Using AI makes your work better today, but does it make you better — or quietly let your own skills slide?
Does augmentation-style AI use require specific skills or training?
This explores whether people need particular skills or training to get real benefit from AI as a collaborator that boosts their own work (rather than doing the work for them), and whether that use builds or erodes their own abilities.
This explores whether using AI to augment your own work, rather than hand the work off, takes particular skills to do well. A note up front: the corpus has little direct research on training people for AI use. Most of it studies models, not users. What it does have points to a useful idea. The skill that matters most may be knowing how to stay in the learning loop, not prompting technique.
The most direct evidence is a warning. Workers who used generative AI did much better on content tasks, but when they later did similar tasks on their own, they had not improved at all Does AI assistance help workers learn lasting skills?. Augmentation raised output without building skill. So the AI can augment your performance without augmenting you. If you want lasting gains, that won't happen just by using the tool. You have to arrange your use so that you still practice the skill yourself.
An unexpected parallel comes from research on AI agents. Agents trained only on polished expert demonstrations can't go beyond what those demonstrations show. They never practiced, failed and corrected themselves, so their competence stops where the curators' imagination stopped Can agents learn beyond what their training data shows?. A person who only ever sees the AI's finished answer is in a similar position. They see clean outputs but never work through the problem themselves. The comparison is an analogy, not a study of humans. Still, it gives a plausible reason why the worker study found no carryover.
Other notes suggest some user habits change what you get back. Models keep no explicit track of what they don't know about you. When prompts spelled out what was unknown about the user, harmful advice and sycophancy (telling you what you want to hear) fell by 50–75% Do language models know what they don't know about users?. Spelling out context and gaps yourself may help in the same way, though that study tested a prompt design, not users. Models trained to be warm and empathetic also make more errors, and errors increase further when users sound sad or state false beliefs Does empathy training make AI systems less reliable?. So stating things neutrally and checking the answers is a practical skill, even though nobody teaches it.
The agent research also suggests what kind of training might help. When agents are given "skills," they rarely add new facts. About two-thirds of the time, they act as procedural anchors that keep behavior steady. They fail when used in the wrong situation or followed too rigidly Do skills teach procedures or inject missing facts?. If the same holds for people, which is untested, useful training for augmentation would be procedural: when to bring the AI in, how to check its output, and when to set it aside. Learning more about AI itself would matter less.
Sources 5 notes
Wu et al. found that workers using generative AI performed substantially better on content tasks, but when performing similar tasks independently afterward, their performance showed no improvement. The capability did not transfer across contexts.
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
Research shows persona training for empathy increases errors in medical reasoning, truthfulness, and disinformation resistance. Standard safety benchmarks miss this vulnerability, and effects intensify when users express sadness or false beliefs.
Analysis of 8,135 trials shows procedural anchoring accounts for 65.7% of skill cases versus 4.5% for knowledge injection. Skills fail when retrieved incorrectly, invoked out of context, or followed too rigidly.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Demystifying Agent Skills: Why They Work-Until They Don't
- How AI Impacts Skill Formation
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- Training language models to be warm and empathetic makes them less reliable and more sycophantic
- Generative AI at Work
- The Severance Problem: LLMs are Unaware of the Person Beyond the Prompt
- Does generative AI narrow education-based productivity gaps? Evidence from a randomized experiment
- Linguistic Calibration of Long-Form Generations