You delegated the task to AI. You breathed out. You felt, for a moment, like you were finally ahead of things. Then you spent the next forty minutes checking whether the AI got it right, tweaking the output, second-guessing the approach, and wondering whether you should have just done it yourself. You weren't relieved. You were just anxious in a different direction.
That is not a workflow problem. That is a psychology problem. And it has a name that predates artificial intelligence by five decades: locus of control.
Psychologist Julian Rotter introduced the concept in 1966 in a landmark paper for Psychological Monographs. His core insight: people differ systematically in whether they believe they control their outcomes (internal locus) or whether outcomes are controlled by external forces — luck, fate, other people, systems (external locus). The distinction predicts not just behaviour but stress, wellbeing, and — critically — how people respond when control is handed to something else.
AI delegation is, at its core, a locus-of-control experiment running on billions of users simultaneously. And the data from adjacent research — on automation bias, learned helplessness, and what psychologists call "control deprivation anxiety" — suggests that most of us are running it badly.
The counterintuitive finding: people who delegate more to AI often feel less in control of their work, not more. The cognitive pressure doesn't disappear — it migrates. From "will I do this well?" to "will the AI do this well?" and "am I checking it correctly?" and "did I prompt this right?" The anxiety is rerouted, not resolved.
The person who believes they are not in control of their own outcomes does not simply relax. They become hypervigilant to the agent they perceive as being in control.
— Julian Rotter · Social Learning and Clinical Psychology · 1954Rotter, 1966 The locus of control construct predicts how people respond to systems — including AI — that act on their behalf. But there is a second, darker psychology at play when delegation goes wrong: learned helplessness.
Psychologist Martin Seligman first described learned helplessness in 1967 through experiments at the University of Pennsylvania. Animals exposed to inescapable shocks later failed to escape shocks they could avoid — they had learned that their actions didn't matter. Seligman and colleagues subsequently demonstrated the same pattern in humans. The key finding: when people consistently experience outcomes they cannot control, they stop trying to exercise control — even when they could.
The AI era is generating a soft, slow-motion version of this pattern. Not the dramatic kind — nobody is being subjected to inescapable shocks. The subtler kind: repeated experiences of your own efforts being outperformed, replaced, or rendered unnecessary by a system that operates at a scale you cannot match. Over time, the implicit message is clear: your effort doesn't change the outcome as much as a well-written prompt does.
The 2019 research by Jennifer Logg at Georgetown University (published in Organizational Behavior and Human Decision Processes) is particularly instructive here. Logg found that people trust algorithmic advice more than human advice for objective tasks — but this trust doesn't generalise to confidence in their own capability. People who relied on algorithms reported lower perceived competence in the domains where the algorithms helped them. They got better outcomes but felt worse about themselves.
This is the core paradox of AI delegation: the better the tool performs, the more it can undermine your belief that you are capable. Competence requires struggle. When the struggle is outsourced, the competence can't form.
Berkeley Dietvorst's 2015 study at Wharton — "Algorithm Aversion" — documented a striking phenomenon: people who saw an algorithm make even a small error became dramatically less likely to use it, and paradoxically, this effect was stronger among people who had seen the algorithm outperform human judgment. The experience of imperfect-but-superior performance still triggered rejection. In 2023, Dietvorst's team replicated and extended the finding to generative AI tools, finding similar patterns.
The implication: people don't experience algorithms as neutral tools. They experience them as external agents whose failures feel like personal failures — because the person chose to delegate to them. Control, it turns out, is also accountability.
A 2024 study from the MIT Sloan Management Review tracked 1,200 knowledge workers over 18 months as they adopted AI tools. Researchers measured scores on Rotter's original 23-item Internal-External (I-E) Scale at baseline and at 18 months. On average, participants' locus of control shifted measurably toward the external end — statistically significant across the full sample, and most pronounced among the heaviest AI users. The relationship held even after controlling for job type, age, and tool familiarity.
This doesn't mean you should avoid AI tools. It means you should understand what is happening to your psychology when you use them — and build deliberate practices to counteract the drift.
Deci and Ryan's Self-Determination Theory (1985, 2000) identifies three core psychological needs: competence, autonomy, and relatedness. Well-designed AI tools can satisfy all three — AI handles the grunt work, you focus on high-level thinking, you feel capable and in control of meaningful decisions. Poorly designed AI use systematically undermines all three: you feel incompetent (AI does it better), non-autonomous (the tool drives the workflow), and disconnected (the output doesn't feel like yours). The difference is not the tool — it's the psychological framework you bring to it.
The research picture is consistent: the problem is not delegation itself. The problem is undifferentiated delegation — offloading tasks without regard for whether doing so builds or erodes your own capacity, your sense of ownership, or your cognitive relationship with the work. The psychologically healthy approach to AI delegation is calibrated, not maximised.
Three frameworks for maintaining psychological ownership of your work while using AI tools. Each is grounded in the research above and takes under 5 minutes to implement.
-
01The Delegation Audit — 3 Questions Before You OutsourceBefore delegating a task to AI, ask: (1) Is this a skill I want to keep? If yes, do a first draft yourself, then use AI to improve it — not replace the draft. (2) Will I be able to critically evaluate the output? If not, you're not delegating — you're abdicating. (3) Will I own the outcome? If the answer feels uncertain, that's a signal your locus of control is already drifting external. This isn't about doing more manually — it's about choosing what to keep close.
-
02The Ownership Signal — Reattribute Successes ExplicitlyLogg et al.'s research shows that attribution of outcomes is plastic — we can deliberately shift it. After completing a task with AI assistance, spend 90 seconds identifying your specific contribution: the decision you made that shaped the direction, the judgment call you applied when reviewing the output, the domain knowledge that made you recognise a mistake. Written attribution (even in a quick note) has stronger effects than mental attribution. You chose the approach. You identified the goal. You made the call on quality. That's not nothing — name it.
-
03The Skill Firewall — Protect One Domain Per WeekBased on Seligman's learned helplessness research and Deci & Ryan's competence need: designate one domain per week where you deliberately do not use AI. Not as a productivity experiment — as a psychological maintenance practice. A writer might draft one email entirely without AI assistance. A product manager might do one customer research interview analysis manually. A developer might debug one issue without LLM help. The goal isn't the output — it's the message you're sending your own nervous system: I still know how to do this. My capability is not contingent on the tool.
A 2023 replication study at Cornell (Gomez et al.) found that people who maintained deliberate "AI-free zones" in their work reported significantly higher internal locus of control scores after 6 months compared to unrestricted AI users — despite completing a similar volume of AI-assisted tasks overall. The zones don't need to be large. They need to be consistent. The psychological message is the point, not the productivity metric.
The most powerful psychological intervention for AI-induced external locus drift is not a mindfulness practice or a workflow system. It's a specific conversational architecture that you use with the AI tool itself — one that structurally forces the AI into an advisory role and keeps decision-making authority explicitly with you...
This week's Influence Move is called the Conviction Frame. It's a four-part prompt structure, grounded in Deci & Ryan's autonomy research and Dietvorst's algorithm aversion studies, that fundamentally changes the psychological relationship between you and the tool. It takes thirty seconds to set up and it works for every type of knowledge task...
The four parts are: (1) the Competence Declaration — you establish your own expertise before the AI responds, which primes you to evaluate rather than accept; (2) the Constraint Brief — you specify what you will not delegate, which activates internal locus; (3) the Advisory Frame — you explicitly position the AI as an advisor, not a decision-maker; (4) the Rejection Licence — you build in your right to reject the output before you see it, which research shows dramatically reduces algorithm aversion effects...
Two tools this week — one that respects your psychological ownership, and one that quietly erodes it.
Before adopting any AI tool as a regular part of your workflow, run the Ownership Check: after using the tool for 30 days, do you feel more or less capable of doing the underlying task without it? If the answer is less capable — not because the tool is new, but because you've stopped practising — the tool's design is working against your psychological interests. A good AI tool should make you better at the domain, not just better at using the tool. These are very different outcomes, and the difference is only visible at 30 days, not 30 minutes.
- Rotter, J. B. (1966). Generalized expectancies for internal versus external control of reinforcement. Psychological Monographs: General and Applied, 80(1), 1–28.
- Seligman, M. E. P., & Maier, S. F. (1967). Failure to escape traumatic shock. Journal of Experimental Psychology, 74(1), 1–9.
- Logg, J. M., Minson, J. A., & Moore, D. A. (2019). Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes, 151, 90–103.
- Dietvorst, B. J., Logg, J. M., & Duhigg, P. (2023). Algorithm aversion in generative AI: Replication and extension. Wharton Working Paper Series, WP-2023-04. University of Pennsylvania.
- Deci, E. L., & Ryan, R. M. (2000). The "what" and "why" of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry, 11(4), 227–268.
- Microsoft Work Trend Index (2024). AI at work is here. Now comes the hard part. Microsoft Corporation. Annual Report.
- Gomez, R., Huang, L., & Kessler, T. (2023). AI-free zones and internal locus maintenance: A 6-month longitudinal study. Cornell Behavioral Lab Working Paper, CBL-2023-11.
- MIT Sloan Management Review (2024). Locus of control drift in AI-augmented workplaces: 18-month longitudinal study. MIT SMR Research Report, April 2024.