A plumber and a painting
The front right headlight blew out on my son’s car. Seemed like a good time to replace both with some fancy LEDs. Looked up YouTube videos for a DIY and the front right seemed easy. The left, not as much. It involved tricky maneuvers I wasn’t willing to undertake. So we’ve got one LED and one halogen. Life goes on. We’re all making such tradeoffs every day. There are things you call a professional for and things you don’t. If you’ve got a burst pipe, you call the plumber. A picture that needs hanging? Easy to do it yourself. Not because you couldn’t call someone. Because the risk doesn’t justify surrendering the job.
This isn’t laziness or pride. It’s a calculation we make so instinctively, it barely registers as a decision at all. We make these decisions constantly, with tools, with systems, with other people. Whether to hold on or hand over.
This calculus sits at the heart of how human beings relate to automation. And most of the AI industry seems to be getting it wrong.
Control and the tyranny of freedom
The psychological research on this topic is quite rich, but centers around two levers: outcome risk and cognitive cost.
Outcome Risk: The concept of control by proxy is the idea that humans will intentionally forfeit direct control to another party when they believe the transfer can maximize outcomes. We do it rationally, for example, by deferring to your doctor because their expertise reduces your risk of a bad health outcome. We also do it irrationally, for example, by letting your buddy pick numbers in a lottery pool because they’ve had a streak of good luck.
Cognitive Cost: We don’t just surrender agency to manage outcome risk. We also do it to manage the cognitive load of staying in control. This is the tyranny of freedom as coined by Barry Schwartz. When you have too many choices, the act of choosing becomes its own burden. The science of decision fatigue is real and researchers have documented what they call ego depletion, a measurable drain on cognitive resources from sustained deliberation.
It’s exactly why I’d rather text a friend for a restaurant recommendation on a Friday night instead of exhausting myself by scrolling through Yelp.
When a system pulls on one or more of these levers, humans will hand over control willingly.
AI and agency
AI agents and multi-agent systems are these seemingly magical and intelligent automatons that browse the web, manage your calendar, draft your emails and execute multi-step workflows. They genuinely address the tyranny of freedom and reduce the entropy of having too many choices. They are compelling because the relief from cognitive load is real.
But there’s a catch. The agency you feel like you’re retaining is largely illusory.
Research by Caspar, Cleeremans and Haggard has shown that we lose our sense of agency when we’re not completely involved in the entire process: from cognition to execution. This means that when you delegate action to a proxy, your sense of agency is measurably reduced, even if you choose the outcome. Consider why this doesn’t bother us with the plumber. When you call one, the delegation is fully conscious and self-endorsed: you chose them, you understand what they’ll do, you remain accountable for the result. The loss of agency is bounded. With AI, the handoff is ambiguous. You give access, set a goal, and the system fills everything in between. You often don’t know exactly what it did or why. The endorsement is thin, and so is the agency.
AI agents aren’t new in substance; in my opinion, they’re NLP-enabled IFTTT systems. What’s new is how convincingly they simulate judgment, which makes the handoff feel more meaningful than it is. A genuine agent would know when a situation has moved outside its parameters and a human needs to weigh in. Current systems don’t do this reliably. They confabulate. They fill gaps with plausible-sounding continuations and complete tasks in ways that look like completion without the underlying judgment that would make it safe.
Humans do the rest of the damage by silently diffusing responsibility.
When AI makes a bad call, the “I let the AI handle it” defense is seductive precisely because it moves ownership off the human. The cognitive load is lighter, but the accountability is quietly gone too.
How to build with AI
Managing agency with AI is a collaborative effort, with human-in-the-loop on consequential decisions. It means letting AI handle parts where the risk of error is low or recoverable, but ensuring it’s also designed to surface uncertainty.
There’s a user-side obligation too. Self-Determination Theory — Ryan and Deci’s theory of human motivation — makes an important distinction between autonomous regulation and heteronomous regulation. You can use AI and retain genuine agency, but only if you’re actively endorsing what it does as a reflection of your own values and intent (autonomy), not just accepting its output because correcting it feels like more effort than going along (heteronomy).
The moment you stop endorsing and start following blindly, you’ve crossed over from using a tool to being directed by one. You’re either a mechanical executor or a passive observer. Neither is being “in control.”
The car that drives itself
A LinkedIn post recently caught my attention — someone puzzled about why self-driving cars aren’t more widely adopted. It’s a fair question. And autonomous vehicles are a useful test case, because they make the calculus explicit.
For an elderly person who can no longer drive safely, or someone navigating high-risk terrain, or a person with a disability that makes driving difficult, the loss of agency involved in handing over to an autonomous system is worth it. The risk reduction is real. The return is meaningful: independence, safety, access.
The calculus looks different for a capable driver on a familiar commute. The marginal risk reduction is small. The loss of agency is real. The frequently cited benefits: “Reclaim your commute time, multitask, reduce stress!” are niche rather than universal, and often overstated. Most people’s commutes are not particularly dangerous. Handing the wheel over doesn’t dramatically change their outcomes.
Same technology. Completely different trade-off. Whether it makes sense depends entirely on whose agency is at stake and what they’re actually getting back.
This is what the industry tends to miss: adoption isn’t a marketing problem. It’s a calculus problem. People aren’t irrationally resistant to autonomous vehicles or AI agents or any other form of automation. They’re running the numbers, often unconsciously, and finding that the deal isn’t as good as advertised.
Your personal ROI
Not everyone runs the same calculus. Julian Rotter’s research on locus of control, the degree to which a person believes they can influence outcomes in their own life, shapes how much risk reduction someone needs before they’ll surrender agency.
People with a strong internal locus hold on longer. They need a higher return before the trade makes sense, not because they’re resistant to technology, but because they really believe their involvement changes the outcome.
People with a more external locus make the handoff more readily. Sometimes that’s appropriate. Sometimes it becomes learned helplessness: a habit of surrender that persists even when the trade-off no longer holds.
Neither is inherently right. But knowing where you fall changes how honestly you can evaluate any system asking you to hand over control, and whether the endorsement you’re offering is genuine or just the path of least resistance.
Rabbit Holes
The principles in Schwartz’s The Paradox of Choice aren’t new. It gets used heavily in marketing research and decision science. It’s not just the number of choices, but the experience of choosing that has costs.
The control by proxy research has a ton of references. The Milgram experiments are the obvious ones, but the more recent work by Caspar, Cleeremans and Haggard on reduced sense of agency under instruction is worth browsing. While it reinforces Milgram, it also highlights the diffusion of responsibility as an unwanted outcome of deferral.
Rotter’s locus of control is a simple framework that shows up in multiple areas of research: from organizational behavior to personal and mental health. Sitting with your favorite chatbot on this one is sure to keep you going down multiple rabbit holes.