What NN/G's Editorial Process Teaches About Reviewing AI-Generated Code

Nielsen Norman Group published a piece authored by Raluca Budiu, an editor at the organization since 2013, describing how it uses AI in its editorial process without giving up one thing: human responsibility for what goes out under its name. The hook is legal, but the lesson is technical, and it applies just as much to those who write articles as to those who accept (or reject) what a copilot spits out in the code editor.

The starting point is the European Union's AI Act, whose transparency provisions take effect in August 2026. The law requires “AI-generated content” labels on certain texts of public interest, but it creates an important exception: content that has “undergone human review or editorial control” and for which a person or organization takes editorial responsibility is exempt. In other words, the law doesn't distinguish based on who typed the words, but on who answers for them.

That distinction is the heart of the argument, and it's exactly where the discussion about AI-generated code usually gets lost. The relevant question was never “did AI write it?” but rather “who guarantees it's correct?”

Where AI Actually Helps

Budiu is candid about the concrete uses, and they translate almost line by line to the day-to-day of engineering. It's worth listing what NN/G delegates to the machine, because the pattern is recognizable:

• Clarity and conciseness: rewriting paragraphs to say the same thing with fewer words, using Copilot in Word or Grammarly. In code, it's the equivalent of asking for a refactor of a verbose method or renaming variables to something readable.
• Rigid formats: NN/G has summaries with a fixed character limit (currently 160), study guides, and glossaries with a defined structure. AI nails this kind of tedious constraint effortlessly. For the dev, it's generating boilerplate, a docstring in the team's standard, or a schema that follows convention.
• Adapting content across channels: extracting from a report or video what's useful for an article. The parallel is migrating a piece of logic from one context to another, or translating a snippet between languages.
• Critique and checking: and here's the most interesting part for reviewers.

AI as a Critique Partner, Not an Oracle

The use Budiu describes as most valuable isn't AI writing, it's AI questioning. She recounts that an earlier draft of the very article argued that “thinking shouldn't be outsourced to AI,” but another passage described AI as a good reasoning partner. When she asked ChatGPT to critique the draft, the model pointed out the tension between the two ideas, which forced her to articulate the final principle more precisely.

Her method is the detail that matters: instead of asking “what do you think?”, Budiu shares her own assessment of the text and asks the AI to challenge or complement it.

Sometimes it finds things I missed; sometimes it disputes my assessment; and sometimes it makes recommendations I disagree with. In every case, I weigh the suggestions instead of accepting them as answers.

>

-- Raluca Budiu, Nielsen Norman Group

It's the same move as a healthy code review. A model reviewing a pull request can flag an unhandled edge case, a race condition, or an inconsistency between what a function's name promises and what its body actually does. That's gold. What it doesn't do is certify that the alert is valid. Budiu is categorical: when AI flags a problem in a citation to a study or law, it's up to the editor to go back to the source, check it, and correct it. AI raises the hypothesis; the human closes the verdict.

Translated into engineering terms: when a copilot suggests a security fix, the right response isn't to accept the diff, it's to go to the CVE, understand the vector, and validate that the change closes the hole without opening another. The model can hallucinate the problem as easily as the solution.

What Can't Be Outsourced

NN/G's editorial process, according to the piece, hasn't changed with the arrival of generative AI. An article goes through 2 to 3 rounds of review (sometimes more), with two editors checking logical coherence, UX accuracy, and text, and the cycle repeats until author and reviewers agree. Budiu compares this to academic publishing, not a blog. And it applies to everyone: guest author or senior executive, one day or twenty years at the organization, same rules.

The sentence that sums up the policy is the one that titles the original article: AI can help write, but it can't answer for the text. That's why NN/G doesn't name AI as an author. Human experts decide whether an idea deserves publication, whether an argument is sound and, ultimately, whether the organization is willing to put its name on it.

There's an honest final provocation: NN/G's own AI content policy was, in large part, generated by AI using that very article as input, then reviewed, edited, and approved by humans. The transparency about the process is part of the point.

The Parallel That Matters to Builders

Budiu's argument, applied to software, dismantles two lazy positions at once. The first is that of those who treat AI-generated code as inherently suspect and refuse the tool. The second, more dangerous, is that of those who merge agent output without reading it, trusting that “AI knows.” Neither takes responsibility, which is the only criterion that the AI Act (and serious engineering) recognizes.

The definition of “done” doesn't change because a model did the typing. A piece of AI-generated code that goes into production needs the same tests, the same review, and the same person willing to show up when the incident hits at 3 a.m. If no one can explain why that code does what it does, it isn't done, whether it was written by a human or a machine.

It's worth noting a difference in scale between text and code that the article doesn't cover. Bad prose that slips through review causes embarrassment; bad code that slips through causes data leaks, downtime, and technical debt that quietly piles up. That pushes the bar for reviewing AI output toward more rigor in software, not less. The more an agent writes, the cheaper it gets to generate volume, and the more expensive it gets to carefully review every line. The bottleneck shifts from writing to judgment, and it's precisely judgment that can't be outsourced.

For those who want the full text, including the list of channels where NN/G repurposes content and the reference to the organization's AI policy, Raluca Budiu's original article is published on the Nielsen Norman Group website.