> ## Content Index > Fetch the complete content index at: https://piotrmechlinski.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI Needs Verifiers. We Are Training Ourselves Not to Think. - URL: https://piotrmechlinski.com/writing/ai-needs-verifiers/ - Published: 2025-11-14T23:00:00.000Z - Updated: 2026-09-08T07:54:09.000Z - Description: AI changes knowledge work from producing answers to supervising them. Make verification a skilled, named responsibility before judgment fades into a final check. - Author: Piotr Mechlinski - Tags: Judgment AI can produce a plausible answer before you have finished forming the question. That speed feels like progress. It can also remove the part of the work where judgment develops. You no longer struggle with the blank page. You receive a polished draft, scan it for obvious errors and move it onward. The document looks finished. Your thinking may have barely started. This matters most in the work that carries consequences: an investment case, a customer promise, a hiring decision, a board recommendation or a public claim. Make verification a real job with time, skill and authority. ## AI moves the work upstream A 2025 study from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers who used generative AI at least weekly. The findings were self-reported, so they do not prove that AI causes weaker thinking. They do reveal an important shift. People described spending less effort on some critical-thinking tasks and more effort on checking, integrating and supervising AI output. Higher confidence in the AI was associated with less critical-thinking effort. Higher confidence in their own ability was associated with more. The researchers call the new role [task stewardship](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/01/lee%5F2025%5Fai%5Fcritical%5Fthinking%5Fsurvey.pdf?ref=piotrmechlinski.com). I prefer a blunter word. Verifier. A verifier proofreads and tests the decision behind the draft. The verifier knows what the work is meant to achieve, which evidence deserves trust, where the model is likely to fail and who carries the consequence when the answer is wrong. This is skilled work. Treating it as the final five-minute check guarantees shallow review. ## A polished draft can hide an empty decision Imagine an executive team considering a new AI service for customers. A model can summarize market reports, compare competitors, draft the business case and prepare the slides. Every artifact may be useful. Together, they can create a dangerous illusion of completeness. The model does not know which customer pain the company is willing to own. It cannot decide whether a forecast deserves capital. It cannot accept the reputational cost of a bad promise. Those decisions remain with the people in the room. If nobody writes down the assumptions, checks the original evidence and challenges the recommendation, the team has produced a presentation without completing the decision. The risk grows with fluency. Awkward output invites scrutiny. Smooth output invites trust. NIST describes this as automation bias: people may defer too much to automated systems or assume their output has higher quality than other sources. Reliability can make this worse because repeated success lowers attention just before an unusual failure arrives. The [NIST Generative AI Profile](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf?ref=piotrmechlinski.com) therefore treats the human and the system as one operating configuration, with risks created by the way they work together. Your control system must examine the people, workflow and model together. ## Use a verification contract For any AI-assisted decision with material consequences, assign one named person to answer five questions before the work moves: 1. **What is the claim?** State the recommendation in one sentence. A verifier cannot test a cloud of polished language. 2. **What is the evidence?** Open the original sources. Check dates, definitions, denominators and missing context. A citation supplied by a model is a lead until a human verifies it. 3. **What reasoning connects them?** Name the assumptions between evidence and recommendation. Ask what else could explain the same facts. 4. **What happens if this is wrong?** Set the review depth according to the consequence. A meeting summary and a credit decision need different controls. 5. **Who owns the release?** Record the person who can approve, reject or return the work. Accountability disappears when everyone is merely reviewing. This contract adds friction where friction has value. It leaves low-risk drafting fast and makes consequential work deliberate. It also changes training. Employees need practice forming an initial view before asking for an answer, locating primary evidence, testing a counterargument and explaining the final choice in their own words. Otherwise, the organization trains people to recognize acceptable output while their ability to create and challenge it fades. Review capacity also needs a budget. If AI multiplies the volume of proposals, reports and customer messages, the organization must decide which output deserves human attention. Sending five times more material through the same approval queue turns accountability into scanning. Use consequence to set the depth. Low-risk internal drafts can move quickly. Public claims, regulated decisions and capital commitments need original sources, recorded reasoning and an accountable release owner. ## Keep authorship where judgment lives Writing forces you to choose the claim, order the evidence and notice the gap in your own reasoning. AI can help with each step. Ask it to expose contradictions, generate objections or test whether a reader could misread your conclusion. Those uses keep the human inside the reasoning loop. The executive decision is simple: choose which parts of the workflow must remain actively authored by a person. For a material recommendation, keep the purpose, evidence judgment, final claim and ownership human. Let AI compress research, suggest structure and challenge the draft. Then require the accountable person to explain the decision without the document in front of them. If they cannot, the work is still borrowed. Your move.