Designing Humans In

Back in 2019, when I was building photorealistic AI avatars, the obsession was with when avatars would be indistinguishable from the humans they were based on. The small tell-tale glitches were off-putting, and crossing them seemed like the gap that would resolve the dissonance.

This reflected the uncanny valley theory: that the discomfort would intensify as the resemblance climbed, and then resolve once the valley was crossed.[^1] But as the fidelity climbed, model by model, for many, the discomfort, as they stared face to face with these AI-generated replicas, remained.

What did resolve it was something else. When we gave the avatars a clear, complementary role alongside humans, extending an instructor into courses they couldn’t teach live, augmenting a service the live version couldn’t reach, the resemblance stopped being the question. The technology had a place alongside the people it was built among, and the unease dissolved into the work.

The same pattern is playing out today at far greater scale, but the question has flipped. The avatar question was whether AI fit alongside humans. The agent question is whether humans fit at all.

The obsession has moved from physical resemblance to capability benchmarks: the bar exam, PhD level, the “billion PhDs in a data center” labs use as a North Star. AI has already cleared many of them. That is why the unease has tipped into fear, especially alongside a narrative that frames replacement as inevitable, that highly capable AI will simply replace you. The feelings are signaling the real worry: being measured against something that has no place for you.

How did the worst sales pitch ever, it will take your job and possibly kill you, as Noah Smith summed it up, come to pass?[^2] Because every inherited structure points the same way. Public companies are rewarded for cutting costs. Private markets need substitution-shaped returns. The labs compete on benchmarks built around what humans do, which keeps pushing the technology toward replacement. Substitution is where the forces lead, and they aren’t going away. Complementarity is what we have to build instead.

What humans must reject is not AI. It is this narrative.

What we saw in miniature with the avatar instructors, technology given a clear, complementary place alongside the people it was built among, is now being designed for, intentionally, across very different contexts.

In physical AI, designers at frog are building robots around the principle of automating the drudgery, not the joy: form chosen for the human context, not to imitate or intimidate.[^3] A senior designer there told me the towering humanoid form embodies, in itself, the idea that we are replacing humans. The contextual form embodies the human at the center of the system. The choice between the two is a choice about whether humans are at the center, or in the way.

At the organizational level, AE Studio works with leaders to put agents into the org chart from the workflow up. The order matters. Most companies start with the org chart and ask where they can cut. Starting with the workflow asks a different question: where does AI change the shape of the work, and what becomes possible that wasn’t before? The answer reshapes what people in the workflow do. With the routable layer absorbed by agents, every engineer is now leading a pod of agents. Every PM is coaching more PMs. Every individual contributor is shipping work that used to need a manager’s review. The workflow requires the people in it to step up. The choice between the two starting points is between cutting and rebuilding.

Substitution’s returns are predictable: the same work, fewer people, lower cost. Complementarity’s returns are unprecedented and not so easily measured, but innately the only path forward. One way to tell them apart: if the goal is fewer people doing the same thing, it is substitution. If the goal is the same people doing things that were not possible before, it is complementarity.

Substitution is the default. Complementarity has to be designed in. It is the hard human work of reinvention, the kind we must claim. Do that work, and we get the leverage AI offers. Skip it, and the work routes around us, instead of through us, and for us.

• The uncanny valley is a concept from robotics, named by Masahiro Mori in 1970. As a humanoid form gets closer to looking human, comfort rises, then collapses into revulsion at the last gap. The theory predicted that crossing the gap would resolve the discomfort.

• Noah Smith, “AI has the worst sales pitch I’ve ever seen,” Noahpinion, March 2026.

• See frog’s recent Nome project, a domestic robot designed around the principle of automating drudgery rather than joy, with form chosen for the human context.