The Human Layer
Every generation gets its version of the same anxiety. The machines are coming. This time, though, the anxiety feels different, because what AI is beginning to automate isn’t just physical effort or routine calculation. It’s starting to touch something closer to cognition itself. And that’s where it gets interesting.
What I noticed reading through this issue is that the articles don’t really support the replacement narrative. They complicate it. Again and again, contributors working in completely different fields arrived at the same uncomfortable finding: the closer AI gets to something genuinely human, the more visible the gap becomes.
Kelly Pyrpyli makes this point through engineering. AI copilots can generate designs and calculations faster than any human team. But speed isn’t the problem engineering was ever really trying to solve. The hard part is knowing which answer to trust, and why. That judgment hasn’t been automated. If anything, it’s become more central.
Eleni Kli sees the same thing in coaching. The tools exist now to track goals, personalise feedback, nudge behaviour at scale. What they can’t do is replace the quality of attention between two people in a room. Coaching works because of trust and emotional attunement, and those things don’t transfer to a dashboard. The technology can support the relationship. It can’t stand in for it.
The Speech AI contributions were the ones that stayed with me longest. Somi Jeong’s article documents something embarrassing: persistent accuracy gaps across dialects, years after the field promised to address them. A fairer model and a more accurate model, it turns out, are not the same thing. The second piece goes further. Some of the hardest problems in speech recognition aren’t acoustic at all. They’re about what silence means. What background sounds carry. Context that no training dataset has cleanly captured.
Saurabh Shewale’s piece on physical AI and digital twins is in some ways the most optimistic in the issue. The engineering is genuinely impressive : entire industrial environments simulated, tested, and refined before anything is built. But even here, someone has to decide what the system is optimising for. AI can refine almost any objective function you give it. Choosing the right objective is a different kind of problem entirely.
Eya Arfaoui closes the issue by following accountability into distributed AI systems, where responsibility is shared across developers, deployers, data providers, and increasingly autonomous models. Her argument is structural: you can’t locate accountability after harm occurs if it was never built into the architecture to begin with. Governance can’t just be a layer added at the end.
Different problems. And underneath all of them, the same pressure point: not whether AI can perform a task, but whether it can carry the weight of what that task actually requires. Judgment. Trust. Context. The ability to decide not just what can be done, but what should be.
That’s not a soft argument for human exceptionalism. It’s a practical observation about where AI keeps running into the same kind of wall.
My thanks to every contributor, reviewer, editor, and volunteer who made this issue what it is. Every month I’m reminded that the most interesting thinking about AI doesn’t come from inside any single discipline. It comes from the edges, where fields rub up against each other and produce friction.
Last, but not least, many thanks to our featured guests, Andrej Savin and Elsa Sklavounou for their valuable insights.
The Human Layer is still there. It took AI to make it visible.
