Life on the Uncanny Precipice

Masahiro Mori's 1970 Uncanny Valley concept has been inverted by modern AI, leaving people on an 'Uncanny Precipice' where every mediated encounter—calls, emails, voices—feels potentially synthetic, eroding trust in real human interactions. The author describes calling a hospital urgent line and being unable to tell if the coordinator was an AI, and cites Harry Harlow's 1950s monkey experiments to argue that even routine human contact matters, especially for isolated individuals.

Life on the Uncanny Precipice I called the hospital’s urgent line today. A new coordinator answered the phone. She was warm, professional, and direct. Her voice was melodious. She barely paused in her list of questions, and as I gave details to relay to the on-call physician, I tried to ignore a sick itch in my brain: Is that a fucking AI? In 1970, roboticist Masahiro Mori first described the “Uncanny Valley”. According to Mori, as simulations of life—robots, animations, illustrations—become more lifelike, they become more sympathetic. However, when a simulation is too lifelike, but not entirely convincing, it becomes grotesque. This space between the alien and the fully human is the Uncanny Valley: the repulsive arena of corpses and RealDolls. This problem was always going to be temporary. As artificial creatures became more realistic, they would venture deeper into the Uncanny Valley. One day, they would become perfectly realistic. They would exit the Uncanny Valley, re-entering the Familiar. Nothing would be uncanny. Everything would feel safe. Sure enough, the simulations climbed out of the valley. The Turing Test became yet another saturated benchmark, as LLMs and image generators became indistinguishable from reality. But the simulations didn’t leave the uncanniness behind; they carried it with them. It infected everything, eroding the boundaries of the Uncanny. And so you don’t stand in the Familiar. You stand at the edge of a cliff. Every mediated encounter—every call, every email, every voice saying your name—is a precarious ledge. At any moment, it can collapse under the weight of an em-dash or a familiar phrase. And when it does, you discover that you were alone the whole time. There you are, performing sincerity for nobody, talking to yourself. You now live on the Uncanny Precipice. In the 1950s, the psychologist Harry Harlow separated infant rhesus monkeys from their mothers and gave them two substitutes. One was a frame of bare wire holding a bottle of milk. The other was covered in soft terrycloth and offered nothing. The behaviorists of the era predicted the monkeys would attach to the wire mother, since she was the one who fed them. Instead the infants took their milk from the wire mother and spent the rest of the day clinging to the cloth one. When something frightened them, the cloth mother was where they ran. Nutrition, it turned out, was the smallest part of what a mother was for. Your day is full of routine functional interactions which carry the possibility of a small human connection. A cashier smiles at you, a receptionist nods at you, a nurse thanks you as she accepts your clipboard. For most people, most of the time, these interactions could disappear without effect. All we need is a wire mother. But for some people, those interactions form most of their daily human contact. Their loss is felt. Do you remember COVID isolation? And on some occasions—when you call a hospital’s urgent line—you do not want to be alone. You look to the cloth mother. Is she still there? When I called the hospital, I didn’t know if the coordinator was really an AI. I still don’t know. A few years ago, I would have reacted normally to the same melodious voice. This year, a sense of unreality touches every encounter, a sort of preemptive dissociation. Patients with dementia can experience Capgras syndrome, a horrifying delusion that their loved ones have been replaced with identical impostors for whom they feel nothing. They recognize faces, but the familiarity is gone. I am losing the ability to see real humans. Recently, I received a project proposal. It took a bit of effort, but eventually I grasped the idea. It might be worth pursuing. Later, I returned to the proposal, ready to reply—but this time I registered the distinct signature of synthetic text. Suddenly I wasn’t sure if their message really described this idea at all. Had they merely gestured at a connection, and allowed me to fill in the blanks? Had I effectively authored the proposal myself in an effort to read meaning where there was none? A sentence means something because a person meant something by it, and we interpret by making inferences about our communication partner. Meanwhile, our partner chooses words carefully to convey that intent over other meanings. The philosopher Paul Grice called this the cooperative principle: the speaker is conveying some intent as well as they can. According to Grice, this assumption is the reason why humans are not overwhelmed by possible interpretations of every utterance. When we struggle, we might be trying to recover a challenging concept or we might be accommodating our partner’s language skills. But we trust that the meaning is there, in our partner’s intent. If we have no partner, if there is no intent, we eventually fabricate meaning from the vacuum. Of course, LLM outputs are not entirely empty of meaning. They can reason through open math problems, after all, so synthetic text must convey some valuable suggestions and correct claims. But they often fail to comply with Grice’s rules of pragmatic communication because they are not engaged in a cooperative exchange. LLMs will generate extraneous filler, add irrelevant caveats, confabulate, and compress claims beyond easy recovery. To find a nugget of meaning, the user must sift through mounds of slop—and some simply won’t bother. They’ll serve the slop unfiltered to another party, who performs this interpretive labor themselves, only to discover that the underlying meaning does not exist . When someone expects me to perform their share of this interpretive labor, I get angry. Grice’s rules are, in essence, rules of etiquette —their unwarranted violation is rude . The loss of intent is acutely felt in art. Art demands the same interpretive labor from its audience, who might scan for hidden meanings in a film’s scenery or re-parse a winding passage by Virginia Woolf. This work reconstructs the intended meaning behind numerous artistic decisions. When the audience discovers that the meaning is entirely absent, they feel betrayed. Where we chose to engage based on markers of effort—a high word count or a surprising metaphor—these same signatures have come to indicate lack of effort and overreliance on tools. The strange poetry of literary classics looks, at a glance, not so different from the cartoonishly ungrounded metaphors a smile like “sunrise over a sink” in the likely-synthetic winner of the 2026 Commonwealth Short Story Prize https://www.newyorker.com/books/page-turner/did-a-chatbot-write-a-prize-winning-story-does-it-matter . Increasingly, art comes to feel pointless—at least, art that demands something of its audience. In 2004, a Bush adviser mocked the “reality-based community”. Empires, he explained, create their own reality. In 2026, this view feels almost quaint: the ability to shape reality is no longer a privilege of empires. Each individual now has the right to their own reality. Your preferred reality is algorithmically curated from a menu of bot swarms; it is then trimmed by selectively applying justifiable skepticism. You can dismiss inconvenient evidence as a deepfake, an encroachment from the Uncanny Valley. Your world is populated by synthetic neighbors who agree with you and peppered with implausibly evil synthetic enemies. Troublesome opponents and embarrassing allies are dismissed as bots. It’s easy to believe you really won an election if you believe your opponents are synthetic. And so our consensus reality has fragmented into so many little worlds. That consensus reality was the invisible infrastructure of democracy. Life on the Precipice breeds paranoia, and paranoia makes you easy to manipulate. They have you coming and going. Trust, and you might fall into the Uncanny at any moment. Distrust, and the Familiar becomes the Uncanny; reality ceases to exist. I am a routine user of LLMs. They simplify my searches and literature reviews. They write my code. I structured and planned this blog post in conversation with Claude Fable. It’s a pleasant way to work; I am constantly assured of the brilliance of my ideas. You’re absolutely right to address your own use as an AI researcher—it’s exactly what the audience is waiting for. When I open the chat window, I know that what I am talking to is not human, however fluent or congenial. This experience is localized to my session. It does not erode the Familiar. But it makes certain signs instantly recognizable, those famous em-dashes, the “load-bearing”, the “delve”. When I leave my session, I see these new cliches again in messages and paper abstracts. The ground erodes. It’s getting harder to tell. It’s not just that models are improving, or that users are better at prompting stylistic variety. My colleagues and students adopt the writing quirks they read throughout the day, and their own writing becomes more like an LLM’s. I get it: In a chat about my research on categorical model behaviors, Claude quoted Plato’s imperative to “carve nature at its joints”. I now use this phrase frequently. A sticky phrase can come from anywhere, and if you read enough from any single author, your tastes start to shift towards the new familiar. While the machines become more human, the humans also become more machine. It’s getting harder to identify the boundary, to find a solid hold. Where can we go from here? Only backwards. We retreat to the high ground of the Familiar. Traditional in-person encounters are safe for now, so we look locally. You’d better be well-connected in this new world. We hire only from our vouched networks. We ask our friends who’s single. We return to the office. This strategy will hold up until smart glasses become the norm, leaving us surrounded by humans parroting a stream of perfectly crafted answers. More than anything, we withhold. Surrounded by untested ground, we become miserly with communicative effort. Why reply to that email? Why incorporate that reviewer’s advice? Why try to understand? Any charity could precipitate another plunge into the Uncanny. Acknowledgements Thanks to everyone who talked with me about this grumpy manifesto. This post was shaped by conversation with Jacob Andreas and Michelle Li. Thanks to Sarah Wiegreffe, Jacob Andreas, Michelle Li, Eugene Vinitsky, Gin & Toxic, Sophie Hao, Angelica Chen, Matthew Leavitt, and Faye Wang for feedback and discussion on early drafts.