Is it rational to stop using AI after observing its errors?

Dietvorst et al. experimentally demonstrated conditions in which people avoid algorithms more than predictions from equally imperfect humans after observing the algorithm's prediction errors.[1] This is a phenomenon where error visualization asymmetrically changes usage choices.

This result does not mean that people who avoid algorithms are always irrational. When there are losses, accountability, distribution shifts, or when the user's own information is not included in the model, the decision not to use can be rational.

Before measuring avoidance, we classify the errors observed by users. The meaning of trust updating differs depending on whether the error is a minor numerical difference, a major failure, insufficient input, out-of-scope for the model, or operational error.

We also verify whether human errors and AI errors are presented under the same conditions. If only AI is shown a large amount of history while humans are shown only one success case, it is not a fair comparison.

When the usage rate drops, we examine changes in quality, explanation, remedy, and role before labeling it as resistance. We ask for the cause of the behavior and cross-check it with performance records.

• Error type: Distinguish between minor, major, out-of-scope, and operational failure
• Comparison conditions: Show the same history and criteria to both humans and AI
• Loss: Observe the impact of both misuse and non-use
• Reason: Link the basis for use and avoidance to results

Logg et al. also demonstrated algorithm appreciation, where people weigh algorithmic advice more heavily than human judgment under certain conditions.[3] The view that avoidance is a universal human response does not align with the entire body of evidence.

Avoidance and evaluation vary by task, expertise, judgment of self versus others, and algorithm explanation.[1] [3] We do not fixate on either "people dislike AI" or "people blindly trust AI."

Users may view AI advice lightly in their own area of expertise while recommending AI for others' judgments. We evaluate personal use, manager-assigned allocation, and advice for customers separately.

We measure not only the percentage of advice adopted, but also the adoption of correct advice and rejection of incorrect advice. A high adoption rate is not calibrated if it remains high even when AI makes mistakes.

We record users' pre-existing attitudes and separate them from post-implementation experience. If expectations from promotion or peer evaluation mix with actual observed performance, we misjudge the effect of the intervention.

Subsequent research by Dietvorst et al. showed that the option to slightly correct algorithmic predictions increases the intention to use.[2] This is a result where limited sense of control, rather than complete override, changes behavior.

Even if correction rights increase usage, it does not necessarily mean the final prediction accuracy improves. We compare whether human corrections are beneficial or merely satisfying against results.

We align the correction scope with business needs. We do not grant the same authority to corrections that add known customer circumstances as to those that move model values in any direction without basis.

We briefly record the reasons for corrections and cross-check them with results later. Using reason records for punishment obscures true overrides, so we separate the purposes of learning and evaluation.

If correction rights are merely formal and automatically revert to the AI value, trust is damaged. We clarify who the final decision-maker is and under what conditions overrides are maintained.

• Supplement: Add field information not in the model
• Correction: Clarify the acceptable range and basis
• Rejection: Retain the option not to use AI in cases of major concern
• Re-review: Return results and reasons to learning later

With optional use only, only those proficient in it will use it, biasing effectiveness comparisons toward self-selection. With mandatory use, there is a risk of increased compliance even out-of-scope and superficial usage.

In phased implementation, we first present advice and comparative performance, recording adoption, correction, and rejection. Next, we make AI advice the default only for limited tasks that meet the minimum, while preserving justified refusal, leaving high-impact cases to human decision.

We categorize reasons for non-use into rejection, lack of time, insufficient quality, unclear usage, and anxiety about authority. We do not collectively label those who do not use as resisters.

Even when mandating use, there is a method to take a provisional judgment before seeing the AI output. This reveals the difference between independent judgment and AI advice, reducing mere ratification.

Directly linking usage rate targets to individual evaluation makes safe rejection disadvantageous. We evaluate outcomes and appropriate judgment, not using AI as a proxy target.

Average accuracy alone does not allow users to judge whether they should use it for their own cases. We present performance by business, difficulty, input completeness, and time point, clearly defining out-of-scope.

In comparing humans and AI, we use the same set of cases, the same information, and the same time constraints. If humans can obtain additional information but do not pass it to the AI, we treat it as an improvement problem for the combined system.

We do not hide failure cases, showing how they were discovered and what was changed. However, we list frequency and impact together to avoid distorting the whole by emphasizing only rare dramatic failures.

Explanations prioritize input, output, reference basis, known limitations, and past performance over plausible internal model reasons. We provide information that users can re-verify.

After display, we conduct understanding checks to measure if users can answer the correct usage boundaries. We do not consider trust calibration complete simply by posting information.

In tasks where prediction or advice results become known later, we link adoption, correction, and rejection to final results. Users can learn the performance of both their own judgment and the AI.

When results are delayed, we do not substitute short-term satisfaction for accuracy. We separate intermediate indicators from confirmed results and do not count uncertain cases as successes.

Sharing only cases where humans exceeded AI, or only cases where AI exceeded humans, politicizes trust. We report representative examples and failure distributions in the same format.

If an override was correct, we return the effective field information to model inputs or procedures. If it was incorrect, we improve judgment conditions rather than blaming individuals.

If feedback is used for individual evaluation or sanctions, override reasons will be suppressed. We define the boundary, access rights, and objection processes between learning records and personnel evaluation.

The balance between avoidance and evaluation changes due to model updates, business changes, and user experience. We do not determine permanent measures based solely on pre-implementation attitude surveys.

In monitoring, we observe usage by business, correct adoption, incorrect adoption, correction, rejection, results, and reasons for non-use. We do not consider increased usage alone as success.

If usage is low in specific groups, we examine ability differences, insufficient explanation, access barriers, and past failure experiences. We do not attempt to solve this solely with attitude-correcting training.

If increasing correction rights lowers accuracy, we review the acquisition of necessary information and feedback before revoking rights. For categories that still cannot meet the lower bound, we reduce the scope of AI use.

The answer to algorithm avoidance is not to make people believe, but to show performance under the same conditions for humans and AI, provide limited control rights, and create a mechanism to update usage boundaries from results.

If users are given correction rights, we do not leave ambiguous who is responsible for the final result. We record roles such as AI proposing, humans changing, and the organization approving, ensuring responsibility is not returned solely to individuals after failure.

If we narrow the correction scope, we explain whether it is due to technical constraints or control reasons. Instead of allowing free input, we prepare alternative channels for additional information, exception requests, and human re-review, avoiding the mere appearance of control.

We do not treat a user's failure to correct as automatic consent. We investigate the possibility of ratification due to time pressure, screen design, or supervisor expectations. For major cases, we require explicit confirmation and basis verification separately.

In result reviews, we compare where value was generated among the AI draft, human correction, and final result. If humans always apply the same correction, we improve inputs or models; if corrections are random, we review rights and education.

Transparent correction rights are not psychological tricks to make people use AI. They are designed as formal controls to safely integrate field information, stop errors, and allow the organization to learn from judgment reasons.

The override ledger separates workflow, model version, and user group from decision criteria: maxima for severe errors and suppressed correction and a minimum quality after appropriate overrides; override reasons remain categorical. A workflow that exceeds either maximum or misses post-override quality stops forced automation and returns final judgment to people. Restart requires repaired correction rights and explanation paths plus joint approval by the model-risk and workflow owners of a passing group-level calibration test.