaceutical from laboratory to pharmacy shelf.
When a company chooses not to deploy a capability it has built, because the societal implications are unclear — that is not a failure of ambition. That is the impartial spectator at work. Smith would have recognised it immediately: the practice of imagining how your actions appear to someone with no stake in your success.
The challenge is that these practices remain exceptional. They are celebrated precisely because they are rare. The engineer who refuses to build a surveillance tool is lauded as a whistleblower, when they should simply be recognised as a professional — someone practicing their craft with the same baseline responsibility as a doctor, a lawyer, or a structural engineer.
The IEEE and ACM recognised this decades ago when they jointly published a Code of Ethics for software engineers modelled on exactly this principle — that software engineering is a profession, with the obligations that word implies. And yet the code remains largely ceremonial. The engineer who follows it is still treated as exceptional rather than baseline.
From Extraction to Contribution
Extraction is not a Silicon Valley invention. Spain deforested its peninsula to build armadas. Russia hunted the sea otter to near-extinction for pelts. Mining has always taken more than it returned. What the last two decades of technology added was not the impulse but the scale — and the abstraction. It is easier to extract when what you are taking is invisible: attention, data, engagement, the slow erosion of a person’s capacity to sit with boredom. This is not a conspiracy. It is a default — the natural consequence of optimizing for growth in the absence of the moral framework Smith described.
Generative AI does not have to inherit this default. It is new enough, and its architecture is malleable enough, that different choices are still possible. But the window is not indefinite.
What would contribution as a default look like in practice?
It would mean that an AI assistant’s success is measured not by engagement time but by whether the person using it accomplished what they needed and moved on with their day. It would mean that a content generation tool is evaluated not by output volume but by whether it helped someone think more clearly. It would mean that an AI-powered search engine is judged not by ad revenue but by whether it consistently led people to what was true.
These are not utopian fantasies. They are design choices. They are the choices that Stef made when he gave MIME to the world. That Cerf made when he designed protocols for a planet he would never visit. That every doctor makes, every day, when they place the patient’s interest above their own convenience.
The question for this generation of technologists is simple, and it is the same question Smith posed in 1759: Can you imagine yourself in the position of the person affected by your work? Can you see your product through the eyes of someone with no stake in your success?
If you can, the design follows naturally. If you cannot — or will not — then no amount of safety tooling, ethics boards, or “AI for Good” programs will compensate for the missing foundation.
The impartial spectator is not a feature. It is the architecture.
In Part 3, we turn to the question that Keynes posed nearly a century ago — the one that generative AI makes inescapable: if machines can do the work, what do humans do with the time?