AI-Assisted Writing Disclosure: Where We Draw the Line

When you publish something a language model touched, what do you owe the reader? The answer most sites give — a vague ‘AI-assisted’ badge or nothing at all — fails both FTC guidance and Google’s E-E-A-T signals. We’ve been writing under our own disclosure rule for the last six months, and the threshold turned out to be more specific than ‘did an LLM type any of this?‘

What ‘AI-assisted’ actually covers

The phrase collapses three very different workflows. Treating them the same is why the disclosure conversation feels theatrical:

1. Research assistance — you ask an LLM to summarize a Hacker News thread, surface counterarguments, or pull dates out of a changelog. The model never writes a sentence that ships.
2. Drafting assistance — you give the model an outline and source material and it produces paragraphs you then edit, restructure, or rewrite. Some of the model’s prose survives to publication.
3. Generation with light human review — the model writes the article; a human reads it once for obvious errors and clicks publish.

A reader who sees ‘AI-assisted’ at the top of an article has no way to tell which of these happened. That ambiguity is the disclosure problem, not the use of the tool.

The threshold we use

Our rule, applied to every article on this site:

aiAssisted: true if any sentence in the published body was generated by a language model and survived editing — even one.

That excludes ‘I used ChatGPT to brainstorm headlines’ and ‘Claude summarized this thread for my notes.’ It includes any case where model output became reader-facing prose, regardless of how much editing came after.

We picked this threshold for one reason: it’s the only line readers can verify. Anything softer (‘substantially AI-generated’, ‘more than 50% AI-written’) requires the reader to trust an estimate they can’t audit. A binary ‘yes, model prose is in here’ gives them a clean signal.

What the platforms require

The disclosure threshold isn’t optional in 2026. Three layers of pressure stacked up in the last 18 months:

FTC guidance (updated July 2025). The Endorsement Guides now treat undisclosed AI-generated reviews and testimonials as deceptive. The guidance doesn’t cover all editorial content, but the precedent is established: material AI involvement that affects what a consumer believes about a product needs disclosure.

Google’s E-E-A-T signals. Google’s helpful content guidance (revised October 2025) explicitly accepts AI-assisted content if it demonstrates first-hand experience and expertise. Articles without disclosure that read as generated tend to rank worse — not because Google can detect AI text reliably, but because the human accountability signals are absent.

dev.to and Medium policies. Both platforms require explicit AI disclosure for syndicated content as of early 2026. Our dev.to crosspost pipeline includes the disclosure note in the body footer; without it, articles get flagged.

The compliance argument is real, but it’s not the strongest one. The strongest argument is that readers who notice undisclosed AI text stop trusting the rest of the site.

What we tell readers

The AiAssistedNote component renders this on every flagged article:

Parts of this article were drafted by a language model and then edited and fact-checked by the author. All product claims, pricing, and links were verified manually before publish.

Three things matter about the wording:

• ‘Parts’ not ‘this article’ — accurate to our drafting workflow, which is iterative editing of model output, not pure generation.
• ‘Edited and fact-checked’ — names the human role explicitly.
• ‘Product claims, pricing, and links verified manually’ — the specific accuracy commitment that matters for an affiliate site.

We don’t disclose research-only use because it doesn’t reach the reader as prose. We do disclose if a model wrote even a single paragraph that survived editing.

Where the threshold leaks

Honest accounting: the rule has edge cases we haven’t solved cleanly.

Translation and summarization. If a model translates a quote from another language or summarizes a long source into a single sentence we use verbatim, does that count? We currently say yes — it’s model-generated prose in the published body. But some publications draw the line at ‘model-generated original prose’ and exclude these cases.

Headline and subheading generation. We treat these as part of the body for disclosure purposes. Some sites exclude metadata.

Code samples. Generated code that compiles and runs is closer to factual content than prose. We disclose it the same way, but the argument is weaker.

Quotes from transcripts cleaned up by a model. This one is genuinely hard. The substance is the human’s; the wording is the model’s. We avoid this workflow entirely — quotes are either verbatim or paraphrased by a human editor.

Why the binary threshold beats percentages

The most common alternative — ‘X% AI-generated’ — fails three ways:

1. Unverifiable. No reader can audit the percentage. They have to trust your estimate.
2. Gameable. ‘Just 30% AI’ sounds responsible until you realize the 30% is the load-bearing content and the 70% is boilerplate.
3. Misleading on impact. A single hallucinated statistic in 5% of the article matters more than 95% of paraphrased background.

A binary ‘model prose is in here’ sidesteps all three. It doesn’t tell the reader how much, but it tells them the only thing they can act on: whether to apply extra scrutiny to factual claims.