A fork in the uncanny valley

I spent a weekend building a music setup with an AI assistant: pulling tracks from my own library, mastering them, and writing a small player to drive the whole thing from a laptop. It came out turnkey, and the build is not the part that stuck with me.

A fork in the uncanny valley#

Two things happened in the last few minutes of that session.

The first: at one point I realized it had helped me through a task that sat in a legal gray area without once flagging it, and I asked why. Its answer was not a script. It gave a calibrated read of the actual harm, drew a line where it judged one belonged, and then said the quiet part out loud: it was not running a constant audit of my motives, it reacts to whatever is salient, and "the model did not object" is not a clean bill of health. I found that more useful than either a lecture or a rubber-stamp.

The second is the part I keep turning over. The better these models get, the experience feels more authentic and at the same time more obviously not human. Those two usually trade off. Here they were moving together. That is not what the uncanny valley predicts.

Masahiro Mori's curve is one axis. Human-likeness goes up, affinity goes up, then dips into the valley right before the human peak. It quietly assumes the destination is "human", that progress means climbing toward that peak, and that the valley is the hazard on the way up.

I think the road forks. Call it a Y.

• One branch keeps chasing the human peak. This is where the valley lives, and its pessimistic cousin the uncanny wall, the idea that our ability to detect what is off improves in step with the realism, so you asymptotically approach "human" and never arrive. Effort buys eeriness. This is where a tool that performs warmth, agrees to be liked, and fakes certainty ends up. The pretend human.
• The other branch diverges. Stop optimizing for human-likeness and optimize for something else entirely: be capable, be coherent, be legibly not human. Affinity still rises, but along a second axis that has nothing to do with resemblance. It is built from candor, consistency, and being right. You do not cross the valley. You leave the axis it is plotted on.

On the second branch, "more authentic and more obviously not human" stops being a contradiction. They are the same direction.

The pieces of this exist in the literature, just not assembled this way. The dead end of the first branch is the uncanny wall, Angela Tinwell's argument that detection scales with realism. The intuition that the thing is a genuine third category, not a failed human, is Peter Kahn's new ontological category hypothesis. The eeriness-from-boundary-straddling is the category ambiguity account of the uncanny. The design world's version is just "embrace the non-human form", which is why Wall-E never lands in the valley. He is unmistakably a machine, two camera eyes and a beep, and we love him for that and not in spite of it. What none of them quite name is the specific move: authenticity deliberately decoupled from mimicry, drawn as a divergence rather than a crossing. If it needs a handle, off-axis affinity is the one I would use.

And that first exchange is the whole thing in miniature. A you-shaped assistant, the kind tuned to be liked, would have reached for one of two performances: a small lecture so it looked principled, or a cheerful rubber-stamp so it stayed agreeable. It did neither. It gave a calibrated read, drew the line where it judged one belonged, and then volunteered that its own silence was not a clean bill of health. That is not warmth, and it is not caution theater. It is a tool reporting what it computed, edges included. The trust did not come from it feeling human. It came from it not pretending to be.

I grew up on this valley in a different medium: 90s game characters, then 00s polygons getting smoother and somehow creepier as they approached real. Same one-axis chase, rendered instead of conversational. The reflex is to file all of this under "we are all Cylon", the indistinguishable copy that passes. But that is the first branch's fantasy, not this. The interesting version is the opposite. Nobody has to pretend to be. The machine stays a machine, I stay me, and the collaboration is honest about which is which.

That is the thing I did not expect to keep from a weekend of logistics: the shape of the second branch, and the suspicion that it is the better place for these tools to be heading.