What AI Can't Edit: How Human Editors Catch What Automated Tools Miss

Scientific manuscripts have always had vulnerabilities that are hard for authors to see in their own work. We wrote about those enduring problems in our previous post. In recent years, something new has entered the picture, and it's made certain problems harder to spot, not easier. Many researchers now use generative AI tools to help with translation, grammar, or composition. The results can read fluently and look correct on first pass. That's precisely what makes the problems they introduce so difficult to catch.

What AI Adds to the Picture

The issues we see in AI-assisted manuscripts are distinct from the ones that have always existed. In scientific writing, some carry serious consequences.

Hallucinated references

Errors in AI-generated references occur frequently. Parts of a reference—such as the title or DOI—may be incorrect, or the reference may be entirely fabricated. A reference could also be real but inappropriate to the context. In scientific writing, a fabricated or inaccurate citation isn't just a writing error. It's a credibility problem that can undermine a reviewer's confidence in everything that follows.

Generic text

AI systems can produce grammatically correct text, but the output often lacks specificity. It looks right, but it doesn't really say anything. Certain words and turns of phrase signal AI-generated text—an early example is the word "delve," which appeared disproportionately across AI-assisted writing. More broadly, AI systems replicate patterns from their training data, which results in repetitive phrasing and overused structures. One such structure is the tricolon, or constructions that list three items where one well-chosen example would be sharper.

Our editors describe what AI-generated text looks and feels like in practice:

"I find AI-generated content very middling, even though it reads correctly and fluently on my first pass. AI writing is unsurprising and inoffensive, which is perhaps the most offensive thing about its proliferation. Unlike human writers with real experiences and strong opinions, AI hedges and fails to innovate on or commit to ideas. It also spends a lot of word count on insubstantial ideas—ones that make me say 'well, obviously!' and wonder why a whole paragraph was dedicated to it."

"Repetitive, circular, and contradictory information. Often, AI-generated writing uses a lot of similar sentence structures, with information that doesn't seem to lead anywhere but instead just gets repeated in slightly different ways again and again. Sometimes the text contradicts earlier sentences, especially in longer documents; even basic information can be subtly different."

What Editors Catch That AI Tools Miss

A grammar-checking AI catches surface errors: spelling, punctuation, basic sentence structure. What it can't do is catch reasoning gaps, assess whether an argument is actually supported by the evidence presented, or judge whether a claim requires more than the author has provided. Those distinctions require understanding the subject matter. That requires a human.

In scientific manuscripts specifically, that gap matters in ways that are hard to automate. Multiple parenthetical citations can appear complete and still fail to support the claim they're attached to. A methodology section can read clearly and still describe a procedure that couldn't be replicated. For example, our editors frequently find that manufacturer information is incomplete or inconsistent when materials and equipment are reported in the methods. 

Although AI tools can assist with translation or flag surface-level errors, they aren't a replacement for the reasoning and attention to nuance that editorial review provides. We stay on the lookout for writing with AI hallmarks by naturalizing phrasing, identifying text with empty meaning, and asking for more information where an argument doesn't hold. As we edit, we see where claims are unclear or unsupported. We also look at what's already working, so the manuscript can build on its own strengths.

Our editors describe what that work looks like from the inside. It's editing that protects the author's voice just as much as it improves it:

"Making changes to improve word choice and flow is probably the aspect of editing that I find the most engaging. It's satisfying to select the precisely correct word for a sentence or to change an awkwardly organized sentence into something that flows well and logically connects to the rest of the paragraph."

"I like discovering a writer's quirks when they're already able to effectively argue a point. Writers usually aren't aware of what makes their work stand out, such as surprising but effective analogies, unique choices of evidence to support abstract claims, and distinct and memorable cadences. It's worth encouraging writers to lean into their voice: it shows how they think and helps their argument land differently from another writer making the same point."

"I enjoy finding ways to maintain how writers already effectively express themselves. The goal is to improve the text in ways that don't strip it of the author's point of view. Editing for correctness—or overediting with AI—can result in writing that's clean but flavorless. When the traces of the person who wrote the piece are gone, what's the point? Good revisions should tighten up a text without flattening it, so that the final draft sounds more like the writer."

What the Manuscript Reveals

There's a version of AI-assisted editing that produces text that is clean, fluent, and empty. It's correct on the surface but hollowed out underneath. That version strips a manuscript of the researcher's voice, replaces judgment with pattern-matching, and leaves behind writing that says little with great confidence. What human editorial review does is different. Editors read the manuscript the way a skeptical reviewer will and seek out where the argument holds and where it doesn't. They work to protect what makes the writing distinctly the author's while surfacing what the author needs the manuscript to say.

A manuscript that's been through that rigorous review process will read better; more importantly, it demonstrates that its author understands and can show what they're arguing.

Get a skeptical read before it matters. Explore Scientific Editing.