AI at Work

Learn AI by Comparing the Draft With the Finished Work

By Owen Marsh / October 11, 2026 / Updated November 8, 2026

Adult comparing notes and paperwork beside a laptop

Quick answer: Don’t just save the polished result from an AI tool. Save the task, the first output, your edits and the reason for each edit. That small log shows you what you checked, what you fixed and what you still had to decide yourself, which is where the real learning happens. It is our suggested exercise for low-stakes work, not an official method, and it works best on tasks where you can verify the answer without asking the same tool.

Key takeaways

Keep the correction, not just the answer. A folder of finished outputs shows what the tool produced; a log of edits shows what you can judge.
Pick practice tasks with a clear reference you can check yourself, and use invented or shareable material only.
Write down four things per exercise: what you asked for, the first output, your correction and why you made it.
Time the whole job, including prompting, checking and fixing. In METR’s 2025 study, developers who were slower with AI still believed they had been faster.
No log or score ranks AI products. It tells you about your own task and workflow.

The idea in plain terms

Most people judge an AI session by the last thing on the screen: a clean paragraph, a tidy agenda, a decent email. But the interesting part is the gap between the first output and the version you were willing to put your name on. Every change you made is a small decision about accuracy, tone or purpose.

So the exercise is simple. For each practice task, keep the request, the first output, the edits needed to make it usable and the reason for each edit. After a few rounds, that record shows how you review, where you slip and what to practice next.

One honest caveat: this is our editorial suggestion. No standard or study we cite endorses this exact exercise. We are applying a broader idea, that evaluating AI output matters, to a small everyday habit.

What the evidence says, and what it doesn’t

NIST: evaluation is part of trustworthy AI

The U.S. National Institute of Standards and Technology describes its AI Risk Management Framework as a resource for building trustworthiness into how AI products and services are designed, developed, used and evaluated. Version 1.0 was released on January 26, 2023, and the framework is intended for voluntary use. Its four functions are named Govern, Map, Measure and Manage. NIST also released a Generative AI Profile (NIST AI 600-1) on July 26, 2024, also voluntary, to help organizations identify risks specific to generative AI. As of November 2026, NIST’s page says the framework is being revised, so check it for the current state. Neither document certifies a consumer tool or promises a correct answer, and they are written for organizations rather than individual learners.

METR: your sense of speed can be wrong

In a randomized trial published July 10, 2025, the nonprofit METR had 16 experienced open-source developers work through 246 real tasks, each randomly assigned to allow or disallow AI. With AI allowed, tasks took 19 percent longer. Before starting, the developers had forecast a 24 percent speedup, and afterward they still estimated about a 20 percent speedup. That gap between feeling and measurement is the reason to write times down.

Treat that result carefully. METR itself says it came from one setting and an early-2025 snapshot of tools, with a small sample, and its page now carries a notice that the results are out of date. In a February 24, 2026 follow-up on late-2025 tools (57 developers, more than 800 tasks), METR said selection effects, such as developers declining to work without AI, made its data an unreliable signal of current productivity, and it was redesigning the study. The lesson is not “AI slows you down” or “AI speeds you up.” It is that your own measurements beat anyone’s vibes.

Anthropic: how you use the tool matters

In a randomized trial published January 29, 2026, Anthropic studied 52 mostly junior software engineers learning an unfamiliar Python library. The group using an AI assistant averaged 50 percent on a follow-up quiz versus 67 percent for the hand-coding group. The AI group finished about two minutes faster, but that difference was not statistically significant. Participants who used AI to ask for explanations and conceptual questions tended to score higher than those who mostly delegated the work, although the authors say that pattern does not prove cause and effect. The study was small, was about coding rather than writing, and measured understanding shortly after one task. Still, it hints that staying engaged with the output, as a draft-versus-final log forces you to, is worth building in.

Choose work you can independently check

The log only teaches you something if you can tell right from wrong without asking the same tool to grade itself. Short tasks with a clear reference work best.

Good practice tasks	Why they work	Tasks to leave out
Organize a short, public event description into an agenda	You can compare each line against the supplied details	Anything involving a sensitive medical, legal or investing decision
Improve an invented practice email	You know what the message is supposed to accomplish	Workplace documents you are not sure you may share with a service
Summarize a public page you have just read	You can spot what was added, dropped or distorted	Private information about other people

Use invented practice material or information you are permitted to share. For anything from work, check your employer’s data rules before pasting it into a service.

Name your edits in ordinary language

Use a four-column note: requested result, first output, your correction and your reason. Plain words are fine. The reason column is the valuable one, because writing “unsupported information” or “too long for the audience” forces you to name the problem. Here is an illustrative example, invented for this article.

Requested result	First output	Correction	Reason
Agenda from a short public event description	Lists a speaker who is not in the description	Removed the name	Unsupported information
Same agenda	A long paragraph under each item	Cut to one line each	The audience needs a quick scan
Same agenda	A start time shifted by half an hour	Restored the original time	Changed a supplied fact

Notice that the three reasons differ: two are about accuracy and one is about format. That difference is the next step.

Read your log like a coach

After a handful of exercises, look for repeats. Are you mostly checking facts, fixing format or deciding what the result should accomplish in the first place? Each answer points to a different skill to practice. If most edits are factual, build a faster verification habit. If most are about purpose, spend more time on the request before you send it.

This is about your tasks and workflow, not a ranking of AI products, and a few exercises are not a scientific test.

Measure the whole job, not just the first answer

A fast first response does not tell you how long the finished work took. If you record time, include writing the prompt, checking the answer and making corrections. Then compare the full process with how you would handle the same small task yourself, and note where that comparison is uncertain.

Finish each exercise with two short sentences: what the tool contributed and what you remained responsible for. That explanation is often a better sign of learning than a big folder of outputs. You are building the ability to judge work, not collecting examples that merely looked convincing.

A quick exercise checklist

The task is low-stakes, and I can verify the result myself.
The material is invented, public or something I am allowed to share.
I saved the request and the first output before editing anything.
Each edit has a reason in plain words.
I noted the total time, including prompting, checking and fixing.
I wrote one sentence on what the tool did and one on what I decided.

Frequently asked questions

Do I need technical skills to try this?

No. The exercise works with everyday writing and organizing tasks, such as an agenda or a practice email. The skill you are building is judgment, not coding.

Does a long edit log mean the tool is bad?

Not necessarily. A log reflects the task, your standards and how clearly you asked. It is not a score for any product.

Is this based on an official method from NIST?

No. NIST’s frameworks are voluntary guidance for organizations that design, use or evaluate AI. They do not certify consumer tools or endorse this exercise. We are applying the broader idea that evaluation matters.

What if the AI task is for work?

Check your employer’s data rules before entering anything into a service, and keep your practice work separate from sensitive material. If you are unsure, practice with invented examples instead.

Before you decide

If you try this, start small: one low-stakes task, one page of notes, one honest look at what you changed. Treat any single study, including the ones above, as a snapshot rather than a verdict. The goal is to be able to say what the tool did, what you checked and why the finished version is yours.