What Is Human-Centered AI?

Stanford's Institute for Human-Centered Artificial Intelligence — co-founded by Fei-Fei Li and John Etchemendy in 2019 — has a page that answers a question a lot of companies are now asking out loud: what is human-centered AI? The definition is clear, well-reasoned, and, if you actually run it against a real deployment, incomplete in a specific and useful way.

Not wrong. Incomplete. It tells you what a good AI system has to optimise for. It doesn't tell you how the organisation around that system has to be built for the optimisation to survive contact with a live rollout, a budget cycle, and a manager under pressure to ship. That second part is a different problem, and it's the one most companies actually fail on.

What Stanford HAI means by “human-centered”

The core claim is straightforward: AI development should be organised around human needs, values and well-being, not around what the model is technically capable of doing. Human-machine collaboration is the operating assumption, not automation for its own sake. Empathy, ethics and user experience sit inside the design process from the start, rather than getting bolted on as a compliance review before launch.

It's explicitly an interdisciplinary framing — computer science on its own doesn't get you there. HAI's own account of the field draws in psychology, ethics and design, on the reasoning that a system can be technically excellent and still fail the people who have to use it, be governed by it, or live with its decisions. The standard the field sets for itself is that systems should end up trustworthy, inclusive, and aligned with the goals of the society deploying them — not just accurate on a benchmark.

That's a design philosophy, and a reasonable one. It answers the question "what should this AI system be like." It's silent on a different question: "what does the company that builds or buys this system have to look like for that standard to actually hold once real incentives, real deadlines and real org charts get involved?"

What the definition asks for — and what it doesn't specify

Read the human-centered AI literature closely and you'll notice it operates almost entirely at the level of the system: interface choices, model behaviour, evaluation criteria, governance principles applied to a given AI product. That's the right level for the questions it's trying to answer. It is not the level at which most human-centered AI actually fails inside companies.

Here is the pattern, and it repeats often enough to be worth naming precisely: a company adopts every human-centered principle on the list — ships an AI feature with careful UX, writes a responsible-use policy, runs a bias audit — and the system still ends up used in a way nobody who wrote those principles would recognise. Not because the principles were wrong. Because the organisation around the system was never redesigned to hold them. The customer service team is measured on tickets closed per hour, so the empathetic, human-reviewed AI workflow gets skipped under load. The product team owns the model, the compliance team owns the policy, and the two groups meet once a quarter. The well-being the system was designed to protect gets traded away by people who were never in the room when the design happened, because their incentives, their reporting lines and their tools were never touched.

Human-centered AI, as a design discipline, doesn't have a mechanism for that failure mode, because it isn't a discipline about organisations. It's a discipline about systems. The gap isn't a flaw in the definition. It's the edge of what the definition was built to cover.

AI-Human Systems: the organisational answer to the same problem

This is the specific gap AI-Human Systems is built to close. The doctrine's claim is narrower and more mechanical than it sounds: technology, people and business are not three separate domains to be managed in sequence — a tech rollout, then a training programme, then a strategy review — they are a single system that has to be architected together, or the parts quietly work against each other.

It comes from watching the same failure pattern recur across twenty-five years of unrelated organisations: football clubs with advanced technology and teams unprepared to use it, schools with a progressive vision and teachers without the tools to execute it, companies with million-euro ERP implementations that staff routed around with Excel within a year. Different industries, same structure — strategy, technology and people funded, planned and evaluated as three separate projects instead of one system, so the system's actual behaviour ends up decided by whichever of the three has the least resistance, not by whichever design was intended to win.

Applied to human-centered AI specifically, the doctrine reframes the diagnosis. A human-centered AI system that fails in production usually isn't failing because the interface or the model was badly designed. It's failing because the strategy that approved the project, the technology stack it was built on, and the people expected to operate it inside it were never architected as one system with a single owner. Human-centered AI describes the target state of the artefact. AI-Human Systems describes the organisational condition that has to exist for a company to actually reach and hold that target — who owns the decision when the empathetic design and the throughput target conflict, what the system is allowed to do without a human in a specific role approving it, and what changes in how people are trained, measured and organised when the AI layer goes in.

Human-centered, responsible, ethical: three labels, three different jobs

Stanford HAI's own materials flag, without fully spelling out, that human-centered AI sits near two other labels that get used almost interchangeably in most corporate AI documentation: responsible AI and ethical AI. They're not the same job, and collapsing them is part of why so many AI governance programmes end up doing none of the three well.

Ethical AI is the narrowest of the three — it asks whether a specific decision, output or system design is right or wrong by some stated standard: fair, non-discriminatory, honest about its limitations. Responsible AI is broader and more procedural — it's about the governance apparatus around the system: who's accountable when it fails, what gets audited, what gets documented, what regulatory obligations get satisfied. Human-centered AI is broader again, and less about compliance than about design intent — it asks whether the system was built around what the humans involved actually need, not just whether it avoids specific harms or satisfies a specific audit trail.

A system can pass an ethics review, satisfy a responsible-AI governance checklist, and still fail the human-centered test — technically fair, formally accountable, and still built around what the model can do rather than what the people using it actually need. That's a common enough combination that it's worth checking for directly rather than assuming good marks on one label imply good marks on the others.

Where the two overlap, and where they don't

It's worth being precise about this, because the honest version of the relationship is more useful than a tidy one. Human-centered AI and AI-Human Systems are not the same concept wearing two names, and treating them as interchangeable would misrepresent both.

Human-centered AI is a design philosophy for AI systems specifically — it applies at the level of a model, a product, an interaction. AI-Human Systems is an organisational doctrine that predates AI as a category and applies to any technology-people-strategy problem; AI is simply the sharpest current instance of it. One is scoped to the artefact. The other is scoped to the company that builds, buys or operates the artefact.

Where they meet is the point that matters for anyone actually running an AI programme: a human-centered AI system, deployed inside a company that hasn't architected itself as one system, degrades into exactly the kind of technically-compliant, practically-ignored initiative described above. And an AI-Human System that gets the organisational architecture right but ships an AI product with no regard for the human-centered design principles — opaque, extractive, indifferent to who actually has to work with it — has solved the org-chart half of the problem and handed the result to a badly built system. Neither one substitutes for the other. Used together, one sets the standard for what the system should be; the other is what makes that standard survive being deployed by a real company with real incentives.

The five layers a human-centered AI system actually has to pass through

The AI-Human Systems doctrine breaks the single-system claim into five layers that have to be architected together rather than handed off in sequence: strategic direction (what the organisation is actually trying to achieve, and what trade-offs it has already decided it won't make), decision architecture (who or what has authority over which decisions, and on what information), human capability (whether the people expected to operate alongside the system have the skill and the standing to do so), operating system (the processes, incentives and workflows the system has to live inside), and technology — the AI itself, last, not first.

Run a human-centered AI initiative through those five layers and the usual failure point becomes visible immediately: the technology layer gets built to a genuinely high human-centered standard, and the other four layers are either skipped or delegated to a different team with a different budget and a different set of incentives. Decision architecture, in particular, is where the earlier reviewer example breaks — the AI was designed to defer to human judgment, but no one redesigned who has the authority, the time or the performance incentive to exercise that judgment. The technology layer did its job. The other four didn't get architected at all.

What this looks like in practice

The pattern, stated generally rather than as a specific case study: a company brings in an AI tool built to every human-centered principle on the list — transparent reasoning, a clear human-override path, careful attention to who's affected by its outputs. Three months after launch, usage data shows the override path is almost never used, not because the AI got everything right, but because using it costs the reviewer time they don't have and isn't tracked anywhere that affects their performance review. The system is human-centered by design. The organisation around it never redesigned the reviewer's incentives, authority or workload to make exercising that design possible. Diagnosing that gap, and rebuilding the ownership, incentives and workflow around the tool rather than the tool itself, is AI-Human Systems work applied to a human-centered AI problem — not a contradiction between the two, but the second half of a job the first half never claimed to finish.

Is Human-Centered AI the same as AI-Human Systems?

No. Human-Centered AI is a design philosophy for AI systems — how a model or product should be built to serve human needs, values and well-being. AI-Human Systems is a broader organisational doctrine about how strategy, technology and people have to be architected together as one system, which applies to AI deployments among other things. They address different layers of the same underlying problem.

Who defined Human-Centered AI?

The term is most closely associated with Stanford's Institute for Human-Centered Artificial Intelligence (Stanford HAI), co-founded by Fei-Fei Li and John Etchemendy in 2019, though the underlying idea — that technology should be designed around human needs rather than technical capability alone — has roots in human-computer interaction and design research that predate the institute.

Why isn't a human-centered AI design enough on its own?

Because a design principle only holds inside an organisation that's structured to enforce it. Without a clear owner, without incentives that reward using the human-centered features rather than routing around them, and without authority given to the people expected to exercise judgment, a well-designed system degrades into a compliant artefact that nobody actually uses as intended.

How do I know if my organisation is ready to build human-centered AI systems?

Check whether the three pillars are already treated as one project: does the team accountable for the AI system's design also have authority over the incentives and workflows of the people expected to use it, or are those decisions split across departments that meet quarterly at best? If they're split, the organisational layer needs work before the design layer can hold.

António Martins is an AI-Human Systems Architect and founder of Bitsapiens, working on the architecture that connects strategy, decisions, human capability, operating systems, and technology.