AI adoption masks eroding skills, experts warn in 2026 study

The capability mirage: what it is and why it matters

AI adoption risks creating an organizational mirage where workplaces appear highly capable while people's real skills quietly erode beneath polished AI-generated output, according to AI experts interviewed for a joint Anthrome Insight–Axialent study in September 2026. The problem is not that AI produces bad work. It is that AI-generated output can look competent even when the person behind it is not, and leaders have no reliable way to see the gap until a critical moment exposes it.

Stephanie Antonian, founder and CEO of Aestora, described the mechanism in the 2026 study: "The upside [of AI tools] is that everyone can produce a level of work that's pretty good for basic tasks. … It looks pretty good. But then you don't know what's underneath it, how resilient that piece of work is, or whether it's going to give you an additional liability." The mirage is structural. AI lets people deliver polished outputs for routine tasks without developing or showing the underlying capability. Leaders see the output, not the process, and cannot assess what happens when the task becomes non-routine.

Amir Michael, professor of accounting and deputy executive dean for executive and professional education at Durham University, noted in the same study that AI does not know when it is out of its depth. The tool will generate plausible-looking work even when the task exceeds its competence, and the user may not recognize the boundary either. The result is a double blindness: the AI does not signal its limits, and the person relying on it may not have the skill to notice when the output has crossed from adequate to inadequate.

AI amplifies capability, it does not level it

Gábor Szórád, CEO of AI Business Impact, observed in the September 2026 study that AI amplifies existing capabilities: "If you are a fantastic software engineer or a great manager, AI allows you to do more with the same energy. If you're a bad one, you're just going to create more crappy instructions, longer ones, more bad ideas." Strong performers use AI to extend their judgment and output. Weak performers use it to produce more work that looks competent but is not.

That amplification effect explains why the capability mirage is not a temporary adjustment problem. If AI magnifies both strong and weak performance, then the gap between what an organization appears capable of and what it can actually do under pressure will widen over time, not narrow. The polished surface hides a growing divergence in underlying skill, and the divergence is invisible until the work is tested in a way that AI cannot assist.

The study does not provide quantitative data on how quickly skills erode with regular AI delegation, or which functions or industries are most vulnerable to the mirage. What the experts describe is a pattern: delegation becomes habit, habit becomes dependency, and dependency erodes the judgment needed to know when AI is producing work that will not hold up.

The trust collapse when misrepresentation is discovered

Elisa Farri, vice president at Capgemini Invent Management Lab, stated in the 2026 study that when AI-assisted work is discovered to have been misrepresented, the impact on trust is severe: "The impact on trust is 0 [doubt] to 100." The collapse is not gradual. It is binary. A leader who discovers that a report, analysis, or recommendation was generated by AI and presented as judgment does not recalibrate their trust incrementally. They move from assuming competence to assuming none.

That trust collapse is distinct from the normal risk of poor work. Poor work can be corrected, and the person who produced it can be coached or reassigned. Misrepresented AI-assisted work breaks the assumption that the person understands what they delivered. The leader no longer knows whether the person can do the work, or whether they can recognize when the work is inadequate. The relationship becomes unworkable, and the damage extends beyond the individual to the team's credibility.

Albert Durig, cofounder and partner at Triviam Consulting, expressed concern in the study about losses in the "muscle of critical thinking" as a result of AI adoption. The metaphor is precise. Critical thinking is a skill that atrophies with disuse, and AI delegation is a form of disuse. The person who relies on AI to draft, analyze, or recommend is not exercising the judgment needed to evaluate whether the output is sound. Over time, the ability to make that evaluation weakens, and the person becomes less capable of recognizing when AI has produced work that will not survive scrutiny.

The study does not specify how long it takes for critical thinking skills to erode, or what proportion of organizations have experienced trust collapse after discovering misrepresented AI-assisted work. What the experts describe is a mechanism: delegation reduces practice, reduced practice weakens skill, and weakened skill makes it harder to detect when AI has failed.

AI-first versus purpose-first: the framing matters

Stephanie Antonian stated in the September 2026 study: "When you go AI-first, you have already told your organization it's not human-first." The distinction is not semantic. An AI-first approach treats the technology as the starting point for decisions about how work is done. A purpose-first approach treats the goal as the starting point and asks whether AI helps achieve it. The difference shows up in how people use the tool.

Thierry Kahane, an AI entrepreneur, stated in the study: "Small groups of people that get together for a specific purpose may outperform corporations because they are more nimble, flexible, fast-moving," emphasizing that cohesion around a goal is more important than technology focus. The observation aligns with Antonian's warning. Organizations that organize around AI adoption risk optimizing for the tool rather than the outcome. Organizations that organize around a clear goal can evaluate whether AI serves that goal or distracts from it.

Amir Michael noted in the study: "AI can be the lead. That's the problem. As long as AI is your follower — it follows your requests, your orders — we're fine. The time that AI jumps to be your lead, that's the downturn." The distinction between AI as follower and AI as lead is operational. AI as follower means the person sets the direction, makes the judgment calls, and uses AI to execute or accelerate. AI as lead means the person accepts AI's suggestions without interrogating them, or structures their work around what AI can do rather than what the task requires.

The shift from follower to lead is often invisible. It happens incrementally, as people learn that AI's suggestions are usually adequate and stop checking them as rigorously. The habit of checking weakens, and the person begins to treat AI's output as a draft that needs only light editing rather than a starting point that requires critical evaluation. At that point, AI has become the lead, even if the person does not recognize it.

The inappropriate application risk

Andrea Jones-Rooy, a data scientist and visiting associate professor at New York University's Center for Data Science, cautioned in the September 2026 study against applying AI tools inappropriately, comparing it to the saying "If you're a hammer, everything looks like a nail." The risk is not that AI is a bad tool. It is that AI is a general-purpose tool, and general-purpose tools invite misuse. A person who has learned to rely on AI for one type of task will be tempted to apply it to other tasks, even when those tasks require judgment or context that AI cannot provide.

The hammer-nail analogy is precise because it describes a cognitive bias, not a technical limitation. The person who has a hammer does not see more nails. They see more problems that might be solved by hitting them. The person who has learned to use AI for drafting or summarization does not see more drafting tasks. They see more tasks that might be solved by asking AI to generate a plausible-looking output. The tool shapes the perception of the problem, and the perception of the problem determines whether the tool is applied appropriately.

The study does not provide examples of inappropriate AI application, or data on how often it occurs. What the experts describe is a pattern: people learn to use AI for tasks where it performs well, then extend that use to tasks where it performs poorly but plausibly. The extension is invisible because the output still looks competent, and the person may not have the expertise to recognize that it is not.

Active engagement versus delegation

Elisa Farri advised in the 2026 study that people need to engage in active, back-and-forth interaction with AI, interrogating its suggestions, pushing back, and cocreating outputs, rather than using AI as a delegation tool. The distinction is behavioral. Delegation means asking AI to produce a result and accepting it with minimal revision. Active engagement means treating AI's output as a draft that must be tested, questioned, and refined through multiple iterations.

Active engagement requires the person to have enough expertise to recognize when AI's suggestions are inadequate, incomplete, or off-target. That expertise is the same skill that erodes when AI is used as a delegation tool. The paradox is that the people who most need to interrogate AI's output are the ones least likely to have the skill to do so, because they have been delegating to AI long enough that their judgment has weakened. The people who have the skill to interrogate AI effectively are the ones who least need AI's help, because they already know how to produce the work themselves.

The study does not describe what active engagement looks like in practice, or how organizations can train people to interrogate AI outputs effectively. What the experts describe is a principle: AI should be used to extend judgment, not replace it. That principle is easy to state and difficult to operationalize, because it requires people to maintain a level of skill and vigilance that AI is designed to make unnecessary.

What this means for consulting leaders

The capability mirage is a diagnostic challenge, not a technology problem. The risk is not that AI produces bad work. It is that AI produces work that looks good enough to pass initial review, while the people behind it lose the ability to recognize when the work is inadequate. Leaders cannot solve that problem by banning AI or by auditing every AI-assisted output. They can solve it by assessing not just deliverables but the capability that produced them.

That assessment requires a shift in how leaders evaluate work. The traditional approach is to review the output and ask whether it meets the standard. The capability-mirage approach is to review the output and ask whether the person who produced it can explain the choices they made, defend the assumptions, and describe what they would do differently if the context changed. If the person cannot do that, then the output is a mirage, regardless of how polished it looks.

The amplification effect described by Gábor Szórád means that the gap between strong and weak performers will widen as AI adoption spreads. Strong performers will use AI to do more high-quality work. Weak performers will use AI to produce more work that looks competent but is not. Leaders who rely on output quality alone will not see the divergence until a crisis forces people to work without AI assistance, at which point the gap will be undeniable and the cost of closing it will be high.

The trust collapse described by Elisa Farri means that leaders cannot afford to discover misrepresented AI-assisted work after the fact. The damage is too severe, and the recovery is too uncertain. Leaders need to create conditions where people can be transparent about when and how they use AI, without fear that transparency will be interpreted as incompetence. The alternative is a culture of concealment, where people hide their AI use and leaders lose the ability to assess what their teams can actually do.

The distinction between AI-first and purpose-first approaches, described by Stephanie Antonian and Thierry Kahane, is a framing choice that leaders control. An AI-first approach signals that the organization values adoption and efficiency. A purpose-first approach signals that the organization values outcomes and judgment. The signal shapes behavior. People in an AI-first organization will optimize for using AI, even when it is not the best tool for the task. People in a purpose-first organization will optimize for the goal, and use AI when it helps and avoid it when it does not.

The difference between AI as follower and AI as lead, described by Amir Michael, is a boundary that leaders must enforce through culture and practice. AI as follower means people use AI to execute their judgment. AI as lead means people accept AI's judgment without interrogation. The boundary is not technical. It is behavioral, and it erodes through habit. Leaders who want to prevent that erosion need to create regular opportunities for people to demonstrate their judgment without AI assistance, so that the skill does not atrophy and the organization retains the ability to assess who can do the work and who cannot.

The inappropriate application risk described by Andrea Jones-Rooy means that leaders cannot assume people will use AI only for tasks where it is appropriate. People will extend AI use to tasks where it performs poorly but plausibly, because the tool is available and the output looks competent. Leaders need to create explicit guidance about where AI should and should not be used, and they need to audit not just whether the guidance is followed but whether people understand why it exists.

The active engagement principle described by Elisa Farri is the operational core of the solution. People who interrogate AI outputs, push back on inadequate suggestions, and cocreate results through iteration will maintain their judgment and avoid the capability mirage. People who delegate to AI and accept its output with minimal revision will lose their judgment and become dependent on a tool that does not know when it is out of its depth. Leaders cannot force active engagement, but they can model it, reward it, and create conditions where delegation is recognized as a risk rather than an efficiency gain.

The study does not provide metrics for detecting the gap between output quality and underlying capability, or data on how long it takes for critical thinking skills to erode with regular AI delegation. What the experts provide is a framework for understanding the mechanism: AI enables people to produce work that looks competent without developing or demonstrating the underlying skill, and leaders who rely on output quality alone will not see the gap until it is too late to close it without significant cost.

Sources

• How AI Creates a Capability Mirage, MIT Sloan Management Review, September 14, 2026