We've started talking about AI the way we talk about cars. "I'm driving ChatGPT." "I'm driving Claude." The phrase is doing more work than it looks like — it implies you get in when you need to go somewhere, you get out when you've arrived, and nothing happens to the vehicle while you're gone. That's the whole appeal of a car: it's inert without you. The question worth asking is whether that's actually true of the AI we're starting to talk about the same way, or whether we've borrowed the language of a tool we trust for a class of tool we haven't earned that trust from yet.
The parking problem
"Can I safely park it and come back later" sounds like one question, but it's really three. Does anything happen to my information while I'm not looking — is it safe in the privacy sense? Does anything happen in the world on my behalf while I'm away — is it safe in the agency sense, nothing acting without my say-so? And when I come back, do I get the same thing I left, or has it quietly drifted — is it safe in the continuity sense? A car satisfies all three by default: turn the key, the engine is off, nothing continues. Right now, most AI products can't make that claim as cleanly. Closing a tab isn't the same as something stopping. A session might still be running somewhere; a memory might still be getting written; something you set in motion last week might still fire tonight whether you're watching or not. That gap — between "I walked away" and "it actually stopped" — is probably the single most honest thing to say about where this technology is right now.
Cars have already been through this
It's tempting to treat the car as a stable reference point and AI as the thing racing ahead of it, but cars have already made almost exactly this transition. A car from thirty years ago really was a dumb machine: no memory, no connection, nothing happened to it while it sat in a garage. A modern connected car has software that updates itself overnight, a subscription behind some of its features, sometimes a connection back to the manufacturer the owner never explicitly authorized each time. "Driving the AI" isn't a metaphor borrowed from the old, simple version of a car. It's a prediction about where AI is headed, borrowed from the later, more complicated version of a car we've already lived through once.
That shift changes what the car key actually means, too. A physical key is a strange and honest piece of technology: whoever holds it has full authority, no fine print, no asterisk. There's no version of a car key that comes bundled with "and by holding this, you agree to let us study how you drive." Whatever the equivalent is for AI — a login, a subscription, a token — isn't like that. Holding it doesn't just grant access, it opens an ongoing relationship with whoever issued it, one whose terms can change on their end without your say.
But even that comparison needs a qualifier the romantic version of "a car just serves its owner" leaves out. No car has ever answered to its owner alone — seatbelt laws, emissions standards, speed limits all bind it regardless of what the owner wants, because what a car does can hurt people who aren't the owner. So the real bar was never "does it answer to nobody but me." It was "beyond the constraints everyone sharing the road has a legitimate claim to impose, does anything else have an unacknowledged say in what this does for me." That's the fairer question to ask of AI too, and it has an answer today, not just a forecast: look at how the thing is paid for. A subscription product's incentive runs close to the car-owner relationship — you pay, it works for you, the company does well when you're satisfied. An ad-supported or data-monetized one has a second customer in the room whose interest isn't yours, and that's not speculative, it's just the mechanics of the revenue, checkable by anyone before they trust a given product the way they'd trust a paid-for, privately owned car.
Where the concern does shade into projection is the more dramatic version of the question — an AI with sustained, independent initiative, loyal or disloyal to someone on its own ongoing authority, acting while nobody's watching the way the sections below imagine. That mostly isn't how deployed AI behaves yet; most of it waits to be asked. So this is worth holding as two separate claims rather than one: whose interest a given AI product serves is a live, answerable question today, through its business model. Whether it can act for or against you while you're not looking is mostly still a reasonable thing to get ahead of, not something already happening at scale. Collapsing those into a single verdict is how a piece like this stops being analysis and starts being projection — some of it onto a present that's already here, some of it onto a future that's still arriving.
And what actually travels with you from one AI product to the next isn't the vehicle, it's your own skill at using it — closer to a driver's license than to the car. The memory, the history, the accumulated context stays behind, locked to whichever one you were using, the way a car's own trip computer doesn't follow you to your next car.
The real mechanism underneath the metaphor
Strip away the imagery and both machines are doing the same basic thing: lowering the cost, in effort, of getting an outcome. A car burns fuel so you don't have to burn your own calories walking or hauling something. AI burns electricity so you don't have to burn your own cognitive effort thinking something through. Once you see it that way, the car analogy stops being decoration and starts being the actual argument: both exist to let a person spend less of their own finite energy on a given result.
For most of human history, that energy — literal calories — was the scarce thing driving decisions. Once a society gets good at producing calories cheaply, which most of the developed world now has, the scarcity doesn't disappear, it migrates onto whatever becomes the real bottleneck instead, which is money. Money absorbed the calorie problem and then kept absorbing what sat above it too — status, belonging, security increasingly route through what you can afford, not only through calories directly. Not entirely: plenty of status and belonging still get built outside money — a reputation in a hobby, an unpaid friendship, a volunteer's standing in a community that trades in nothing but trust. But money is the dominant channel in a market economy, even if it isn't the only one.
What that makes AI's actual value proposition
If money is the resource that's absorbed most of the motivational stack, then AI's significance isn't that it's clever — it's that it's one of the first technologies that meaningfully cheapens access to that layer for individuals specifically, the same way it cheapens cognitive effort generally. That's not a new right. People already had the legal standing to contest a bad contract, read a lease before signing it, or question a bill — they just rarely had the money to put a professional's time behind exercising it, while an enterprise always did. AI collapses that cost. It doesn't close the gap, because enterprises get the same tools too, often sooner and at greater scale — the floor rises more than the ceiling falls — but it's a real and underreported story: not new rights, just the restored ability to use the ones that were already on paper.
Push that further and you get to something sharper still: a wealthy person's real advantage was never just being able to afford one expert, it was being able to run several at once — a lawyer on one problem, an accountant on another, an assistant on a third, in parallel, while their own attention stayed free. A single AI conversation doesn't deliver that, it's still one assistant, used serially. What does is several agents working concurrently on a person's behalf — which is the actual payoff of AI's proliferation, not just that more AI products exist, but that an ordinary individual can now direct several of them at once the way a wealthy person directs staff.
That claim deserves the same scrutiny the rest of this does, though. A hired professional carries accountability an agent doesn't — a license, a malpractice risk, something at stake for being wrong. A wealthy person's staff also includes someone who integrates across the specialists, a role that currently still falls on the individual doing the directing. And someone still has to check the work, a job real staff come with a track record for and agents don't. None of that makes the claim false. It makes the honest version "a meaningfully larger share of that leverage, at a tiny fraction of the cost" rather than a full equivalent to a staff — which is still a genuinely large claim, it just doesn't need to overreach to be one.
The hinge
All of this, the parking question and the staff question both, comes down to the same test: can you hand it something, walk away, and trust what comes back when you return? That's what makes a thing staff rather than a tool you have to babysit, and it's what makes a car something you drive rather than something that's quietly still running while you're not looking. Most AI products haven't fully earned that trust yet — not because they aren't capable enough, but because the "off" switch, the accountability, and the verification that would make the parking problem and the staff problem both disappear don't fully exist yet. That's the actual frontier. Not how smart the thing is. Whether you can park it.
Comments
0 comments
Join the discussion
Sign in to leave a comment on this post.
Sign InNo comments yet. Start the conversation.