The intelligence that remains

A mindset for using AI without surrendering judgement

Andrea Viliotti

An analytical article for entrepreneurs, executives and readers interested in the quality of human–AI interaction

ARTICLE THESIS The real competitive advantage is not how many answers a machine produces, but how much human capability remains — and grows — when the machine is not there.

21 June 2026

An analysis by Andrea Viliotti based on three empirical studies

THESIS

AI use can be considered transformative only when improved performance with assistance is accompanied by human control and responsibility, verification in the real world, and capability that remains after the assistance is removed. The most informative test is not how impressive the answer sounds, but what the individual and the organisation can still reconstruct, judge and do on their own.

How to use AI without losing human judgement

PERSPECTIVE

The wrong question we are asking

The debate on artificial intelligence is dominated by technological questions: which model to choose, which licence to buy, which processes to automate, how many minutes to save. These questions matter, but they are not the deepest ones. The decisive question is different: what kind of human beings and what kind of organisations are we becoming as we use AI?

A system can produce a brilliant answer while making the person receiving it less capable. It can accelerate a decision while impoverishing the understanding behind it. It can improve the output without improving judgement. It can even increase confidence precisely as the ability to verify declines. Digital maturity therefore cannot be measured solely by assisted productivity: it must also include the quality of thought that remains available after the assistance ends.

The perspective adopted in these pages treats human–AI interaction as a single process: the person, the tool and the real-world context influence one another. AI is therefore not an external oracle, but a tool that can reveal alternatives, connections and scenarios. The human being must remain in control of the reasoning: framing the problem, declaring the criteria, making the decision and answering for the consequences.

The criterion used in this article

Here, the most robust approach does not mean a universal formula already validated in every context. It means a method that, in the context at hand, improves performance without sacrificing human control and responsibility, quality of evidence, the ability to correct the decision, safety or the capacity to act without AI.

EVIDENCE

Three signals from research

1. Adoption begins with usefulness, but growth is not the same as use

How to read β. The study by Fabio Ibrahim, Johann-Christoph Münscher, Monika Daseking and Nils-Torge Telle, involving more than one thousand people with at least some previous experience of AI, uses the standardised coefficient β, read as “beta”. This number makes it possible to compare, within the same statistical model, the strength and direction of associations: a positive value indicates that two variables tend to rise together; the further the value is from zero, the stronger the relative association. β is not a percentage and, on its own, does not demonstrate cause and effect. What emerged. Perceived usefulness was the factor most strongly associated with a favourable attitude towards AI (β = 0.34), followed by the belief that AI can support the growth and development of one’s capabilities (β = 0.28). The belief that using AI does not erode skills — described by the authors as non-deskilling — showed a very small and statistically non-significant association (β = 0.03).

Study limitation. The research measures attitudes and intention to use, not independent capability after AI is removed; the largely well-educated, German-speaking sample limits automatic generalisation to other contexts. (Ibrahim et al., 2025, Table 1, pp. 4 and 7)

The managerial lesson is clear. Encouraging adoption takes more than imposing tools or promising that “no one will lose skills”. It requires visible usefulness, credible use cases and a framework in which people see AI as an opportunity for development. But a growth-oriented mindset chiefly predicts willingness to use AI: on its own, it does not show that capability has been retained. Adoption shows that a tool has been accepted; it does not show that people have learned to reason better without it.

2. A flattering response can seem more credible

Eunhae Lee’s MIT thesis studied 238 participants exposed to predetermined, fictitious predictions presented as if they had been generated by AI, astrology or personality analysis. Positive predictions were perceived as 36% more valid, 42% more personalised, 27% more reliable and 22% more useful than negative predictions. The study measures perceptions in a simulated task: not the truth of the predictions, nor an actual change in behaviour. (Lee, 2024, Table 3.2, p. 50)

This evidence touches on a sensitive issue in both business and personal life: linguistic fluency and personalisation can create an illusion of being accurately recognised. When AI reflects back the image of ourselves that we would like to see, we tend to confuse feeling understood with having been described correctly. A favourable response can be psychologically persuasive without being better grounded in fact.

3. Better with AI, without evidence of lasting skills

A longitudinal study — one that follows the same people at several points in time — by Anku Rani, Valdemar Danry, Paul Pu Liang, Andrew B. Lippman and Pattie Maes followed 67 participants for four weeks as they distinguished between true and false news items. With a persuasive chatbot, accuracy increased by an average of 21.3 percentage points. In week four, when participants assessed new items without AI immediately after the dialogue, accuracy was 15.3 points lower than in the initial week. Unaided performance measured before each new interaction had fallen by 6.8 points, but this second decline was not statistically conclusive: the 95% confidence interval — the range of values compatible with the data under the model — extended from −15.3 to +1.7 points and included zero. How to read p = 0.127. The p-value is an indicator used to assess whether a result is clear enough to be distinguished from variation in the sample. In this case it did not cross the conventional threshold used in the study for statistical significance; it does not mean that there was a 12.7% probability that the result was due to chance. (Rani et al., 2026, Sections 6.2–6.4, pp. 8–10)

The finding concerns a specific misinformation-discernment task, a small US and UK sample, four weeks of observation and no parallel longitudinal group without AI: it does not demonstrate permanent skill loss. It does, however, expose a general risk: assisted performance can improve without producing lasting independent improvement. The exploratory associations between some guided-questioning strategies and later outcomes do not prove that any particular method is causally superior in every context.

What the three studies show — and what they do not show

Perceiving AI as useful supports adoption. Favourable responses can bias judgement. Better results with AI do not guarantee capability without AI. Together, the studies define the problem and identify conditions that must be managed; they do not, on their own, demonstrate the causal superiority of the proposed method.

MODEL

The guiding principle: AI broadens the possibilities; people decide

A simple principle follows from these risks: AI should broaden the range of possibilities, not take over the decision. Its role is to propose alternatives, counterarguments, scenarios, sources to verify, competing hypotheses and questions the person had not yet formulated. The human task is to retain ownership of the criteria, choose, state the risk they are prepared to accept and remain answerable for the decision.

This distinction rests on four ideas: consider the person, the AI and the context in which the decision has effects, rather than only the generated text; separate immediate advantage from the learning that remains over time; alternate between exploring alternatives and deciding, instead of delegating continuously; and remember that a persuasive sentence is not yet a fact until it has been compared with sources, actions and observable results. This is neither a universal law nor a validated psychological scale: it is a practical framework to be tested in different contexts.

Test	Question	What would show it
Coherence of the response	Is the answer clear and internally coherent?	The text produced by AI
Human decision	Do the objective, criteria and responsibility genuinely belong to the individual or team?	A decision explained and defended in their own words
Real-world verification	What happened outside the conversation?	Independent sources, a pilot, consequences and feedback
Capability that remains	Can the person still reason or act when AI is removed?	Reconstruction of the reasoning and a similar case handled without AI

Immediate performance and learning are two different things

An intuitive example

Imagine someone who, without AI, gets 60 out of 100 cases right. With AI, they get 81 right: an immediate gain of 21 points.

Later, on new cases and without AI, they get 45 right. Compared with their starting point, independent capability has fallen by 15 points. It is therefore possible to perform better while the tool is present without having learned to perform better alone.

The pattern observed in the task studied by Rani et al. was similar: approximately +21.3 points with AI and −15.3 points on new cases without AI in week four. The four checks above cannot compensate for one another: three positive results do not demonstrate learning if the capability disappears when AI is removed.

PROTOCOL

Seven steps for a mature interaction

The proposed way of thinking is not an immutable psychological trait. It is a dynamic practice, to be used in proportion to what is at stake. For a purely operational task — formatting, translating or transcribing — a light process may be enough. When the task involves decisions, learning, identity, people or future capability, a full cycle is needed.

1	Start from your own position Before consulting AI, write down what you think, your objective, your criteria, your uncertainties and what you will not delegate. Without an explicit starting point, it is easy to mistake the first plausible answer for your own thinking.

2	Use AI to broaden the possibilities Ask for genuinely different alternatives, other ways of framing the problem, sources, scenarios, overlooked people or viewpoints, and the cost of each option. AI should add useful perspectives, not merely rewrite the same idea more elegantly.

3	Examine objections and trade-offs Do not rush to a synthesis. Record contradictions, missing evidence, costs and benefits, and the facts that would show each hypothesis is wrong. A good interaction leaves enough time to see what does not fit.

4	Make a human decision The individual or team states the criterion, makes the decision, explains the risk they are prepared to accept and establishes what would cause the case to be reopened. AI does not decide for the person who bears responsibility.

5	Test it in the real world Turn the decision into an action, a simulation or a reversible pilot. Gather sources and feedback independent of the conversation: quality, errors, consequences, reactions and data.

6	Check what remains without AI Reconstruct the reasoning without the chat or the transcript, or tackle a similar case. This is the test that distinguishes temporary help from capability that has genuinely been acquired.

7	Review the reasoning in light of the results Record the facts first; then use AI again to understand what worked, what created dependence and what needs correction. This allows real observations to guide the review without retrospectively rewriting what was thought at the start.

LEADERSHIP

At work: from productivity to organisational capability

For a business, the most common mistake is to treat AI as a simple output multiplier. If time saved is the only metric, the organisation may celebrate an operational gain while accumulating cognitive debt: fewer people can reconstruct the reasoning, check the sources, challenge a recommendation or take responsibility for a choice.

Good governance — a clear system of roles, controls and accountability — begins with one question: who is the human owner of the decision? It may be a manager, a team, a committee or an authorised function, but it must be identifiable. AI can support the analysis; it cannot become the invisible authority everyone follows because “the model said so”.

Five management practices

·  Separate the stage in which many options are generated — the divergent phase — from the stage in which one is selected — the convergent phase: AI proposes alternatives and objections; a person or human decision-making body decides.

·  Keep a concise record of sources, assumptions, rejected alternatives and decision criteria.

·  Prefer reversible pilots to immediate large-scale adoption, especially when the impact concerns customers, employees or reputation.

·  Include an AI-free step: the team must be able to explain the decision and reconstruct the reasoning without consulting the conversation transcript.

·  Protect dissent: a fluent AI response must not weaken the right and duty to challenge it.

This approach also changes the transformation dashboard. Alongside productivity and assisted quality, organisations should measure the ability to reconstruct the reasoning independently, source integrity, clarity of accountability, the ability to correct the pilot, the number of errors identified by the team and the quality of the objections raised. An AI-enabled process creates more value when it improves not only the outcome of a decision, but also the human system that will have to decide again.

Metric	Control question	Useful signal
Assisted quality	Is the result better with AI?	Accuracy, completeness, errors, time
Independent capability	Can the team reproduce the reasoning without AI?	Reconstruction, a similar case, retention over time
Decision accountability	Who chose, and by what criterion?	Named decision owner and explicit rationale
Evidence	Are the sources independent of the generated text?	Source verification, data and traceability
Reversibility	Can we correct course without disproportionate harm?	Pilot, ability to roll back, escalation procedure
Diversity of viewpoints	Has the system preserved objections and the voices of affected people or groups?	Dissent, employee voice, alternatives

A non-negotiable boundary

Decisions about hiring, dismissal, performance assessment, safety, health, legal matters or high-impact financial matters must not be made autonomously by AI. They require an authorised human decision owner, professional expertise, verifiable accountability and a clear procedure for stopping the process or escalating it.

PERSONAL

In everyday life: an assistant, not an all-knowing interpreter

In personal life, the risk of delegation is more subtle because AI enters areas in which we want reassurance: identity, relationships, life choices and other people’s motives. This is precisely where Lee’s personal validation effect becomes relevant: the tendency to judge favourable, apparently personalised descriptions as truer and more accurate. A response may feel intimate, precise and even liberating; the fact that it resonates with us does not make it true.

Mature use begins by stating our values before receiving the answer. It then separates three levels: facts, interpretations and unknowns. AI can help generate hypotheses, but it cannot know another person’s intentions with certainty, define us through a clinical label, pronounce our destiny, or replace observation of the body and relationships.

A proportionate personal practice

·  Write down what genuinely matters before asking for advice: values, limits and acceptable risk.

·  Ask AI to distinguish what it knows, what it infers and what it cannot know.

·  Ask for the counterargument and the condition that would make the most reassuring answer wrong.

·  Turn the reflection into a reversible small action and observe the consequences.

·  Compare the outcome with a trusted person or a professional when the stakes require it.

·  Reconstruct the choice without AI: if we cannot explain it, we probably do not yet own it.

The method does not require us always to distrust AI. It asks for trust proportionate to the available evidence, the stakes and the possibility of verification: neither technological superstition nor reflexive rejection. AI can be a remarkable conversation partner for broadening perspectives, preparing difficult conversations, structuring a plan or revealing alternatives. It becomes dangerous when its voice replaces contact with facts, the body, other people and consequences.

The blackout test

Imagine that the system is unavailable tomorrow. What remains? A decision you can defend? A criterion you can apply to a new case? A skill you can exercise? Or only a persuasive text you could not reconstruct?

GOVERNANCE

The most robust approach is proportionate, not ritualistic

A serious protocol should not become bureaucracy. A direct mode may be enough to correct a draft, organise notes or translate. The full cycle is needed when the stakes, responsibility, risk of dependence or value of learning increase. It must also be open to being disproved by results: if simple Socratic questions or a reasoning worksheet completed without AI produce the same independent capability with less effort, AI is not necessary.

By “Socratic questions” I mean questions that do not immediately provide the solution, but require the reasoning to be made explicit: “What fact supports this conclusion?”, “What would disprove it?”, “Which alternative are we overlooking?”, “Which criterion will we use to choose?” Their value is not in making the conversation more philosophical, but in returning the work of analysis and choice to the human being.

The principle is proportionate: the more a decision affects people, identity, money, safety or future capability, the greater the need for an explicit human decision, robust sources, a real-world test and an AI-free step. Optimisation does not mean maximising either automation or caution; it means choosing the right degree of pause and mental effort to protect judgement, the ability to correct course and learning.

Work and personal life remain distinct contexts. A method that helps a strategy team does not automatically transfer to a relationship or health decision: sources, responsibilities, risk thresholds and the meaning of error all change. The common element is that human beings retain control and responsibility for the reasoning, not that the procedures are identical.

Seven questions to ask before ending a session

1	What was my position before using AI?
2	Which genuinely new alternatives emerged?
3	Which tensions, missing sources or facts that could disprove our hypotheses remain unresolved?
4	What is my decision criterion — stated in my own words?
5	What real, proportionate and reversible test will I run?
6	Which consequences will I observe, and who is accountable for them?
7	Can I reconstruct the reasoning without this conversation?

CONCLUSION

A mature AI culture

One managerial proposition worth testing is that resilience in the age of AI depends less on the number of licences than on people’s ability to use the machine without turning it into an authority; to increase speed without losing judgement; to experiment without mistaking a pilot for an established truth; and to decide even when assistance is unavailable.

In personal life, competence does not mean never delegating. It means distinguishing what can be entrusted to the machine from what must remain our own: the origin of the question, the meaning of our values, responsibility for the choice and contact with the consequences.

This leads to a discipline of co-intelligence — conscious collaboration between human and artificial intelligence — in which responsibility remains human. AI broadens the possibilities; the human being chooses the path and answers for the decision. The context provides feedback; an AI-free test checks what has actually been learned. The goal is not only to produce more, but to retain the capacity to remain the authors of our decisions. It is a practical framework to be tested in different contexts, not a universal law.

Andrea Viliotti

June 2026

TRANSPARENCY

Methodological note and limits of the evidence

This article develops a practical framework from three empirical studies and a conceptual analysis of human–AI interaction. It does not present a validated psychological scale or a universal law. The cited evidence concerns, respectively, adoption, perceptions of fictitious predictions and misinformation discernment: on its own, it does not demonstrate the causal superiority of the protocol, its transferability between work and personal life, or its effectiveness in clinical, legal and financial settings. The numerical examples clarify the difference between assisted output and independent capability; they do not measure a psychological trait.

REFERENCES

Essential sources

Ibrahim, F., Münscher, J.-C., Daseking, M., & Telle, N.-T. (2025). The technology acceptance model and adopter type analysis in the context of artificial intelligence. Frontiers in Artificial Intelligence, 7, 1496518. DOI: 10.3389/frai.2024.1496518. Data cited: Table 1, pp. 4 and 7; limitations, p. 11.

Lee, E. (2024). The Power of Perception in Human-AI Interaction: Investigating Psychological Factors and Cognitive Biases that Shape User Belief and Behavior. MIT, arXiv:2409.15328v1. Data cited: Table 3.2, p. 50; limitations, pp. 57–58.

Rani, A., Danry, V., Liang, P. P., Lippman, A. B., & Maes, P. (2026). Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills. CHI ’26; arXiv:2510.01537v2; DOI: 10.1145/3772318.3790656. Data cited: Figure 6 and Sections 6.2–6.4, pp. 8–10; limitations, p. 14.