When AI Does the Thinking
A warning about AI that schools know about, but enterprises haven’t learned yet
In March 2026, Professor Jason Lodge of the University of Queensland and Professor Leslie Loble AM of the University of Technology Sydney published research examining what happens when students offload thinking to AI. Their findings were clear: AI enhances learning when it supports cognitive effort, checking grammar, brainstorming ideas, or testing understanding. But when AI does the thinking… writing essays or solving problems outright, it bypasses the mental work that builds lasting expertise.
What is metacognitive laziness?
The result is what the researchers call metacognitive laziness. AI's polished fluency creates the illusion of understanding by human authors who have prompted it. This reduces the need, or motivation, to think critically. Students feel they understand. They don't.
While this research focuses on students, its findings are already showing up in the workplace. MIT research, reported by Forbes, found that 95% of enterprise generative AI initiatives fail to deliver measurable ROI. The report points squarely at a workforce learning gap, not a technology failure, as the primary cause. The MIT study points to the failure of technology systems that don't retain feedback, adapt to context, or improve over time, as well as organisational integration failures
What is the “Performance Paradox?”
Enterprises are now grappling with a new dilemma dubbed the performance paradox: AI boosts short-term output but diminishes long-term learning and capability. This is a question of short-term gains versus securing long-term organisational readiness and resilience.
Education researchers have already mapped a vicious cycle that is, if anything, more pronounced at work.
The student (or in the enterprise case, the employer and the worker) seeks efficiency.
AI's fluency creates an illusion of competence.
The illusion triggers metacognitive laziness, and they stop planning, monitoring, and revising.
This leads to more mental outsourcing, which erodes their actual retained knowledge… and the organisation’s knowledge base.
This eroded knowledge makes the enterprise more dependent on the AI tools.
Greater dependence makes people (and the organisation) less capable of judging the quality of the tool's output. The cycle accelerates.
In the workplace, deadlines, KPIs, billable hours, and performance reviews all incentivize output over understanding, making this worrisome cycle the default. Workers are not lazy for choosing efficiency. They are rational. But when repeated day after day, this behavior becomes the mechanism of cognitive atrophy.
The cost of that atrophy is not abstract. McKinsey identifies the skill gap as the single largest driver of lost employee productivity. It’s larger even than disengagement or wasted time. This has the potential to cost a median-size S&P 500 company roughly $116 million a year. Errors climb and teams spend more time fixing problems than moving forward. People can look "ready" on paper, as L&D teams so often measure by completion rates and certification scores. But their “readiness” still falls apart the moment a real business conversation or workflow gets tense.
What workplaces are most at risk for metacognitive laziness and the performance paradox?
Experienced professionals typically use AI as an amplifier of their existing expertise.
Less experienced employees are more likely to use AI as a substitute for developing expertise in the first place. Their work may look polished, but the underlying judgment has never been built. Managers see strong outputs, not missing capability. The gap only appears when the AI is wrong, unavailable, or confronted with a novel situation: a regulatory question with no script, an irate customer, a negotiation that goes off the rails.
As one 2026 study of knowledge workers observed, the routine tasks most easily delegated to AI are often the very activities that develop the foundational skills needed for complex work. This isn't limited to entry-level roles either. It shows up in regulated, high-stakes industries too, where a completed course is not proof someone can perform the job on day one.
How do enterprises harness AI’s efficiency while not “dumbing down” their organisations?
The instinct is to solve a technology problem with a technology answer. However, the fix isn't less AI. It's using AI the right way and ensuring that AI supports deliberate friction built into how people learn and practice.
Lodge and Loble propose a cognitive mirror that asks questions, challenges assumptions, and encourages reflection instead of replacing thought. In practice, that means giving employees structured opportunities to reason through a problem before AI hands them the answer, and testing them on judgment, not just recall.
How does a ”cognitive mirror” work in fast-paced enterprises?
One way to operationalise a challenge aimed at learning is through simulation. Put employees inside realistic, high-stakes scenarios like a difficult customer conversation, a tense negotiation, or a complex stakeholder briefing. Require them to think, respond, and adapt in real time, and give them feedback on how they reasoned rather than simply whether they landed on a scripted right answer.
This is the design principle behind Cicero, CGS Immersive's AI-powered workforce performance platform. There are three mechanisms purpose-built into Cicero make that friction possible:
Unscripted persona challenges: Cicero Roleplay uses human-powered avatars and unscripted, lifelike personas that the employee talks to directly, to recreate the experience of engaging a customer, stakeholder, regulator, or difficult colleague… all with genuine conversational unpredictability.
Real-time reasoning diagnostics: After each simulation, Cicero's AI Replay mentor reviews the transcript and evaluates how the employee reasoned, not just whether they landed on the right line, surfacing what moved trust or satisfaction and where the reasoning broke down. It also looks for distinctly human capabilities like listening, acknowledging, and demonstrating empathy.
Compounding practice: Performance insights carry forward across practice sessions, with each session becoming more challenging than those that came before, so judgment is built cumulatively rather than tested once and forgotten.
What is the “Protégé Effect” in L&D?
The protégé effect is the learning phenomenon in which people learn material more deeply and retain it longer when they expect to teach it to someone else versus studying it only for their own benefit. Cicero automates and scales the Protégé Effect, capturing employee knowledge and paying it forward to other learners. Explaining, defending, and adapting your team members’ logic in simulated environments strengthens the same cognitive infrastructure that passive AI use erodes.
Importantly, this is also why a general-purpose AI chatbot cannot substitute for a purpose-built AI training solution like Cicero. A standard LLM is built to be agreeable and to resolve the user's request as quickly as possible. It does not offer a cognitive mirror. Cicero introduces productive friction to create lasting capability. The difference matters enough that it deserves its own side-by-side: see Cicero vs. Generative AI LLMs: Which Will Deliver Enterprise Resilience and Growth? for the full comparison.
Why are the decisions about how to deploy AI for L&D so important right now?
The skills most vulnerable to cognitive atrophy are also the hardest to rebuild: judgment, adaptability, critical thinking, and the ability to navigate difficult conversations. These capabilities are not developed by reading AI-generated answers. They are forged through experience, reflection, and learning from mistakes. Deloitte found that AI is automating precisely the entry- and mid-level work that once built proficiency, creating what it calls a "broken skills ladder." There is a widening gap between the expertise organisations need and the pathways to build it.
To be clear, this isn't an argument for using less AI. Using AI at work is inevitable, and in most cases, it's the right call. It makes people faster, sharper, and more capable when it's used well. The risk isn't the tool. It's treating AI as a replacement for thinking rather than a partner in it. Organisations that make this distinction see the difference in hard numbers.
Cicero's approach was recognised with a Silver Stevie Award for AI Breakthrough of the Year, tied to a 55% reduction in training costs and 30-day retention climbing from roughly 10% to over 85% among organisations using it. And the unrivalled impact of Cicero on driving real business outcomes for clients is what’s behind CGS Immersive being named among Fast Company’s 2026 Most Innovative Companies in HR.
The question every leadership team should be asking isn't whether to use AI. That decision has already been made. The question is whether your people are still building judgment, or quietly losing it, every time they use AI. That's a question worth putting in front of your own L&D and operations leaders this quarter, before the gap shows up at a moment that matters.
It’s also a question to ask now because it’s the same question your Board will soon be asking as well.
References:
Lodge, J. M. & Loble, L. (2026).Artificial Intelligence, Cognitive Offloading and Implications for Education. University of Technology Sydney.
Link: https://doi.org/10.71741/4pyxmbnjaq.31302475Temkin, B. (2026, July 21).Six Human Skills Kids Need to Thrive in the Age of AI. Human Experience Luminary.
Jack, A. (2026, July 24).Is AI killing critical thinking in the classroom? Financial Times.
Link: https://www.ft.com (paywalled; the article was provided as an attached document)Dell'Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023/2026).Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Organization Science.
Link: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321Lee, H-P. (Hank), Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025).The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. CHI 2025.
Link: https://doi.org/10.1145/3706598.3713778Asisof, A. (2026).The Invisible Gap: How AI Productivity Masks Eroding Expertise in Knowledge Work — and what to do about it. AMCIS 2026.
Link: https://aisel.aisnet.org/treos_amcis2026/7/Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O., & Mariman, R. (2025).Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26).
Link: https://doi.org/10.1073/pnas.2422633122
