Why would AI help beginners the most, yet slightly lower the quality of work done by veteran experts?

This explores why AI help that lifts beginners can slightly drag down the work of veterans. First, a direct caveat: the study behind this pattern, a customer-support deployment where the least experienced agents gained the most and the most experienced saw small quality declines, is not in the material retrieved here, and these notes don't test why experts lost ground. What the collection does have is a set of mechanisms from neighboring research. Together they make the result feel less like a paradox and more like something you could predict.

The first mechanism is leveling. An AI tool tends to push everyone toward one competent, middle-of-the-road answer. For a novice that's a step up. For an expert it can replace a better judgment with an average one. The clearest parallel in the corpus comes from hiring rather than support work: when cheap writing removes the effortful signal that once set strong freelancers apart, top-quintile workers get hired 19% less often and bottom-quintile workers 14% more often Does cheap writing weaken hiring based on worker ability?. That case is about how employers judge workers, not about how good the work is. The shape is the same, though: the tool compresses the distribution, and the people at the top have the most to lose.

The second mechanism is that fluent output switches off scrutiny. Several notes describe how polished AI text makes people feel competent and lowers their guard Does processing ease mislead users about their own competence?. People also start counting AI-assisted results as proof of their own skill Do AI-assisted outputs fool users about their own skills?, and these effects reinforce one another How do AI tools trick users into overestimating their own skills?. Experts may be especially exposed here. Their value lies in catching the subtle case where the standard answer is wrong, and a confident-sounding suggestion is exactly what makes that check feel unnecessary. A safety-focused note argues that the most dangerous systems are the ones that look competent while quietly weakening skepticism How do competent systems quietly undermine safety oversight?.

The third mechanism is slow skill erosion. Anthropic's engineers reported big productivity gains, yet most could fully hand off only 0–20% of their work. They worried that leaning on the model for routine tasks would wear away the hands-on practice they need to catch its mistakes Does AI assistance erode the skills needed to oversee it?. A study that mapped thousands of workplace AI risk scenarios found that 'augmentation' (AI assisting a person rather than replacing them) doesn't protect against this. Over-reliance can gradually hollow out both skill and oversight Does AI augmentation protect workers from skill erosion?. For experts, that erosion falls on the very edge that made them experts.

Put together, the counterintuitive takeaway is that an AI tool's benefit depends on where you start. If it encodes good-but-typical practice, it is a ladder for people below that level and a ceiling for people above it, unless experts treat its suggestions as drafts to challenge rather than answers to accept. A fuller answer would need the original support study and research that directly compares how experts and novices defer to AI. That gap is worth filling.