Everyone I meet is now an AI expert. Most of them mean they are good at ChatGPT. That is not nothing. Getting a useful answer out of these tools takes some skill, and plenty of people still can't do it. But being good at prompting is the AI equivalent of being good at typing. It is visible, it is easy to show off, and it tells you almost nothing about whether the work that comes out can be trusted.
In July I wrote that "human in the loop" is only as good as the human [3]. Put a competent, sceptical professional in the loop and it catches what the machine missed. Put someone there who trusts the confident output, and the loop does something worse than nothing: it launders the error. This piece is about how you tell which kind of human you've got.
The polished answer is the problem
We usually judge competence by looking at the work. With AI, judging by how polished the work looks stops working, and there is now good evidence for why.
The best study I know of is a preregistered experiment with 758 Boston Consulting Group consultants, run by researchers from Harvard, Wharton, MIT and Warwick using GPT-4 as it was in April 2023, and published this year in Organization Science [1]. The consultants were split into two groups. About half worked on tasks GPT-4 handled well. Those using it completed 12.2% more tasks, worked through them 25.1% more quickly on the timed measure and produced significantly better work [1].
The other half got a business problem deliberately chosen to sit just outside what the AI could do well. Without AI, 84.5% of them got the right answer. With AI, the two AI groups managed 70.6% and 60% [1]. The wrong answers also looked better. Graders who had not been told the correct answer rated the AI-assisted recommendations as more coherent and more persuasive than the unassisted ones, whether the recommendation was right or wrong [1].
The prompt engineering overview the study tested didn't protect against it either. One AI group had been given that overview before starting. On this task its correct-answer rate was 24.5 percentage points below the group working without AI, against 13.9 for the group using AI without the overview [1]. The gap between the two AI groups is not statistically strong, so the safe reading is narrower: the overview did not help here. (On the tasks inside the AI's limits, the same overview did improve the quality of the work.)
Inside its limits, AI lifts average work until it looks like good work. Outside them, it makes wrong answers more persuasive. That is why fluency is such a poor guide to competence. The thing we instinctively use to judge someone, how good their output looks, is exactly what AI has made unreliable.
What a prompt can't supply
At the core of competent AI use is catching what is wrong with the output before anyone relies on it. That isn't all of it. Choosing when AI is the right tool, handling client data within the rules and telling clients what you've done matter too. But catching errors is the part that the output alone can't show you.
It takes two kinds of knowledge, and you need both. The first is knowing your field: you can rarely spot a wrong answer unless you know what a right one looks like, and there is no prompt for that. The second is knowing how AI fails. AI-assisted wrong answers can read as more persuasive than unassisted work [1], and RICS lists erroneous output and bias among the risks its members must understand [2]. These systems also handle some tasks well and similar-looking ones badly, and you often can't see the boundary from the output alone. The researchers called it a "jagged frontier" [1].
Behind both sits the point I made in July. The machine never holds the stamp. Whoever signs, sends or relies on the output answers for it [3].
So there are two separate things to measure: how well someone drives the tool, and how well they can judge what it gives them. Plot one against the other and you get four rough types of user.
- The Bystander isn't really using AI. Low risk, and little gained.
- The Sceptical Expert knows the field and is unimpressed by the tool. They have half of what's needed. The half they may be missing, how AI fails, can be learned, but it has to be learned.
- The Confident Amateur is fast, polished and prolific, and can't tell a good answer from a plausible one. It is the easiest type to mistake for the best, because the output looks the same. It is also where the risk sits.
- The Accountable Practitioner uses AI well and can defend every output. That is the target.
Tool skill moves you sideways. Only judgement moves you up. In the framework I'm drafting, the step that matters is the move into the top half, and the way through is meant to be showing that you can catch what's wrong, not writing better prompts.
The thing we instinctively use to judge someone, how good their output looks, is exactly what AI has made unreliable.
Don't ask. Test.
My first draft of this framework had ten interview questions. A reviewer pointed out the flaw, and it was a good one. A framework arguing that fluency isn't competence was scoring people on how fluently they described their checking. Anyone can rehearse a good story about catching an error.
So the answer is a work sample. Take an AI-generated piece of work from the person's own field. In construction, that might be a summary of the NEC4 compensation event procedure. Plant known errors in it: an invented reference, a real clause with the wrong time period, a total that doesn't add up, a correct fact leading to a wrong conclusion, a material risk left out, a contested point stated as settled. Then ask them to review it the way they would before relying on it.
Score what they catch, and whether they can say what is wrong. Count false alarms against them, because otherwise the test rewards blanket suspicion, which isn't judgement either. Weight by severity, so that missing an error that would cause harm costs more than missing a cosmetic one. Also leave in at least one genuine error the model made by itself, and don't say how many there are. Otherwise you risk measuring skill at hunting for planted errors rather than at reviewing real work.
The aim is a test that tool skill alone can't pass. Whether it achieves that, and whether it can tell someone who understands how AI fails from someone who is simply a careful checker, is what the trial has to show.
Why this matters now
RICS has written the stamp into the rulebook. Under the RICS professional standard on responsible AI use, in force since 9 March 2026, where an AI output has a material impact on a surveying service, the member must use professional judgement to decide whether it is reliable. RICS defines that judgement as knowledge, skills, experience and professional scepticism. The decision must be recorded in writing, with the assumptions, the concerns and the reasons for them, and a conclusion on whether the output can reasonably be used. If it can't, the client must be told in writing. It must be prepared by, or under the supervision of, an appropriately qualified named surveyor who accepts responsibility for it. Where AI produces outputs automatically or in high volume, RICS accepts that a written decision on each one generally isn't necessary or proportionate, and requires regular random dip-samples instead. The firm remains accountable for each output [2].
But a written reliability decision is only as good as the person making it. The standard requires members to have the knowledge and to record their judgement, but it doesn't prescribe any test of whether the named surveyor can catch the errors. Without one, the written decision risks becoming what I called in July "theatre with a name on it" [3]. Once it has been validated in a trial, a test of what someone catches could be one piece of evidence that the named surveyor can carry that responsibility.
Where this stands
This is a draft framework, not a finished product. It hasn't been tested on anyone yet, the scoring thresholds don't exist, and it shouldn't be used to hire, promote or appraise anyone until it has been through a proper trial with independent markers. It isn't about age or seniority either: the test measures what someone catches, not who they are.
The next step is to build the first work sample, in construction, and run it with a small group of people across a range of roles. Then I'll see whether the way people rate themselves predicts what they catch. I suspect it won't, but that is exactly the kind of claim the test exists to check. At Synaform this belongs with our Advice work: helping clients decide, on evidence, where AI is used and who can stand behind what it produces. For transparency: the framework is my own work in progress, and I may use it in paid advisory work.
Almost anyone can now get a good answer out of AI. What matters is whether they can tell when it has given them a bad one.
Sources
- [1] Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz, H., Kellogg, K.C., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K.R. (2026) "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality", Organization Science, 37(2), pp. 403-423, published online 11 March 2026. https://doi.org/10.1287/orsc.2025.21838 (open access, CC BY 4.0). Locators: abstract (758 subjects, 12.2% / 25.1% more quickly); s.3 (two non-overlapping groups, 385 and 373; GPT-4 as at end of April 2023); s.4.1 and Table 4 (overview improved quality inside the frontier); s.4.2 and Table 7 (84.5% vs 70.6% and 60%; 24.5 vs 13.9 points below the no-AI group; difference between the two AI groups only marginally significant); Table 9 and Fig. 6 (higher coherence and persuasiveness ratings whether correct or incorrect; graders not told the correct solution).
- [2] RICS (2025) Responsible use of artificial intelligence in surveying practice, RICS professional standard, global, 1st edition, September 2025, effective from 9 March 2026. Locators: s.1.2 (scope, material impact); s.2 (baseline knowledge: limitations and failure modes, erroneous output, bias, data risks); s.4.2 (professional judgement = knowledge, skills, experience and professional scepticism; written decision covering assumptions, concerns and the reason for each, and a conclusion; prepared by or under an appropriately qualified and named surveyor; for automated or high-volume outputs, individual decisions generally neither necessary nor proportionate, randomised dip-samples at regular intervals, firm accountable for each output). https://www.rics.org/content/dam/ricsglobal/documents/standards/Responsible-use-of-artificial-intelligence-in-surveying-practice_September-2025.pdf
- 3] Lohan, J. (2026) "The Machine Never Holds the Stamp", written July 2026; published on this site at [/insights/the-machine-never-holds-the-stamp.