With AI assistants taking on more writing, analysis and problem-solving, workers across all industries will still need the brainpower, judgment and expertise to challenge an answer, explain a decision, and handle situations the technology gets wrong.
Developing those capabilities requires deliberate practice, even when automation offers a faster route to a finished task —the challenge is deciding which work to delegate, and which work provides experience worth preserving.
This means employees and managers must make that distinction part of AI adoption, alongside training in how to use the tools.
Nick Gonios, vice president of learning transformation and company ambassador at Epignosis, the parent company of TalentLMS, says the difference rests on what the technology enables a worker to avoid.
“The key difference is whether it is helping someone build on what they know or helping them avoid what they don’t know,” he says. “When it becomes a substitute for understanding the work itself, the line begins to blur.”
When Completed Work Conceals a Skills Gap
AI can help employees produce work beyond their training without making the underlying knowledge gap visible to a manager—or to the employee.
TalentLMS’s Learning Debt research found 59% of employees reported using AI at least sometimes for tasks they had not been trained to perform. Gonios says the report also found that 29% had delivered work they could not fully explain.
While those findings do not establish that AI causes cognitive decline, they do suggest that output alone is an incomplete measure of employee capability.
“If AI allows someone to complete a task without fully understanding it, the organization may not realize there's a learning need,” Gonios says.
For a developer, that distinction might involve generating working code without understanding its failure conditions. For an analyst, it could mean presenting a recommendation without being able to explain the assumptions behind it.
The practical test is whether the employee can defend the result, investigate a problem and recognize when additional expertise is necessary.
Separate Research Findings from Sweeping Claims
Concerns about AI and critical thinking deserve attention, but the evidence does not support treating all AI use as harmful.
A Microsoft Research and Carnegie Mellon University study found that greater confidence in generative AI was associated with less self-reported critical thinking. The research examined workers’ reported behavior and perceived effort; it did not establish that using AI inevitably causes lasting loss of ability.
Dr. Gina Smith, IDC senior research director for IT skills for digital business, says the effect depends on how people use the technology.
“Claims that AI universally degrades human cognition are premature,” she adds. “Well-designed workflows and training can mitigate these risks.”
Smith describes productive reliance as actively questioning, validating and contextualizing AI output. The worker brings expertise to the exchange and treats the response as one source of information.
“Productive users demonstrate 'cognitive discipline'—they know when to challenge, when to trust, and when to escalate,” she says.
For tech professionals, that makes judgment a capability to practice and demonstrate, rather than an assumed benefit of experience.
Make an Independent First Attempt
One useful distinction is whether the task is simply something to complete or an opportunity to develop a skill essential to the worker’s career.
From the perspective of Kimball Parker, CEO of SixFifty, sequencing matters. Attempting a problem independently gives learners experience with the difficulties they will later need to recognize in AI-generated work.
“If you are trying to learn a new skill, you may need to do that task without AI first,” he explains. “There’s no other way to experience all of the challenges and pitfalls of doing the task yourself.”
That does not require avoiding AI throughout the assignment. A worker could outline an argument, identify likely causes of a technical problem or propose an approach before asking an assistant to critique it.
The initial attempt creates something to compare with the AI response, and differences thereby become opportunities to investigate, rather than instructions to accept automatically.
Parker also distinguishes career-critical skills from incidental tasks. For example, automating presentation formatting carries different developmental consequences for someone whose core responsibility is analysis than outsourcing the analysis itself.
Train People to Challenge Polished Answers
A policy requiring employees to review AI output would be insufficient if employees have never practiced recognizing what could be wrong.
Gonios says he recommends exercises using plausible responses that contain errors or omit important context. Employees should then identify the issue and explain how they would verify it.
“I’d also change how we think about assessments,” he notes. “Rather than evaluating only the final product, ask employees to explain why they made a particular decision, what information they relied on, and how they challenged the AI’s output.”
For technical teams, that could mean explaining why a proposed fix addresses the root cause, which test would disprove an assumption or what evidence supports an architecture recommendation.
Smith similarly recommends scenario-based training and periodic independent exercises. Managers can subsequently assess error detection, escalation decisions and the reasoning employees document when accepting or rejecting AI suggestions.
Preserve Experiences That Build Judgment
The development challenge extends beyond individual habits, as employers must protect opportunities for less-experienced workers to acquire the expertise they will eventually use to supervise AI.
Diana Kelley, chief information security officer (CISO) at Noma Security, says foundational knowledge should be paired with practical experience through labs, internships and real projects.
“The long-term risk is a pipeline that runs dry,” she says. “Cut off the early-career pathways and you lose the next generation of defenders.”
Removing routine assignments without replacing their learning value can leave employers demanding experienced judgment while providing fewer ways to develop it. Kelley says she favors apprenticeships and junior roles in which AI helps people become more capable.
For professionals, the objective is to retain the ability to improve the machine’s contribution. That includes supplying context, questioning assumptions and making decisions that the generated answer leaves unresolved.
Smith says organizations with the highest AI ROI are likely going to be those that embed “cognitive guardrails” — policies and workflows that require human review, periodic skill refreshers, and ongoing measurement of independent reasoning.
“Employers should design workflows that alternate between AI-assisted and independent work, invest in continuous human skills development and use analytics to monitor for signs of over-reliance,” she says. “The goal is to achieve 'augmented intelligence', where AI amplifies human expertise. It shouldn't replace it.”
