Domain expertise doesn't save you from bad AI output
A popular LinkedIn take on AI right now is half right and half flattering. The flattering half is the dangerous one.
I found a LinkedIn post reacting to Gillian Tett’s Financial Times column on AI-native workers very interesting. The author’s argument lands well: “AI native” often means copy-paste over real knowledge. Vibe coding over understanding the big picture. Quoting ChatGPT over domain expertise. The corrective offered is that real understanding (history, physics, mathematics, art, literature, the ability to see the big picture) matters far more than the ability to write a good prompt.
I want to push on this, because the post is half right in a way that matters, and the other half is the kind of comfortable framing that’s going to get a lot of senior professionals into trouble over the next two years.
The half that’s right: prompt skill is overrated. Yes. Anyone who has worked alongside someone with deep domain expertise watching them interact with an AI tool can see immediately that the expert is asking better questions, catching more errors, and producing better output than the novice. Domain expertise still matters, and the post is correct to push back on the framing that prompt engineering is the new literacy.
The half that’s flattering, and wrong: the post implies that domain expertise plus a willingness to “own the outcome” is enough. The harder truth is that the new cognitive skill AI use requires is neither prompting nor domain expertise. It is verification under uncertainty, a skill almost nobody has been trained to do, including the senior experts who feel most equipped to.
The expert’s pattern-matching is the trap
When a senior researcher reads a draft from a junior colleague, they’re using pattern-matching built up over years of evaluating other people’s work. They notice when a citation looks off, when a claim doesn’t quite fit the methodology, when an argument has the shape of someone who didn’t fully understand the source. This pattern-matching is real expertise, and it works because the kinds of errors junior researchers make have particular signatures that experienced eyes can spot.
The senior researcher reading AI output is using the same pattern-matching system. The system fires the same way. The confidence the expert feels reading a plausible AI summary is the same confidence they would feel reading a competent junior researcher’s summary.
But the AI’s errors don’t have the same signatures. The AI hallucinates citations that look correct in form, with real-sounding journal names, plausible author combinations, dates that fit the timeline. The AI produces arguments that have the surface structure of expert reasoning while missing key dependencies between claims. The pattern-matching system was trained on human errors, which clustered around specific failure modes. AI errors cluster differently.
The expert doesn’t notice this gap at first. They read AI output, it feels right, they sign off on it. Most of the time it is right. Occasionally it isn’t, and the error gets through, because the expert’s confidence in their own pattern-matching is at a level calibrated for human errors, not AI ones.
I wrote about this dynamic in The More AI Output You Review, the Less You Can Judge It: exposure builds confidence, but consequences build expertise. The expert who has reviewed five hundred AI-generated reports without anyone telling them which ones contained errors hasn’t become a better AI evaluator than the novice, just a more confident one. Confidence and accuracy are different variables, and AI output systematically uncouples them.
Why “own the outcome” doesn’t do the work
The post’s strongest line is “you need to own the outcome. No matter how good the AI is.” The sentiment is correct, but it does no operational work.
What does ownership look like when the AI generated 80% of an analysis? At what point does verification become rubber-stamping? How does the expert tell the difference between “I verified this and it’s correct” and “I read it, it felt right, and I attached my name to it”? The post doesn’t answer because the post is operating at the level of professional values, not at the level of actual cognitive practice.
The honest version of “own the outcome” requires acknowledging that ownership requires verification, and verification of AI output is a specific cognitive activity that domain expertise prepares you for unevenly. Verifying a claim against your own knowledge is one thing. Verifying a claim against a domain you know well but didn’t generate the analysis for is another. Verifying a long document with multiple interdependent claims, when you’re tired, when there’s a deadline, when the document feels mostly right, is something else entirely.
The senior professionals most at risk are the ones whose careers were built on judgment under time pressure. That confidence to make calls quickly with incomplete information is exactly the cognitive style that misfires when AI output enters the workflow. The signal that worked under human-only conditions stops being reliable.
The new literacy is verification, not prompting
What’s actually happening in AI-native work isn’t that prompt engineering has replaced expertise. It’s that a third skill has emerged alongside both, and it doesn’t have a name yet in most organizations.
That skill is: the ability to systematically check AI output against ground truth, to know when you’ve done enough checking, to recognize the specific failure modes of the models you’re using, and to maintain calibrated confidence in your own verification rather than calibrated confidence in the underlying domain.
Domain expertise is necessary for this skill but doesn’t deliver it on its own. The expert who hasn’t developed verification practice will sign off on AI output at higher rates than the novice who has been forced to check everything. This isn’t hypothetical. It’s what the empirical evidence on AI-assisted work is starting to show across multiple domains. Radiologists at every experience level were affected by automation bias when an AI mammography tool offered incorrect BI-RADS assessments in a 2023 Radiology study. Lawyers have been sanctioned in more than 600 documented cases for filing briefs containing fabricated AI-generated citations, with a 2024 Stanford RegLab analysis finding that general-purpose AI tools hallucinate in roughly one in three legal queries. And a 2026 Clutch survey of 800 software professionals found that 59% of developers say they ship AI-generated code they do not fully understand.
The post’s framing pushes the opposite direction. It tells senior professionals that their existing expertise is what protects them, that AI is just a tool, that ownership is the answer. Each of those claims is partially true. Combined, they describe a posture that will produce a generation of confident sign-offs on flawed output.
What operational ownership looks like
If I were writing the post the LinkedIn author meant to write, the operational claims would be different.
Ownership of AI-augmented work means knowing which parts of the output you verified independently and which parts you accepted because they sounded right. Most professionals can’t currently make this distinction even to themselves.
Ownership means building a personal practice of catching the AI in errors regularly enough to maintain calibration. If you haven’t caught the AI in a substantive error in the last week, either you’re not using it for real work or you’re not checking carefully. There’s no third option for current-generation models.
Ownership means being explicit with collaborators about what level of verification you’ve performed. “I had the AI draft this and I read through it” is different from “I had the AI draft this and I checked every citation against the source.” The distinction matters and most teams don’t currently track it.
Ownership means maintaining the cognitive load of verification even when the AI has been right ninety-five times in a row. The temptation to relax that load is enormous, and the relaxation is exactly when the ninety-sixth error gets through.
None of this maps onto domain expertise alone. Some of it requires expertise. All of it requires a new kind of self-awareness about what your own verification consists of, which is the part that’s hardest to develop in people who were already confident in their expertise before AI arrived.
Where the LinkedIn post lands
The LinkedIn post will get a lot of agreement, because it tells senior professionals that what they already have is enough, and that the new generation’s prompt-engineering enthusiasm is a distraction from the real work. That’s a comfortable position, and the comfort is the tell.
The actual situation is less flattering. Prompt engineering is overrated. Domain expertise matters. And both are insufficient against the specific cognitive challenges of working with output that is locally plausible and occasionally wrong in ways that don’t pattern-match to anything in your training.
The skill that closes that gap is verification under uncertainty, and the people who develop it fastest are the ones who took the post’s comfortable framing least seriously. Owning the outcome requires more than asserting that you do. It requires building a practice of catching the AI in errors often enough to maintain calibrated doubt, and being honest with yourself about the difference between reading carefully and verifying.
Two years from now, the AI failures we’ll be reading about in trade press will not come from junior workers who relied too much on ChatGPT. They will come from senior experts who trusted their pattern-matching too far into a domain where the patterns have changed.