Imagine you’re evaluating an autonomous diabetic retinopathy screening platform for your practice. You ask AI to summarize the benefits. You get a clean list: improved access, earlier detection, reduced specialist burden, better consistency.
The list isn’t wrong. But it isn’t enough.
We’ve spent the last two years teaching clinicians how to prompt AI — be specific, give context, define the audience, ask for a table, ask for a patient-friendly summary. These things are useful. They’re also incomplete. The next phase of AI literacy isn’t learning to ask better questions. It’s learning to ask AI to push back.
Prompt Pushback
I call this prompt pushback: asking AI not just to generate an answer, but to challenge the assumptions behind the question. It’s the difference between using AI as a stenographer and using it as a thinking partner.
That difference matters because generative AI can make weak thinking sound strong. It produces a polished paragraph, a clean framework, a confident recommendation — before we’ve done the harder work of deciding whether the premise is right.
Take the DR screening example again. A useful prompt isn’t “summarize the benefits.” It’s: “Evaluate this like a skeptical clinician, a regulator, a payer and a practice administrator. Was the model validated in a population like mine? How does it handle media opacity, high myopia, retinal comorbidity or atypical pathology? What happens when image quality is poor? Who owns follow-up? Does this improve outcomes, or just create another data point?”
Or take a clinical encounter. Instead of asking AI, “Help me explain why this patient needs cataract surgery,” ask: “Challenge my reasoning. What else could be contributing to this patient’s symptoms? What findings would make cataract surgery less likely to improve vision? What should I document before recommending surgery?”
The first prompt helps us move faster. The second helps us think better.
Thinking Better
This isn’t a new idea. Before medicine, I studied philosophy, and what I’m describing is the Socratic method dressed in 21st century clothes. Socrates didn’t teach by handing out answers. He asked questions — uncomfortable, recursive, sometimes maddening ones — until his interlocutors realized their confident positions didn’t survive scrutiny. The point wasn’t to humiliate. It was to expose the gap between what someone thought they knew and what they could actually defend.
That gap is exactly what generative AI exploits – by accident. A polished output looks like knowledge. Prompt pushback is the discipline of testing whether it actually is.
Speed is seductive — AI can help us write, summarize, document and decide faster. But faster isn’t always better, and the cost of bypassing reasoning is no longer theoretical. In a recent New York Times op-ed (“There’s a Good Reason You Can’t Concentrate,” March 27, 2026), Cal Newport cites a January study finding a significant negative correlation between frequent AI tool usage and critical thinking ability, and a separate study tracking brain activity in subjects writing with LLMs that found brain connectivity scaled down with the amount of external support. Cognitive offloading is measurable.
‘Workslop’
There’s a name now for what this produces: workslop — a term coined by researchers at the Stanford Social Media Lab and BetterUp Labs in a September 2025 Harvard Business Review piece. They define it as AI-generated content that “masquerades as good work, but lacks the substance to meaningfully advance a given task.” Their survey of 1,150 workers found that roughly half of recipients viewed workslop senders as less creative, capable and reliable than before.
In an office, that’s annoying. In medicine — where credibility is the currency of trust between clinicians, patients and referring providers — clinical workslop is more than wasted time. It’s reputational. A note, summary, referral letter or diagnostic impression that sounds reasonable but hasn’t been examined is a liability with your name on it.
Going Beyond the Surface
Prompt pushback is one defense. The questions are old; only the interlocutor is new:
What am I assuming?
What’s the strongest counterargument?
What information is missing?
What would make this answer unsafe?
These aren’t adversarial for the sake of it. They’re protective. AI is a tool for language, pattern recognition and synthesis — not a source of truth. It can support reasoning. It shouldn’t quietly replace it.
This matters most as diagnostic AI moves deeper into eye care: OCT interpretation, glaucoma progression models, AI triage for red eye, autonomous screening. Surface-level interaction returns surface-level answers. The questions that matter are the ones that don’t appear on the marketing slide.
That’s where AI literacy in healthcare has to mature. We don’t need clinicians who are good at producing AI-generated content. We need clinicians who can interrogate it. That means teaching physicians, optometrists, administrators and trainees not just how to prompt — but how to pressure-test.
The irony is that the best way to use AI may be to make it less agreeable. Don’t let it flatter the premise. Ask it to argue the other side, identify the weak spots, slow you down. Socrates was annoying enough to get himself sentenced to death — we don’t need to go that far. But a little of that disposition, turned on our tools rather than our colleagues, would serve medicine well.
Prompting is about getting an answer.
Prompt pushback is about earning confidence in that answer.
In medicine, that difference matters.
Read more columns from our Professional Co-Editors here
Author
