PTDoes easy AI make us learn the wrong lessons?
Sometimes experience teaches the wrong lesson: a person links an outcome to a cause that is not the real one. The thesis measured five signs of that risk in 126 professionals who use generative AI at work and looked at how they relate to job performance. Explained in 12 screens.
Height = average answer of the 126 professionals (1 = never, 7 = always) on each sign’s statements, our calculation from Table 6. Blue: signs about the person. Red: signs about the environment.
Learning wrong means taking a coincidence for a cause
The name comes from Skinner’s experiments in 1947: hungry pigeons got food at fixed intervals, whatever they did. Even so, they began to repeat whatever they had done just before, as if it brought the food.
The same happens at work. The thesis cites the doctor who credits a patient’s recovery to their own treatment, when it may have come from natural recovery or from other professionals, and keeps repeating it. The thesis is not about black cats or lucky numbers, but about these wrong links between what one did and what happened, which the literature calls superstitious learning.
Food drops every 5 seconds, whatever the pigeon does. Each gesture’s bar counts how many times food came right after it.
Since the error itself cannot be measured, the thesis measures five risk signs
Anyone who knew they had learned wrong would no longer be learning wrong. So the author measures the propensity for superstitious learning: three signs about the person and two about the setting in which they get feedback on what they did.
The five signs form the thesis model (Figure 4). Height = number of statements per sign in the final model. Click a block or a button. The statements come from the questionnaire, answered from 1 (never) to 7 (always), thinking about using generative AI at work.
It took three rounds of questionnaires to reach the final model
First, a literature review gathered what was known about the topic. Then two trial runs of the questionnaire showed what had to change. Click each step.
Step height marks the order of the stages, not their size. On top, the number of articles (review) or responses. The funnel is from Figure 1 (search in October 2024, across six databases).
126 professionals who use generative AI at work took part
The questionnaire circulated on LinkedIn, Facebook and Instagram, without asking for names. The invitation sought managers, specialists and entrepreneurs who use tools such as ChatGPT, Copilot and Gemini.
The slices add up to 100%: they split the purposes mentioned.
Answer five survey statements and compare with the average
Think about using generative AI tools at work or on a project. From 1 (never) to 7 (always), how often...
The red line is the average of the 126 (Table 6). On reflecting, the thesis uses the reversed answer to measure lack of reflection (average 2.2); the 5.8 shown is for the statement as written (our calculation: 8 minus 2.2).
The thesis bet on ten paths between AI, the environment and performance
Easy AI would bring overconfidence, a sense of control and less reflection; uncertain work, confusing and slow feedback; and all of it, worse performance. Bet on three first.
Six paths held up, one came out the other way and three had no effect
Each column is a hypothesis. Height is the strength of the link (β, from 0 to 1, in standard deviations). Switch the model: the numbers barely change with controls or with the questionnaire bias test.
People who find AI easy reflect more, not less
The thesis expected ease of use to lead people to accept the machine’s answers without thinking. Before you look: people who find generative AI easier to use...
Everything in standard deviations, from the average. Each bar = β from Table 13 × the chosen value (our calculation). It shows the direction and relative size of the links, not a forecast for a person.
Lack of reflection and confusing feedback weighed on performance, confidence did not
Move the five signs and watch the column for the performance professionals reported (meeting deadlines and commitments, taking initiative, adapting). Only two controls move the column; the other three showed no clear effect.
Illustration with the β from Table 13, in standard deviations from the average, added up (our calculation). The three signs with no clear effect do not change the column. The model explains 23.1% of the variation in performance.
When work changes fast, feedback gets more confusing and slower
Uncertainty was measured with statements such as “tasks changed quickly” and “managers’ and clients’ demands were hard to predict”. It was linked to both kinds of poor feedback, but only the confusing one reached performance.
Arrow thickness = β from Table 13; dashed = no clear effect. Values in standard deviations; the effect passing through confusing feedback is the product of the two β (our calculation, not tested separately in the thesis). illustration
What the research suggests to those who work with generative AI
Managers and teams
Set aside time to reflect before decisions and after results, good or bad, and hold regular moments of clear feedback with the team, within a culture that treats errors as a source of learning.
Organizations and HR
The author suggests a learning unit that analyzes successes and failures, records lessons and works with HR on assessments and training, especially for decision-makers.
People who use AI at work
For the author, knowing how to use the tools and, above all, how to analyze the answers they produce may be decisive for employability. The questionnaire works as a warning sign for the person themself.
Researchers
Apply the model within a single company or sector, compare hierarchical levels and generations, follow decisions over time and test psychological safety as a moderator.
What this research does not say
- Whether people actually learned wrong. The thesis measures risk signs, the propensity, not cases of superstitious learning, which only show up after a failure is analyzed. · according to the author
- That it applies to everyone. The sample is a convenience sample of 126 professionals recruited on social media. · according to the author
- Cause and effect. Everything was answered at a single point in time, and performance is self-reported. The author suggests measuring performance later, at a second point in time. · according to the author and our observation
- A number that depends on the method. The thesis p-values come from Kock’s method, adopted because bootstrapping was unstable. In the software diagram (Figure 8, bootstrap, no controls), the path from confusing feedback to performance has p 0.118. · our observation
Credits
KUSSANO, Walter Shuiti. Propensão à aprendizagem supersticiosa em contextos de percepção da facilidade de uso da inteligência artificial generativa e seus efeitos no desempenho individual no trabalho. Doctoral thesis (Business Administration), Mackenzie Presbyterian University, São Paulo, 2026. Advisor: Prof. Diógenes de Souza Bido.
Supported by MackPesquisa, Mackenzie’s research fund. Analyses in SmartPLS 4. The new feedback statements were generated with ChatGPT, Copilot and Gemini and reviewed by researchers and experts; the author used ChatGPT and Grammarly to revise the text. The two trials became articles (CASI 2024 and SemeAd 2025, Brazilian conferences). This version simplifies the work; the full tables are in the original.