Grounding

The intensity of online discourse about AI has been rather extreme lately, with some big numbers being thrown around, often without a lot in the way of explanation, evidence, or justification.¹

It occurred to me during this time that the average adult is plausibly being affected more by news about AI these days than by generative AI itself. Like many others, I’ve certainly received my share of questions from friends and family asking what’s happening, and how worried we should be.

Personally, I don’t feel like it is especially helpful for me to jump into those debates here, in part because I’m not convinced that it will do any good. Many other people seem to be happy to engage publicly on this topic, and it seems like they mostly suffer a lot of vitriol for their efforts.

Instead, I think it’s interesting to briefly reflect here on how much of our experience with respect to generative AI (or agentic AI, or whatever you want to call it), remains in the virtual world.² Huge numbers of people are now using LLMs and other tools to generate images, send emails, write code, summarize documents, solve math problems, ask for advice, engage in conversations, study all manner of subjects, and on and on, but the vast majority of these interactions seem to remain broadly in the form of electronic messages or code.

The cases that involve a stronger connection into the real world are interesting. If we ask for a recipe, and then cook and eat it, that is a kind of acting on the advice of a model that may well have a very real physical effect. The same goes for planning a vacation, or generating a bedtime story to tell to our kids, or asking for medical advice, but in all these cases, we remain as a kind of active intermediary, taking in the information, and taking actions which might depend on it. Critically, in all these cases, we retain a high level of agency. In many cases, we might choose to rewrite the story, or modify the recipe, or decline the advice.

The cases in which modern AI systems do play a more active physical role seem relatively rare. Waymo is one of the most obvious exceptions, and indeed, I think that is part of the reason it seems so magical. Robots are of course already all around us, but Waymo is closest to a long imagined science fiction form of embodied AI, with a familiar form (a car), carrying out a complicated, flexible task—suddenly navigating a city almost entirely without human intervention.³

Drone warfare is the other obvious (and much more terrifying) case, but thankfully that is not something that many of us have to deal with. As with most such examples, it is hard to avoid conflating various aspects of technology, only some of which involve AI. If we were being attacked by a military drone (or simply annoyed by the noise of a delivery drone), it might not make that much difference to us whether it was being directed, in that moment, by a human or by code. But there is no doubt that integration of sensors and actuators into weapons (as in cruise missiles, for example) changes the terms of warfare in a very material sense.

A very different case is all of the infrastructure on which we depend. Electricity, water, internet, and so on, not to mention food delivery to cities, medical care, transportation, etc.—these are all critical infrastructure without which most of us would be relatively helpless, and all are to some extent touched by software. For the moment, at least, these seem to be more like sites of potential vulnerability to disruption, rather than being actively modulated by generative AI systems. Nevertheless, that could change over time.

Finally, there have been many attempts over time to expand the visions of embodied AI, and to make them a reality. Home robotics, smart cities, personalized cybernetic implants, various forms of imagined virtual reality that would somehow have a physical aspect (such as the Holodeck on Star Trek)—all of these have been dreamt of and in many cases attempted, but so far have remained rather marginal in our world.

The desire to create these things is clearly there, and some of them do exist to some extent, but the blockers seem to be primarily a) cost, in the case of personal devices, and b) social factors, in the case more infrastructural visions.

That’s not to say that we won’t see such things scaled up in the coming years. Rather, it is to make the point that so much of our lives (and arguably the most important parts) remains in the physical realm, in a world that is largely, so far, untouched by anything close to what most people mean by AI.

In pushing back on this, one could very justly point out that many forms of nature that we engage with, such as lakes and forests and rivers and mountains, are themselves socially constructed, and in some ways technological. This is clearly true, but considering our physical experience of these things—the sound of the breeze in the trees, the cold of water or air on our skin, the smell of ocean air—it’s remarkable how primal and foundational and relatively unchanged these things are from the time of our distant ancestors.

Obviously we can’t practically live in a pure phenomenological mode all the time. Nor would I deny that extensive societal dependence on complicated software architecture creates systemic risks. But for those feeling the stress of the present moment, I do think it is remarkably helpful to step back, focus on the experience of one’s physical environment, and reflect on the continuity of the human experience of millions of years.