The Human in the Loop: Salvaging Projects From AI Slop

Last Updated on July 28, 2026

We’re far enough into the generative AI era that most of us who practice medical writing and medical editing have lived through at least one project where AI slop killed the timeline. And when that slop comes to you from a team lead or a client, you don’t get to opt out of the AI — you get to fix what it produced. Here’s how I’ve approached that, and what I’ve learned about where the approach still falls short.

What Is AI Slop?

AI slop is text generated by a large language model without the prompting discipline or contextThe information an AI model has access to within a single se... More it needed to do the job well. It’s not that the tool failed; it’s that the setup around it did. The result reads fluently but doesn’t hold up — thin on substance, wrong in places, or simply not fit for purpose.

Where I Run Into It

Most of my work sits under tight nondisclosure agreements, and on those projects AI use is usually off the table because of the IP exposure it creates. But my clients aren’t operating under the same constraints, and more of them are bringing generative AI into their own process upstream of me. Two recent projects — a journal article and a proposal — arrived with AI-drafted material I was expected to turn into something that can be submitted.

Why AI Slop Is a Real Project Cost, Not a Minor Annoyance

Working with it creates a specific set of problems:

• Hidden inaccuracies — errors that aren’t obvious on a first read
• Fabricated or misattributed citations — a reference that doesn’t exist, or doesn’t say what it’s cited as saying
• Lost efficiency — the whole point of using AI was speed, and that’s gone
• Correction time — finding the inaccuracies is one cost; fixing them is another
• Loss of coherence — across a multi-author, multi-section project, the seams show

The Approach: Be the Human in the Loop

Use a verification checklist. On both projects, my job became verifying four things, deliberately, one at a time:

1. Technical content — is the science, the data, the methodology actually correct?
2. Citations — does every reference actually exist, and does it say what it’s cited as saying? Don’t assume; check each one independently.
3. ContextThe information an AI model has access to within a single se... More — does this material make sense for the audience, the venue, and the argument it’s supposed to support?
4. Presentation — does the structure and framing do what the piece needs it to do? This includes voice: AI-drafted material rarely matches an author’s established tone, so even accurate content often still needs a pass for authorial consistency.
5. Conclusions drawn — do the stated conclusions actually follow from the evidence presented?

That order matters. It’s tempting to start with presentation, because that’s what’s most visible — but presentation is the layer most likely to hide the other three. (Read more about defensible AI workflows here.)

Why This Is Harder than It Sounds

LLMs are language models. They’re built to generate text that sounds right, which means it’s easy to read past a mistake without registering it as a mistake. Fluency is not a proxy for accuracy, but it’s very good at feeling like one.

Where an Adversarial Approach Helps—and Where It Can’t

One tactic I use is cross-model validation — running a second, different model to cross-check the first model’s output, a genuine adversarial check, since two models trained differently don’t share the same blind spots. A cheaper but weaker version of this is having the same model critique its own draft, using techniques like self-refine or chain-of-verification; it can still surface some errors, but it’s checking its work with the same judgment that produced the error in the first place, so it won’t catch everything a second model would. (Read more about AI workflows and tools here.)

Either way, this can catch a meaningful share of errors, but it isn’t a substitute for the human check, because the model is still working from what’s already known. If the material you’re verifying is novel or predictive — a new hypothesis, a forecast, an untested claim — there’s no existing record for a model to check it against. Fact-checking that kind of material is on you, and those are exactly the mistakes that are hardest to see.

The Bottom Line

None of this makes AI slop someone else’s problem to solve before it reaches you. If it lands on your desk, the fix is the same discipline every time: verify technical content, citations, contextThe information an AI model has access to within a single se... More, presentation, and conclusions, in that order, and don’t let a second model or a self-check stand in for your own judgment on anything novel. That discipline is slower than trusting the draft. It’s also the only thing that turns slop into work you can put your name on.

And that’s the part worth sitting with: AI slop is sold as an efficiency gain, but the version that reaches you often costs more time than it saves — the hours spent finding what’s wrong, fixing it, and re-establishing coherence across a project can outweigh whatever speed the draft bought upstream. We can’t always control whether AI slop reaches our desks. We can control whether it leaves them. That means pushing back on the process that produced it where we have standing to — flagging thin prompting or missing contextThe information an AI model has access to within a single se... More to the team lead or client who handed it over — and, where we don’t, refusing to let the discipline slip just because the draft looked done.

---

Set up Your Own Local AI Machine: A Short Course

If you’re weighing how to bring generative AI into medical writing without the slop—or the IP exposure—this fall DCC Cyber is launching a short course on setting up your own local, air-gapped GenAI environment: fast enough to be useful, private enough to keep the work under NDA. More on that soon.

---

Free AI Tool-Vetting Checklist for Freelancers

I’ve created a detailed vetting checklist that walks you through each of these considerations—download it now to start auditing your current tool stack.

---