No Clean Slate

"Innovation-specific helplessness indicates that after a series of failed innovations, an employee becomes and remains passive even after a subsequent innovation is deemed successful." — an organizational variant of the learned helplessness Martin Seligman described as early as 1967: passivity acquired after repeated exposure to situations where one's actions had no bearing on the outcome.

The phrasing is precise and uncomfortable. It means that the intrinsic quality of the next initiative is not enough, on its own, to reverse the effects of past history. You can improve the tool, refine the rollout, pick the right moment — and still run into the passivity of employees who have already spent their psychological resources on previous attempts. This finding, documented in a multilevel sample of 481 Korean and Chinese employees, states in behavioral terms a thesis that the empirical literature on organizational trust confirms through several converging routes: trust in AI does not function like a checking account you can top up before each new initiative. It functions like a partially non-renewable resource, with asymmetric withdrawal rates — withdrawals happen faster, last longer, and cut deeper than equivalent deposits.[1]

Dominant management practice rests on the opposite assumption. Every AI rollout is presented as a fresh start: new vendor, new team, new internal communication, new project name. What this assumption ignores is that employees don't get a clean slate of their own. They arrive at the next initiative carrying an active cognitive schema, shaped by the previous ones, and that schema filters their perceptions of the new AI long before they have even assessed its merits. Understanding why this filter resists rational correction — and why standard repair mechanisms do almost nothing about it — is the subject of this essay.

The Ground Is Tilted Toward Distrust

The asymmetry principle Paul Slovic described in 1993, drawing on the management of technological risk, applies to any opaque technology whose users do not control its inner workings. His central observation is that trust "is typically created rather slowly, but it can be destroyed in an instant — by a single mishap or mistake"[2] — and that once destroyed, it may never return. But it is not just a matter of speed. It is a matter of shape: negative events and positive events do not leave the same trace. "Negative (trust-destroying) events often take the form of specific, well-defined incidents such as accidents, lies, discoveries of errors... Positive events, although sometimes visible, are more often fuzzy or indistinct."[2]

For LLMs, this asymmetry is structural. A hallucination in a client-facing document is a specific event: visible, memorable, tellable to others. The thousands of correct responses surrounding it stay diffuse — the user processed them, used them, but does not encode them as proof of reliability in the sense psychology measures. In Slovic's empirical study (N=103 participants rating 45 hypothetical events tied to a nuclear plant), trust-destroying events consistently score higher on impact than matched trust-building events. No positive event reaches the level of the most striking negative ones.[2]

This asymmetry is amplified for algorithms. Berkeley Dietvorst, Joseph Simmons, and Cade Massey documented it across five experimental studies (N≈2,367) under the name algorithm aversion: watching an algorithm make a mistake triggers a faster, deeper loss of trust than watching a human make the same mistake.[3] In the first study, 65% of participants with no prior exposure choose the algorithm for their forecasts. That rate drops to 26% among those who have watched it work — even when the algorithm outperformed the human on the same task. 83% of those participants had seen the model outperform the human and abandoned it anyway.[3] Superior performance does not protect trust: the observed error weighs more than the overall track record, and this bias holds robustly across five different domains. "As long as prediction errors are likely to occur — which is the case in virtually every forecasting task — people will be biased against algorithms."[3] The authors admit in 2015 that they have no recovery mechanism to offer those who have already watched the algorithm err.[3]

The Question Is Not Whether the Error Happens, But How It Is Encoded

Slovic's asymmetry describes a generic phenomenon. The literature on trust repair helps identify the mechanism that determines its intensity: depending on whether an error is perceived as a competence failure or an integrity violation, the available repair regime is radically different — and so are the consequences for whether trust can be recovered at all.

Kim, Ferrin, Cooper, and Dirks (2004) established this in two experimental studies (N=644 total). They start from a psychological principle: competence-based judgments are forgiving — even an expert can have a bad day, and a single success confirms one can succeed. Integrity-based judgments are unforgiving: "only individuals who lack integrity behave in a dishonest manner"[4], and a single dishonest act is enough to confirm a stable trait. The authors illustrate the asymmetry: "hitting a home run once makes us a home run hitter in the eyes of others, even if we fail to hit one again. In contrast, embezzling money from a company once makes us an embezzler, even if we do not commit any other theft in the following days, months, or years."[4] And once the label is affixed: "the belief that an individual lacks integrity will be hard to disconfirm."[4]

The practical question for LLMs is therefore one of cognitive framing. Technically, hallucinations are competence failures: the model generates statistically probable tokens with no internal fact-checking mechanism. Psychologically, users encode them differently. The model did not hesitate. It did not flag its uncertainty. It asserted something false with exactly the same confidence as something true — behavior that matches precisely the profile Kim associates with a lack of integrity: an entity that knows (or should know) and asserts anyway. The qualitative data collected by Vuori, Burkhard, and Pitkäranta (2025) in their longitudinal case study at TechCo (~600 employees, software sector, Scandinavia) confirm this: employees who lost trust in their AI tool do not describe a broken tool — they describe a surveillance instrument. "I find it dangerous. My fear is that the data will be misused. It's used against you in some cases."[7]

This is not a judgment on the algorithm's competence — it is an attribution of malicious intent. Kim predicts exactly what Vuori then observes: workshops, FAQs, algorithmic transparency — every form of excuse and explanation — repaired nothing.

The integrity-violation framing activates a regime that Schweitzer, Hershey, and Bradlow (2006) characterized in their study of trust and deception: "Trust damaged by unreliable behavior can be repaired by a subsequent series of trustworthy actions. Trust damaged by the same untrustworthy behavior plus deception, however, does not recover — even after apology, promise, and a subsequent series of trustworthy actions."[5] In their protocol (N=262, repeated trust game over seven rounds, 2×2×2 design), the best post-deception repair scenario — promise, plus apology, plus five rounds of trustworthy behavior — produces final trust lower than the baseline of the group with no deception and no repair whatsoever.[5] The mechanism is attributional: deception "clarifies that the untrustworthy actions were intentional"[5], which retroactively transforms the interpretation of everything that came before.

AI improvement roadmaps, iterative deployment plans, and progress communications — standard strategies of AI change management — are all forms of "promise + trustworthy actions." Schweitzer shows they are insufficient once the integrity regime has been activated, even without the aggravating factor of deliberate deception. In the AI context, they are insufficient by construction: the user who has reframed hallucinations as integrity violations is applying exactly the filter Kim and Schweitzer predict.

The Past Does Not Reset Between Initiatives

The mechanisms described so far operate within a single initiative. What structurally worsens the situation is that they compound between initiatives: the history of past deployments creates an active cognitive schema that precedes and filters the evaluation of the next initiative, regardless of that initiative's own quality.

Fiction has an image for this cost. In Malena Salazar Maciá's "The Forgetting Code," a grieving man pours his memories of his deceased daughter into an android's body, convinced he is bringing her back; each transfer erases a little more of his own memory of himself. Believing you can start over cleanly — wiping the slate to rewrite it — ignores that the erasure itself has a cost, a cost the eraser himself pays. The organization that announces "this time we're starting from zero" is making the same bet: it assumes a blank slate where participants carry, in their memory, the accumulated cost of previous attempts.

Bordia, Restubog, Jimmieson, and Irmer (2011) documented this in a longitudinal design: 124 employees of an institution tracked over two years following a restructuring. Their central construct is the Poor Change Management History (PCMH) schema: experiencing poor change management develops "a cognitive schema that captures the essence of this experience (e.g., 'this organization is poor at managing change'). All attitudinal and behavioral effects operate through this schema, because it affects the perception of future organizational events."[6] The study's most striking result: the cognitive schema measured at T1 predicts actual turnover two years later better than declared turnover intentions themselves.[6] Organizational memory of failure is more predictive of future behavior than conscious intentions. Slovic's formula takes on a concrete meaning here: "if trust is lacking, no form or process of communication will be satisfactory"[2] — this is not rhetorical hyperbole, it is the description of a schema that intercepts communications before they can change attitudes.

This temporal contamination also shows up at the individual level, with a particularly dangerous property: it persists even when the next initiative is objectively better. Chung, Choi, and Du (2017) showed that two dimensions of an organization's innovation history — intensity (the frequency of past initiatives) and quality (the failure rate) — independently produce innovation-specific helplessness: a loss of the sense of control over outcomes, with significant structural path coefficients on both dimensions in the SEM model (β intensity→helplessness=.44, β failure→helplessness=.31, p<.001 in both cases).[1] This helplessness translates into innovation fatigue — employees "simply avoid (often unconsciously) anything related to innovation because they are exhausted and depleted of the emotional and cognitive resources to deal with it."[1] The word "unconsciously" is decisive: the resistance is not a deliberate decision but an automatic exhaustion response. It is not reachable through rational arguments about the new initiative's quality.

One of Chung's additional findings is counterintuitive and worth drawing out: intensity alone — the frequency of past initiatives, even successful ones — is enough to generate helplessness if it is high enough.[1] Organizations that accelerate the pace of AI rollouts, even when those rollouts succeed, may deplete psychological capital employees have not had time to rebuild between initiatives. And the practical paradox the authors identify is especially striking: "an excessive emphasis on learning from past failures pushes employees to recall the failure of previous innovations, thereby triggering innovation-specific helplessness and fatigue."[1] Retrospectives, post-mortems, transparency reports — the very devices management practice recommends for learning from failure — activate precisely the schema they are meant to correct.

This is the cycle Vuori et al. document in real time at TechCo. The AI tool has 1,800 active users in November 2019. Eighteen months later, only 97 remain. "An organizational culture of trust and radical transparency does not guarantee high levels of cognitive and emotional trust toward AI."[7] TechCo does "everything it should" — a culture of psychological safety, workshops, FAQs, direct communication. The spiral turns anyway: the first bad results push some employees to opt out of data collection, which degrades the map for everyone, which pushes others to opt out too. "Bad outcomes were the main reason people did not come back to the service,"[7] a leader confirms. By 2021, the tool is dead.

The Condition Adoption Plans Ignore

In 2023, Esterwood and Robert ran the experiment that closes the question Dietvorst had left open in 2015. 240 participants worked with a robot co-worker in a virtual sorting task, exposed to three consecutive errors with intervening repair — apology, denial, explanation, or promise, depending on the assigned condition. The result is unambiguous: no repair strategy restored perceptions of the robot's competence and integrity to pre-violation levels.[8] Where the theoretical distinctions seemed promising — each strategy anchored in a distinct mechanism (forgiveness for apology, forgetting for promise, informing for explanation) — the empirical results converge on the same shortfall. After repeated violations, "apologies, explanations, and promises all appear equally ineffective at restoring overall trust, perceived competence, and perceived integrity."[8] Competence and integrity ratings — the two dimensions most critical to adopting a professional system — never returned to pre-violation levels, whatever the repair strategy used.[8]

This thesis has a serious opponent it would be unwise to ignore. Karunakaran, Vendraminelli, and Narayanan (2024) analyzed two sequential AI projects at the same firm — same team, first success, second failure — and explain the divergence through structural factors: how clearly competence domains are demarcated, how central the target task is to daily work, and how deeply the AI is enacted into existing practices.[9] The authors explicitly argue against a narrative of trust learning or spillover between sequential projects. This forces us to clarify the thesis's scope: it does not claim that trust capital is the sole determinant of AI success, nor that the next rollout is doomed simply because the previous one failed. It claims that this capital is a binding determinant organizations leave out of their calculations, and that the structural factors Karunakaran identifies are themselves harder to build once the available trust stock has been depleted. "Clarity of demarcation" between domains requires stakeholders willing to invest in defining a new boundary — and that willingness is precisely what a history of failures erodes.

What AI adoption plans fail to measure is the trust stock available at the entry point of each initiative. Standard indicators — adoption rates, declared usage intentions, satisfaction scores — are proxies for a capital that is silently consumed across successive deployments and failures. Chung et al. show that even successful initiatives deplete it if the pace is too fast. Bordia et al. show that the cognitive schema of organizational distrust persists for two years without dissolving. Esterwood and Robert show that the conventional repair arsenal is insufficient once violations have repeated. And Slovic summed it up three decades earlier, in a different context but for an identical mechanism: "if trust is lacking, no form or process of communication will be satisfactory."[2]

The question management practice refuses to ask is not "how do we make our next AI rollout succeed," but "how much trust stock is left to attempt it — and have we given it time to rebuild since the last failure?" Asking this question would mean treating organizational trust in AI as a resource to be managed, not a starting state to be assumed — and recognizing that some contexts have already, through an accumulation of failures, consumed what it would have taken for one more deployment to have a real chance of taking hold.

A single organization can thus accumulate both debts at once: a depleted trust stock and an unevenly distributed capacity for framing.

Answering it is a concrete operation, not a posture. Measuring that stock before launching — targeted surveys on the stories circulating, analysis of withdrawal signals, mapping of past experiences and their collective interpretation — and calibrating the pace of initiatives accordingly are not optional precautions: they are the minimum conditions for deployment investment not to be spent in a context where organizational trust has already decided the outcome, before the tool has even been opened.