Originally published in my AI Minds column on UXmatters (October 5, 2026). Reproduced in full with permission.
Eight months into a UX research engagement with an inventory-management team, I reviewed usage data for an artificial-intelligence (AI) assisted–logistics-routing tool. One user consistently overrode system recommendations, took longer than his peers to accept recommendations, and described the AI as unreliable and narrow.
The interaction logs for his first week as a user explained why. During his third session using the tool, the AI had recommended the wrong route for a complex region that he had managed for four years. He caught the mistake, overrode it, and noted in the feedback form that the system had failed to account for regional-carrier agreements.
First Impressions Set the Standard
While the AI had improved substantially over the following months and, by the time I interviewed that user, had produced accurate recommendations in about 91 percent of comparable cases, the user still viewed the tool as unreliable.
The primacy effect helps explain the persistence of the user’s mistrust. Users consider their initial thoughts about interactions with an AI-powered product as a definitive assessment. During their early sessions, they decide what the AI understands, where it fails, and the amount of trust it deserves. Once users form that working model, confirmation bias helps preserve it, even when later evidence contradicts it.
AI onboarding can guide users in understanding when a system has earned their trust, presenting crucial insights for UX teams.
How Early Experiences Become a Lasting Judgment
The primacy effect leads people to weigh the information they encounter first more heavily than information they encounter later. Solomon Asch’s foundational research in the 1940s demonstrated this pattern of impression formation. Asch found that the early presentation of traits shaped how people interpreted subsequent information.
For example, consider a homeowner who hires a plumber for a home repair. On the plumber’s first visit, he accurately diagnoses a pipe leak in minutes and fixes it quickly at a fair price. Thus, the homeowner establishes a high trust prior. Months later, when the plumber recommends replacing an expensive pipe, the homeowner agrees immediately, assuming the work is necessary.
However, if the plumber had failed to fix a simple leak on the first visit, the homeowner would have established a low trust prior. Months later, even if the plumber provided accurate advice on a complex issue, the homeowner would assume that the advice was incorrect and seek a second opinion.
AI users follow a similar thought process. During the first weeks of use, they build a working model from a small and unrepresentative sample of system behavior. They encounter a limited range of tasks and system conditions and still need to learn how to work with the product. But those early interactions establish the lens through which they interpret the system’s later performance.
A user who decides in Week 1 that the AI performs reliably tends to view correct recommendations as expected and mistakes as exceptions. In contrast, a user who reaches the opposite conclusion tends to treat errors as confirmation and successes as either luck or evidence that the task was easy. The system’s measured and perceived performance diverge almost immediately.
Confirmation bias reinforces the user’s initial judgment. Users notice, interpret, and remember evidence that supports what they already believe. But users experience this filtering as accurate perception rather than bias.
This dynamic creates risk in both directions. A negative early impression drives hypervigilance or abandonment even after the system improves. A uniformly positive experience encourages overtrust and reduces appropriate user scrutiny. UX teams must actually aim for calibrated trust.
What Users Build in the First 30 Days
The conventional view treats onboarding as the period when users are learning the product. During AI onboarding, users build three psychological structures that persist past any tutorial or first use of a feature. In the first 30 days, users establish a trust prior, an oversight posture, and an identity frame.
A Trust Prior
Users establish a baseline level of trust, or trust prior, and bring it to every interaction. A high trust prior makes errors seem exceptional, while a low trust prior makes successes seem accidental. Both responses compromise users’ judgment. UX teams guide users to adjust their trust to the system’s demonstrated reliability and the stakes of their decisions.
An Oversight Posture
Users develop habits for checking an AI’s output. Calibrated users increase their scrutiny whenever consequences rise or their confidence falls. Other users check everything, eroding the AI’s promised efficiency, or stop reviewing outputs meaningfully because they no longer value the system.
An Identity Frame
Users decide whether they should act as the AI’s collaborator, supervisor, or passenger. When the system reflects users’ corrections and engages their professional judgment, they see themselves as active participants. However, when the system delivers outputs that users cannot shape, they adopt a passive role and disengage from oversight.
Four Design Principles for the First 30 Days
These design principles target the beliefs and habits that users form during their onboarding window, which extends past feature tours and progressive disclosure.
Engineer the first task. Design teams must select a representative and useful first task that falls within the AI’s stronger performance areas and carries manageable consequences if the system makes a mistake. For enterprise products, design teams partner with implementation leads in routing each role toward a suitable first task. For example, an inventory manager might start with a moderately complex configuration that showcases domain knowledge, but without putting a major shipment at risk. Design teams achieve this goal by introducing a system realistically through work that accurately represents both the system’s capabilities and limits.
Surface uncertainty before users discover it. Flagging uncertainty demonstrates the system’s boundaries to users. For example, the system might state: “I have lower confidence in this recommendation because the region requires custom carrier terms that do not appear in similar deals. Review those terms before accepting the recommendation.” A well-handled early limitation results in better calibration of trust than weeks of seemingly error-free output. It also teaches users where the system needs help, gives their professional judgment a clear role, and reduces their shock when the AI eventually fails.
Create early wins that matter for the user’s role. Generic accuracy fails to create a useful trust prior. Users need evidence that the AI is helping them with relevant work. For example, financial analysts care when the system surfaces a pattern they would otherwise have missed, but they learn little from seeing it automate a task they consider trivial. Develop role-specific onboarding paths around tasks that combine strong AI performance with genuine professional relevance. Ask whether the win demonstrates utility to the user rather than focusing on whether the system simply completed the task.
Make the user’s influence visible. Show users that their actions matter. When users correct or override an output, confirm what the system has recorded it and explained what happens next. Avoid promising learning that the product cannot deliver. If the correction changes only the current output, say so. If it informs future recommendations, explain the scope of the decision.
Visible, truthful responsiveness supports a collaborative identity frame that tells users the system requires their professional judgment and gives them a reason to remain engaged in oversight.
What UX Researchers Should Look For
UX research that is conducted during the first month measures user satisfaction, usability, and the psychological models that users form. The following practical, longitudinal approach to UX research connects what users encounter early with what they do later.
Week 1: Capture the initial model. Ask what tasks users think the AI handles well, what makes them distrust the AI, and how they would respond to a consequential mistake, revealing their early beliefs, perceived boundaries, and anticipated oversight.
Weeks 2–4: Observe updating. Repeat key questions and review users’ decisions after successes, errors, and uncertainty signals, revealing whether this evidence changes their working model or confirmation bias protects it.
Day 30: Diagnose the relationship. Interview users about their role relative to the AI and review behavioral patterns to reveal collaborative, supervisory, passive, hypervigilant, or disengaged postures
Day 90: Connect early exposure to adoption. Compare first-week errors and task types with their later use, overrides, review times, and abandonment to reveal leading indicators that product teams should address during onboarding.
Researchers interpret behavioral data. Low override rates and rapid approvals could signal confidence, disengagement, or automation bias. Correlate these patterns with early error exposure, task difficulty, and interview data before labeling them as successes.
Listen closely to how users describe errors and their own agency. A phrase like “the AI usually gets this right” would frame a mistake as an exception. In contrast, a statement such as “That is the AI being the AI” would frame a mistake as confirmation that the AI is unreliable. Active language such as “I guide the AI” suggests collaboration, while passive language like “It gives me an answer” signals a weaker sense of influence. Treat these phrases as clues, then verify them through observed behavior.
Track a small set of onboarding measures over time, including the following:
error exposure by task type
override and acceptance rates
time spent reviewing
use of uncertainty details
return frequency
These measures help product teams evaluate whether onboarding produces calibrated use rather than simple adoption.
When the First Month Goes Wrong
When the user’s early experience goes poorly, design teams must proactively repair the user’s mental model by presenting clear, unmistakable evidence that directly challenges their previous negative perception. Simply silently deploying underlying model updates or backend algorithmic improvements is ineffective because confirmation bias leads users to overlook changes and continue perceiving the system through the lens of its past failures.
To successfully reverse a poor impression, design teams should execute the following deliberate, four-step recovery process:
Show what changed. Explicitly highlight system updates and bug fixes directly within the user interface, making performance improvements immediately noticeable to users rather than subtle.
Acknowledge past limitations. Transparently own up to the specific flaw, edge case, or system limitation that triggered the user’s initial bad experience, helping rebuild trust through honesty.
Invite targeted retesting. Encourage users to test the upgraded system on a meaningful, high-value task within their actual workflow so they can experience the improvement firsthand.
Provide clear operational boundaries. To set accurate user expectations moving forward, clearly display control mechanisms and make the system’s current capabilities and ongoing boundaries visually prominent.
While repairing trust requires significant deliberate effort, effective product teams view an early negative judgment not as a permanent failure, but as a temporary, repairable state that they can overcome through transparent communication and visible proof of progress.
The First Month as a Design Priority
AI product teams often invest heavily in the core workflow, model accuracy, and feature roadmap while treating onboarding as a tutorial layer. User psychology suggests that a different priority might be more effective. The first month shapes how users interpret every later improvement and mistake.
A highly accurate system underperforms a less accurate one when the first interactions create distrust or overconfidence and the second interactions support calibrated use. The human–AI system succeeds only when users apply the right level of scrutiny, understand the system’s relevant limits, and engage their professional judgment when consequences arise.
UX designers treat AI onboarding primarily as belief formation. UX researchers study the first 30 days as a longitudinal experience, connect early events to later behaviors, and help product teams identify the interactions that disproportionately shape user trust.
Evaluating established users reveals an AI product’s current status, while studying new users indicates the future direction of the human–AI relationship. Few investments offer equal power to UX research for deliberately designing that direction.
References
Solomon E. Asch. “Forming Impressions of Personality.” The Journal of Abnormal and Social Psychology, 1946. Accessed September 15, 2026.
Bennet B. Murdock. “The Serial Position Effect of Free Recall.” Journal of Experimental Psychology, 1962. Accessed September 15, 2026.
Raymond S. Nickerson. “Confirmation Bias: A Ubiquitous Phenomenon in Many Guises.” Review of General Psychology, 1998. Accessed September 15, 2026.
