OnBodyAI — Why the Next AI Needs a Body is a single-author paper by Michael Beigl, posted to arXiv on 15 October. It is a write-up of his talk at the Industrial Panel at UbiComp/ISWC 2026 in Shanghai on 13 October, structured around that talk’s ten key statements.
It opens with a claim, not a result:
The next AI will not win by knowing more about the world. It will win because it knows you: because it lives on your body, works for you and answers to you.
OnBodyAI is the term Beigl proposes for what happens “when large language models are given a body, and what research and industry need to make that happen.”
Taking the three clauses separately
The sentence does a lot of work, and it is worth separating because the three parts have very different standing.
“It lives on your body.” This is the part that is largely a hardware and systems problem, and it is in progress now. On-device inference on power and thermal budgets that a body tolerates, continuous sensing, models small enough to run without a round trip. It is hard and it is tractable, and the evidence that the industry believes it is the volume of wearable silicon shipping.
“Works for you.” This is a claim about whose objective function. A wearable assistant optimising for engagement, or for a platform’s data interests, lives on your body without working for you. Nothing about the form factor delivers this clause; it is a product and governance decision that the hardware makes more consequential rather than more likely.
“Answers to you.” This is the hardest and it is where the paper’s framing is most valuable, because it is the clause almost no shipping product currently satisfies. Answering to someone means they can inspect what you did, correct it, and overrule it. A model that cannot explain why it surfaced something, and cannot be told to stop surfacing it, does not answer to the person wearing it regardless of where the weights live.
Beigl’s framing deserves credit for including the second and third clauses at all. The common version of this argument stops at the first and treats locality as if it automatically produced alignment with the wearer’s interests. It does not. Running locally buys you privacy from third parties; it does not by itself buy you accountability to the user.
Why this is worth reading as a position paper
Position papers get an unfair reputation. The value here is not evidence — there is none, by construction — but a crisp statement of a premise that an enormous amount of work is already assuming.
Look at what has shipped and been published in the last year: silent-speech wearables, gesture-sensing dev kits, neural haptic rings, diffusion models running bare-metal on microcontrollers. Each of those is an answer to a question nobody wrote down. Beigl is writing down the question.
That matters for a specific reason: when a premise is unstated, it cannot be argued with. Naming OnBodyAI makes the thesis falsifiable at the level of strategy rather than engineering. If the next generation of AI is in fact distinguished by capability rather than locality — if a sufficiently capable remote model with good latency makes on-body computation an implementation detail — then a lot of current hardware investment is answering the wrong question, and that is now a disagreement someone can hold.
The sceptical case, stated fairly
There is a decent counter-argument that the paper’s single-author, single-talk format does not engage.
Ubiquitous computing has made this prediction before. Weiser’s calm technology, context-aware computing, the first wave of wearables — each arrived with a strong claim that proximity to the body was the decisive variable, and each was substantially overtaken by developments that were about raw capability somewhere else. The smartphone did not win by being on your body in any ubicomp sense; it won by being a good enough general computer with a network connection.
The counter-counter-argument is that what changed is the sensing, not the computing. An LLM with continuous access to what you are looking at, saying, and doing has an input stream that no remote model can obtain, and that input stream is a genuine capability difference rather than a latency convenience. Whether that is decisive is the real question, and it is empirical rather than philosophical.
Beigl is associated with TECO, the pervasive-computing group at KIT’s Institute of Telematics, established in 1993 and working on wearable computing, embedded AI and human-centred sensing — relevant context for reading a paper that is explicitly an argument rather than a finding, from someone who has been making this argument’s components for a long time. Read it as the opening of a conversation at a conference panel, which is what it is.
Sources
- OnBodyAI - Why the Next AI Needs a Body — arXiv:2610.21137
- UbiComp/ISWC 2026
- TECO, Karlsruhe Institute of Technology
- arXiv listing, cs.HC