If AI can produce a standard first draft, use it to save effort only where your workplace permits, then keep the learning work visible: decide what a good result must contain, check the draft against source material, ask for feedback on one judgment call, and sometimes repeat a similar task without AI. Workplace experiments show that assistance can improve immediate output, but an AI-assisted deliverable is not proof that you can explain, verify, or reproduce the work unaided. This is a practical learning design inferred from task studies, not a protocol those studies tested.
Can AI help you learn while it drafts?
Two things can be true at once: AI can help a newer worker deliver a better draft today, and relying on that draft may leave the underlying skill untested. The useful question is not whether the tool helped. It is what the worker can now judge, explain, or do again when the tool is unavailable.
In a randomized experiment at Boston Consulting Group, researchers gave consultants access to ChatGPT and brief training for bounded data-science tasks outside their usual skill set. The assisted group performed better on the three tasks, but researchers found no evidence of an advantage on related technical questions answered afterward without ChatGPT. An exploratory analysis also suggested weaker accuracy when participants judged which problems the tool could solve; this capability-judgment analysis was not preregistered, so it should be treated as tentative rather than a confirmed study finding. The paper involved BCG and OpenAI collaborators in task design, and tested short technical exercises, not routine first drafts or long-term career development. It supports a narrow point: successful assisted work does not by itself establish retained knowledge.
A different randomized experiment with 453 college-educated professionals found that ChatGPT improved speed and assessed quality on occupation-specific writing assignments. That is evidence for immediate performance on those tasks, not proof that workers learned the subject matter or that employers broadly adopted the tool. For example, a meeting summary may read cleanly while still omitting a decision owner or turning an unresolved question into a settled commitment. Those details require someone to compare the draft with the notes and understand what matters.
Sources: Generative AI and the Temporary Upskilling of Knowledge Workers; Experimental evidence on the productivity effects of generative artificial intelligence
What is the difference between a good draft and good judgment?
A good draft meets the immediate format and content requirements. Good judgment also means knowing whether the task suits assistance, what evidence is missing, which exceptions change the answer, and when to stop and ask someone. Answer-first delegation asks the tool for a finished reply and lightly edits it; bounded assistance uses a draft or critique while the worker retains the checks and decisions.
That distinction matters because capability can vary within one workflow. In a preregistered field experiment involving 758 management consultants, AI improved speed, quality, and task completion across 18 tasks designed to fall within its measured capability range. On a separate complex task outside that range, participants with AI access were less likely to produce a correct solution. The researchers tested consultants on designed tasks, not early-career employees learning through routine work. The practical implication is to identify the boundary before using the tool: drafting a routine summary may be suitable, while deciding what an ambiguous customer request commits the organization to may require escalation.
A useful boundary check is to separate a task’s repeatable format from its consequential judgment. Ask: can each material statement be traced to a supplied source; could a missing exception change a customer, financial, legal, or safety outcome; and who has authority to approve the result? If the answer depends on context the tool cannot see, keep that decision with the responsible person and ask for review. This checklist is an editorial application of the experiment’s task-boundary result, not a validated screening instrument. It helps decide where assistance may save time and where independent expertise or escalation remains necessary.
Sources: Generative AI and the Temporary Upskilling of Knowledge Workers; Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
What should stay yours when the first draft is automated?
The International Labour Organization’s 2026 publication is a review of empirical evidence, not one seven-country experiment. It brings together experiments, platform and firm data, other studies, and worker and firm surveys, including representative surveys in several countries. The review describes productivity gains as uneven and often unverified; it says large-scale displacement remains limited in the evidence it examined and identifies reduced opportunities for younger workers as a risk. That is a concern raised by a mixed evidence base, not a measured causal effect of AI drafts on early-career training or a forecast for an individual. The ILO’s 2025 task-exposure update likewise describes potential transformation, generally not automatic redundancy.
Use that uncertainty to make a work decision, not to predict your fate. For one recurring task, write down the source material, the quality criteria, and the parts that require judgment or approval. Check employer policy and confidentiality before using a tool. If assistance is allowed, keep a record of what you changed and why; that makes your contribution and remaining skill gaps easier to discuss with a supervisor. If a task is sensitive or high consequence, ask for a safe sample or supervised practice instead.
A degree, certificate, or vendor course is worth considering when a specific target role requires foundations or a credential you lack; it is not the automatic response to first-draft automation. For your wider decision, compare your actual task mix and constraints—such as location, salary floor, and learning time—before deciding whether to redesign your current role or explore an adjacent one. The free checker can offer task-level change-pressure signals, while the paid roadmap can compare career scenarios; neither predicts job loss or guarantees an outcome.
To make the discussion concrete, bring one example to that conversation: the source, the draft, one correction you made, and one uncertainty you could not resolve. Ask what evidence would have changed the decision and what a more experienced colleague checks first. This feedback can reveal a specific next learning target, such as identifying commitments in customer correspondence or distinguishing reported facts from interpretation. It also gives a manager a chance to clarify which tasks are approved for assistance. The small project is useful even if the answer is to keep that task unaided.
Sources: Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality; The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence; Generative AI and jobs: A 2025 update
Sources and notes
- Generative AI and the Temporary Upskilling of Knowledge Workers
The randomized BCG consultant experiment found assisted task gains and no detected advantage on unaided follow-up technical questions. Its non-preregistered exploratory capability-judgment analysis suggested weaker accuracy; the study does not test routine writing or long-term development.
- Experimental evidence on the productivity effects of generative artificial intelligence
The randomized writing-task experiment supports immediate speed and quality gains for college-educated professionals on its assigned midlevel writing tasks.
- Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
The preregistered consultant experiment found benefits on 18 tasks within its capability boundary and lower correctness on a separate task outside it.
- The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence
The ILO’s 2026 empirical review synthesizes heterogeneous evidence, including representative worker and firm surveys alongside experiments, platform data, and other research. It describes uneven productivity gains, limited large-scale displacement in the reviewed evidence, and younger-worker opportunity erosion as a risk, not a measured effect of AI drafts.
- Generative AI and jobs: A 2025 update
The ILO task-exposure update describes occupational exposure as transformation potential, generally not automatic redundancy.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map