While a recent vendor survey suggests employees feel their problem-solving skills are weakening due to AI, the more reliable evidence points to a specific risk: over-reliance on AI assistance prevents the effortful practice required to build durable, long-term capability. The problem isn’t the tool itself, but how we use it. We risk accruing significant capability debt if we allow frictionless answers to replace the valuable friction of real thinking.
Reading the AI Skills Warning Signs
#A recent L&D report from TalentLMS for 2026 made waves with the headline that 36% of employees believe generative AI tools are weakening their ability to solve problems on their own. Before we treat that number as solid ground, it pays to look at where it comes from. This is a self-reported perception from a survey, not a measured decline in performance. It also comes from a vendor in the learning management space, who has a commercial incentive to frame AI as a threat to human-centered learning.
So, is the 36% figure a hard fact? No. It’s shaky evidence. But is it pointing at something real? I think so. The perception is valuable because it’s directionally consistent with what we know from much harder, controlled science about how people actually learn and retain skills.
The Competence Illusion
#The survey figure points to a phenomenon we can call the competence illusion: the false sense of mastery we get from having an answer provided for us, without the mental work of retrieving it ourselves. When an AI gives you the perfect line of code, the polished paragraph, or the five-point plan, you’ve solved the immediate problem. But you haven’t necessarily learned how to solve it again next time. You’ve borrowed competence, not built it. This isn’t a new problem tied to AI; it’s a known trap in learning. And the latest tools just make the trap easier and more tempting to fall into.
What Harder Science Actually Tells Us
#The mechanism here is well-established in learning science. A 2025 study by Bastani et al. found that students using an LLM as a tutor often performed worse on later tests once the AI assistant was taken away. The tool helped them get the right answers during the learning phase, but it did so without creating durable knowledge. This happens because genuine learning requires effort. Seminal 2006 work by Roediger and Karpicke in Psychological Science demonstrated the power of what they called “test-enhanced learning.” The key finding was that the act of actively, and even difficultly, retrieving information from your own memory is what makes that memory stronger and more lasting. This is the core idea behind retrieval practice, and it’s the exact opposite of having an answer handed to you.
The High Cost of Frictionless Answers
#When we consistently rely on AI to close the gap between what we know and what we need to do, we stop doing the retrieval work. Each time an employee asks a chatbot for a process step instead of trying to remember it first, a learning opportunity is lost. This accumulates over time into a serious organizational risk. The term for this is capability debt: the growing gap between what your team can verifiably do on their own and what the work now demands of them. You only discover the true size of that debt when the system is down, the pressure is on, or a novel problem appears that the AI can’t solve.
The Fix: Designing for Desirable Difficulty
#The answer is not to ban generative AI. That’s a losing battle and it forfeits the massive productivity gains the tools can offer. The strategic move is to redesign work to ensure people are still doing the thinking. It’s about re-introducing what researchers like Robert Bjork call “desirable difficulties” into the workflow. Instead of using AI to get the final answer, use it to get the raw materials. Have your team use AI to conduct research, but require them to synthesize their findings and write the recommendation themselves, without the tool. Use it to generate practice scenarios, then have them work through the problems collaboratively. The goal is to use AI to augment thinking, not replace it. The question is not whether to use the tool, but how to use it without outsourcing the very effort that makes us capable.
Sources
#- Bastani et al. (2025). Generative AI Can Harm Learning. PNAS.
- Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science.
- TalentLMS. (2026). The 2026 L&D Report: AI and the Skills Economy.