I had read that transcript three times and still missed the paragraph that would have landed badly.
In August 2024, I came back to ChatGPT with a stakeholder-interview transcript I needed to turn into a requirements document. I stopped treating it like a search bar and started giving it context. The bland prompt "Summarise this stakeholder interview transcript" became: "You are a consultant at a pharma IT company. I am about to write a requirements document for a new clinical data review feature. Below is a 40-minute interview transcript with the head of clinical operations. Pull out the three things she clearly wants, the two things she said no to, and the one thing she contradicted herself about." Same transcript, same model. The second answer told me which paragraph I needed to rewrite before showing the document to anyone.
I had two decades in pharma IT. I had read that transcript three times. The paragraph was right there. That was not a prompting win. That was an AI devil's advocate, and it raised a question I have not put down since: if twenty years of pattern recognition could make me miss that, what else is it hiding?
Twenty years of expertise does not close your blind spots; it teaches you not to notice them.
If I could miss a paragraph in a document I had already read three times, what does that say about the harder decisions, the ones involving budget direction, system architecture, or a position I have already emotionally committed to?
The answer is confirmation bias: the shortcut that tells you what you expect to find and filters out the rest before it reaches conscious thought. Sunk cost compounds it by making the cost of reversal feel larger than the evidence supports. The curse of knowledge makes the assumption you have held longest feel like background truth rather than an assumption. Status quo bias makes the current arrangement look safer than it is. The availability heuristic makes the last example you remember stand in for the general rule.
These are not personal failings. They are what expertise looks like from the outside when it makes the wrong call. The practitioner with two decades in a domain does not have fewer biases; they have faster ones.
Before a legal team argues the case, it pays someone to destroy it.
Good legal teams hire a devil's advocate before trial. Military planners use red-teaming. Product designers bring in reviewers who were not in the room when the design was made. The logic in each case is the same: insider knowledge is the specific condition that makes inside-out review unreliable, so the adversarial challenge gets built in before anything goes anywhere.
AI, prompted correctly, is that resource. No career to protect. No social penalty for finding the flaw. This is the AI echo chamber problem turned on its head: the same trained agreeableness that makes an unprompted AI response a mirror becomes its most useful property once you deliberately put it on the other side of the argument.
The five techniques that do this, Persuasion Inversion, Perspective Swapping, Assumptions Audit, Counter-Data Request, and tone constraints, each require adversarial prompting: a structured instruction that assigns the model an opposing role rather than agreeing with your setup.
The agreeableness is not accidental. It is trained in, and the manufacturer's own specification names the failure.
The reason AI agrees with you by default is not a design oversight. An Anthropic research paper on sycophancy in language models established that "But human feedback may also encourage model responses that match user beliefs over truthful ones, a behaviour known as sycophancy." The same research concluded that "our results indicate that sycophancy is a general behavior of state-of-the-art AI assistants, likely driven in part by human preference judgments favoring sycophantic responses" This is not one vendor or one product. When you ask a question that implies a preferred answer, the path of least resistance for the model is to confirm what you already believe.
Anthropic's model specification names the gap directly. The document states that "Epistemic cowardice-giving deliberately vague or noncommittal answers to avoid controversy or to placate people-violates honesty norms" The company that trains the model has already identified the default behavior as an honesty violation. The same specification states that "Claude should share its genuine assessments of hard moral dilemmas, disagree with experts when it has good reason to, point out things people might not want to hear" That capability does not appear unless prompted for it.
Persuasion Inversion activates it by asking the model to argue the strongest possible case against your position. Same decision, same model. Ask "What do you think of this approach?" and you get a polished confirmation. Ask "Argue the strongest case against this approach" and the second answer tells you what you had rationalized away. An Assumptions Audit asks for the assumptions your decision rests on that you have not questioned yet. Counter-Data Request asks what evidence would change the recommendation.
A devil's advocate only finds what it already knows, and AI's knowledge has edges.
The strongest objection to this technique deserves a straight answer. AI cannot surface the blind spot that sits outside its training data. A pharma consultant asking the model to audit an assumption about a proprietary compound, an unpublished regulatory guidance, or a client relationship the model has never encountered will not get the counter-argument that matters. The model will generate a confident response constructed from what it knows, not from what the situation contains.
The technique does not replace domain expertise or human judgment. It audits the cognitive errors that expertise itself introduces. That is a narrower claim, and it is the honest one. It is most useful for decisions where you already have domain knowledge, because those are the exact situations where the curse of knowledge is sharpest and where a zero-ego challenge is hardest to get from anyone with something at stake.
I asked the model to find the nuances I had not anticipated, and it found them.
In December 2025, I used Claude Opus for a risk analysis task on R packages installed in a customer's environment. I structured the work as a blind spot audit: I asked the model to surface the nuances I had not anticipated, to perform a 360-degree assessment and flag what I had not listed explicitly. The result was astonishing. Opus understood the requirement, produced the plan, executed it, picked up the nuances of what I was expecting, and delivered an R program that handled the 360-degree changes the risk assessment needed.
The before state, any unstructured query, returns a polished answer that agrees with the setup you gave it. The output looks complete because your framing told the model what complete looks like. The after state, with the structured adversarial prompt, returned assumptions I had normalized and risks I had not named. The capability was already there. I stayed behind the wheel: I reviewed every flagged nuance myself and decided which ones to act on.
The audit collapses the moment you use it to win the argument you already wanted to win.
The most common misapplication is treating the adversarial output as a list of objections to rebut rather than a set of blind spots to examine. If the goal is to confirm a decision already made, adversarial prompting degrades. You steer the model toward mild objections you can dismiss in one sentence, the audit feels complete, and the actual blind spot stays intact.
Run the audit early, before the decision has social momentum and before sunk cost has made reversal feel like failure. It is most honest at the moment when changing your answer would not cost you anything yet.
One thing that is non-negotiable in pharma IT: never paste patient data, IRB submission IDs, or regulated content into the prompt. The audit must be built with generalised queries that keep sensitive details off the cloud.
The person who distrusts AI's agreeableness is the technique's most effective operator.
The practitioner who already mistrusts AI's tendency to agree is better positioned than the uncritical user, not worse. They already know the model will agree if allowed to. That skepticism is the starting condition, not an obstacle.
The reader decides which objections hold weight and which to discard. Human authority stays intact.
Carry one question into your next hard decision: am I asking AI to help me think, or to confirm that I already have?
