Does AI Help Beginners or Experts More?

Some studies say AI lifts beginners most; others say it can't make novices experts. Both are true — here's the rule that resolves it and what it means for teams.

By Andrej Ruckij · · 6 min read

Does AI Help Beginners or Experts More?

By Andrej Ruckij · May 29, 2026

TL;DR: Both — on different tasks, and the confusion comes from conflating them. On executable tasks inside AI’s capability frontier (drafting, support replies, routine analysis), AI lifts beginners most — controlled studies show low performers gaining 34–40% while experts gain almost nothing, because AI hands the novice the competent-baseline they lacked. But on judgment-heavy tasks outside the frontier (knowing which AI suggestions are wrong, prioritizing, catching confident errors), AI helps experts more — because evaluating the output requires the very expertise the novice doesn’t have. The rule: AI compresses skill gaps where there’s a knowable right answer; it widens them where judgment is the work.

If you’ve read that “AI helps beginners most” and “AI can’t turn novices into experts,” you’re not reading contradictory research — you’re reading two true findings about two different kinds of work. This article reconciles them, because getting it wrong leads to deploying AI to the wrong people for the wrong tasks. It’s the “who benefits” instance of the broader question in when-can-you-trust-ai.

The case that AI helps beginners most

Three of the most rigorous workplace studies all found the same skill-leveling effect — AI raises the floor faster than the ceiling:

StudySettingWhat happened
Brynjolfsson, Li & Raymond (2023), n=5,179Customer support, real production+14% issues/hour on average — but +34% for novices and roughly 0% for the most experienced agents
Noy & Zhang (2023), Science, n=444Professional writing−40% time, +18% quality — low-ability workers gained most; the skill gap narrowed
Dell’Acqua et al. (2023), n=758Consulting knowledge workInside the frontier: +40% quality — and the bottom-half consultants gained most

The mechanism is intuitive: AI distributes a competent baseline. A new support agent doesn’t know the best-practice answer; the AI does, and hands it over. The skill premium that experienced workers used to charge for compresses. This is the empirical core of the wiki’s glossary/ai-skill-leveling finding and the glossary/automation-eats-execution pattern.

The case that AI helps experts most

Now the apparent contradiction. Other research — including Harvard Business School working knowledge — found AI can’t turn novices into experts, and in some settings adds more value for experts:

  • Knowledge distance is the limiter. AI helps you with a task only when you’re close enough to have relevant knowledge — like a woodwind player picking up a new woodwind. Hand the same player a string instrument and AI hits a wall. Novices lack the knowledge to use the AI’s output well.
  • Experts catch the errors. Experienced people can disregard hallucinations and judge which AI suggestions are worth pursuing. Novices waste time chasing confident-but-wrong outputs — exactly the glossary/jagged-frontier failure mode where AI is −19 percentage points less accurate outside its frontier.
  • Experts prioritize. A scientist uses domain judgment to pick which AI ideas to test; a novice tests false positives.

The rule that resolves it

Both are right because they measure different task types. The resolving distinction:

  • Executable, verifiable tasks with a knowable right answer (write this email, summarize this doc, answer this common support question) → AI lifts beginners most. The bottleneck was knowing the competent move, and AI supplies it. Low performers close the gap.
  • Judgment-heavy tasks where evaluating the output is the skill (is this AI answer correct? which of these is worth doing? what’s the right frame?) → AI helps experts most. You need expertise to use the AI safely, and the novice doesn’t have it. The gap widens.

AI compresses skill gaps where there’s a knowable right answer; it widens them where judgment is the work.

A useful test: can the task’s output be checked against a clear standard without expert judgment? If yes, it’s a beginner-leveling task. If checking the output itself requires expertise, it’s an expert-amplifying task. This is the same two-axis logic — frontier position crossed with the cost of a wrong answer — laid out in questions/what-ai-tools-actually-deliver-roi.

What this means for your team

  • Deploy AI to junior staff on executable work first. That’s where the ROI concentrates — new hires reach competent-baseline output faster. Brynjolfsson’s novices gained 34%; your experts will gain little on the same tasks.
  • Don’t expect AI to replace the judgment layer. It can’t make a junior person senior on tasks where catching the AI’s mistakes is the job. The “knowledge distance” wall is real.
  • The human edge moves up, not away. As AI absorbs executable work, the scarce skill becomes evaluation, prioritization, and framing — the things that let you tell when the AI is confidently wrong. (See glossary/automation-eats-execution.)
  • Beware the trap: giving a novice an AI tool for an expert task feels like leveling the field but produces confident, wrong work nobody on the team can catch. That’s negative ROI, not zero.

Will AI replace junior employees?

Short answer: it changes what juniors do, not whether you need them. AI handles the executable tasks juniors used to cut their teeth on — which is exactly why the development path has to change (juniors need to build the judgment layer faster, since the executable layer is automated). The risk isn’t “AI replaces juniors”; it’s “teams stop developing the judgment that AI can’t supply.” Full reasoning in when-can-you-trust-ai.

Key takeaways

  • AI lifts beginners most on executable tasks with a knowable right answer (Brynjolfsson +34% novices; Noy-Zhang low-ability gained most).
  • AI helps experts most on judgment tasks where evaluating the output requires expertise (knowledge-distance wall; experts catch hallucinations).
  • The two findings aren’t contradictory — they describe different task types.
  • The test: can the output be checked against a clear standard without expert judgment? Yes → beginner-leveling. No → expert-amplifying.
  • Deploy AI to juniors on executable work; keep the judgment layer human; expect the human edge to move up the skill ladder.

Sources