The AI Dependency Paradox: You Cannot Stop Using What You Cannot Trust

I caught myself doing something strange last week. ChatGPT gave me a citation that did not exist. I noticed the hallucination, corrected it manually, and then reopened the same chat five minutes later to ask for help structuring a paragraph. I did not trust the tool, but I also could not stop using it. That gap between distrust and continued use is what I keep thinking about, because it breaks the standard story we tell in information systems about why people stick with technology.

Bhattacherjee (2001) built the canonical account of this story. His expectation-confirmation model of IS continuance says that after initial adoption, a user compares post-use experience against prior expectations. If the experience confirms those expectations, satisfaction rises, and that satisfaction, together with perceived usefulness, drives the intention to keep using the system. The model is elegant and empirically robust, and I have cited it in exam answers more times than I can count. But the model assumes a kind of rational exit: if expectations are disconfirmed and satisfaction drops, the user is free to leave. That assumption no longer fits the world I am living in.

What happens when the system disappoints you and you stay anyway? The AI literature offers one side of the answer through algorithm aversion. Jussupow et al. (2024) clarified that algorithm aversion is best understood as a preference for humans over algorithms in decision-making, and they showed that the empirical landscape is fragmented because researchers measure the construct differently across contexts. Some people reject AI after seeing it err. Others appreciate algorithms precisely because they expect humans to be worse. The point is that user responses to AI are not uniform; they oscillate between aversion and appreciation depending on task, transparency, and prior experience. That oscillation does not automatically lead to abandonment.

Lee and See (2004) gave us the trust calibration triangle: overtrust leads to misuse, undertrust leads to disuse, and the goal is calibrated trust matched to actual system capability. AI systems today seem to invite both errors at once. We overtrust them for low-stakes drafting and undertrust them for anything important, but we rarely disengage entirely. The reason is not inside the trust model. It is outside it, in the costs of leaving.

Wessel et al. (2025) examined generative AI in digital platforms and found that hyper-personalization strengthens user engagement and platform lock-in through increasingly tailored interactions. They noted that switching costs become not only technical or economic but also psychological. Users become dependent on content that competitors cannot replicate because the personalized history is trapped inside the platform. Wade and Hulland (2004) identified customer switching costs as one of the five key drivers of IS strategy, and that logic only intensifies when the AI knows your tone, your project history, and the way you like your summaries formatted.

These switching costs interact with a post-adoptive reality that plain-vanilla expectation-confirmation models were never designed to capture. Koh et al. (2010), working on mandatory use contexts, pointed out that intention becomes irrelevant when users must use a system whether they want to or not. AI is rarely mandatory in the formal sense, but it is becoming structurally obligatory. Once your team builds workflows around Copilot outputs, once your writing style is shaped by autocomplete suggestions, the cost of extraction is not just relearning a tool. It is rebuilding a practice.

I think this is why the concept of technology dependency matters more now than a decade ago. In my study-hub notes I have a public-sector AI implementation paper that lists a direct tension among its findings: balance technology dependency and knowledge diffusion. That tension is real. Organizations want the efficiency of AI, but they do not want their employees to forget how to think without it. The individual version of the same tension is what I felt when I reopened ChatGPT after the hallucination. I had become procedurally entangled.

The sunk cost fallacy is part of the mechanism too. I have spent hundreds of hours refining prompts, building custom GPT instructions, and training the model on my voice. Walking away from that investment feels wasteful even when the current interaction is disappointing. The rational choice would be to evaluate each tool on its marginal utility, but humans do not work like that. We protect prior investments. We build routines. We lose the metacognitive distance to ask whether the tool is still worth it.

Continued use intention, in the original Bhattacherjee formulation, was a function of satisfaction and perceived usefulness. In the AI era, I think we need a post-adoption model that adds withdrawal costs, procedural entanglement, and psychological lock-in as direct predictors of staying. The expectation-confirmation mechanism still operates: I am less satisfied after a bad answer. But satisfaction is no longer the dominant path to behavior. Dependency is. I wrote earlier about why trust calibration for AI is not the same as for humans and about how we need to stop counting users and start measuring delegation. This post adds a third layer: even the measurement of delegation assumes the user could choose to stop, and that assumption is dissolving.

Baird and Maruping (2021) argued that agentic IS artifacts require a shift from the language of use to the language of delegation, with appraisal, distribution, and coordination as the three core mechanisms. Delegation presupposes a moment of choice where the human decides whether to hand control to the system. What I am describing is the phase after that choice has been made so many times that it no longer feels like a choice. The system has become infrastructure. You do not evaluate whether to delegate each morning; you simply operate inside the arrangement.

If I were advising a company on AI strategy right now, I would tell them to map not only adoption barriers but also exit barriers. What happens if you stop paying for the platform? How much of your institutional memory lives inside its prompt history? How many of your workflows assume its presence? Vendor lock-in is a strategy, not a problem, and generative AI makes that strategy almost invisible because the lock-in is emotional and cognitive rather than contractual.

Expectation-confirmation theory explains why people stay with systems that work. It does not explain why people stay with systems that betray them. For that, we need the darker vocabulary of dependency, switching costs, and sunk commitments. I am still using the tool that gave me the fake citation. I wish I could say I have a clean theory for why.