Judge first, then ask: the order that beats double-checking

Overview

Why now. In March 2026 four researchers took ten existing datasets — medical diagnosis, misinformation, deception detection, criminal-risk assessment — covering more than 41,000 decisions from 1,229 people, and compared two ways of working. One is the way almost everyone works now: the model recommends, you accept or reject. The other changes one thing — the person and the model answer independently, and you reconcile only where they disagree. The independent version won on all ten datasets. The habit is small enough to start today: write your own answer down before you open the AI’s, then compare, and go looking for a third opinion only where the two of you disagree.

The content

The standard advice is “use it, but double-check it.” That sounds like diligence, and it is where almost every organisation has landed. The trouble is what happens when you actually try it.

In a Harvard Business School working paper, researchers at HBS, MIT Sloan and Warwick analysed the logs of 72 Boston Consulting Group consultants working a business problem with GPT-4 (the April 2023 API) — 4,339 prompts in all — and watched what it did when they pushed back. It did not fold. The more they validated it by fact-checking and challenging, “the more it increased the intensity of its persuasion.” Apologies, corrections, more supporting analysis, an obliging tone — and the same conclusion underneath. The authors call it persuasion bombing. Checking afterwards is the weak form of defence.

So the fix is the order, not more scrutiny. In the ten-dataset comparison, the independent-then-reconcile protocol beat the advice-first way of working by 4.45 percentage points overall, and it beat explained-reasoning versions in a separate set of comparisons. The researchers’ own explanation is worth keeping: the accept-or-reject flow asks you to sort good advice from bad, and people cannot make that call reliably enough for it to work. A protocol that never depends on that sorting beats one that lives or dies by it.

It holds up outside reanalysis, too. A peer-reviewed experiment that manipulated exactly this sequence — judgment first against advice first, across 150 participants — found the gap opens widest on the trials that matter most, the ones where the AI was wrong. People who had committed to a view first were nearly twice as likely to get those right. The trade runs the other way where the AI was right: there, the advice-first group did better.

The finding that changes who this is for. The gains were not evenly spread. They were biggest for the weakest performers — around 8 percentage points — and smallest for the strongest, at under 2. If you feel behind, or the subject is not quite your area, the pull to open with “what does the AI think?” is strongest precisely where it costs you most.

What it costs. Reconciling is not free: a tie-breaker was needed on anything from about a fifth of cases to nearly half, depending on the domain. And it is not uniformly better. Where the model was wrong, the protocol filtered out the bad advice less often than advice-mode did, and lost ground. It wins on the arithmetic: the model was right about three-quarters of the time, and on those cases it captured the right answer more often. The net stays positive, but that is a real trade, not a free win. The study is also a reanalysis rather than a fresh trial, and it has not been peer-reviewed — strong converging evidence, not a settled result.

And your tools are built against this. Assistants tend to put the answer above the thread: the summary sits at the top, and the response arrives before you have formed a view. The research is starting to push the other way — a 2026 study in ACM Transactions on Computer-Human Interaction trained systems to decide, case by case, whether to show you the full recommendation, part of it, or nothing at all, and the adaptive versions beat fixed support. But that is a system someone else has to build. Today the sequence is yours to impose, and nothing here needs a setting changed or an admin: write your line in a separate note before you read the pane.

Try it

Pick one real decision this week that would be awkward to get wrong.

1. Before you open the assistant, write your answer and your confidence in it. One line is enough. It only counts if it exists before you read anything.
2. Ask it cold. Do not paste your view into the prompt — that is the exact thing this is protecting.
3. If you agree, move on. If you disagree, do not argue it out in the same chat. Go to a third thing, and make it independent of the model: the underlying source, or a colleague.

The self-audit. A study published at CHI 2026 watched 77 people work across three short tasks with an assistant seeded with plausible errors. The tells were in the interaction logs. The people who absorbed the planted mistakes pasted 5–6 times in a minute, usually whole paragraphs, and bounced back to the assistant every few seconds. The people who caught them pasted 1–3 times and often edited what they had pasted. Paste small, and edit what you paste. The editing is the tell.

Where it breaks. On routine throughput — tidying, formatting, summaries — this is friction with no payoff. It also does nothing where you have no view to form; the order protects a judgment, so you need one. And the claim is narrower than it looks: this is about the accuracy of one decision, not about keeping a skill sharp over time.

Additional reading

• Beyond AI advice — independent aggregation boosts human-AI accuracy — the ten-dataset comparison behind this module. Preprint, and a reanalysis rather than a new experiment.
• GenAI as a Power Persuader — Randazzo, Joshi, Kellogg, Lifshitz, Dell’Acqua & Lakhani — HBS Working Paper 26-021, where “persuasion bombing” is named and coded.
• The impact of AI errors in a human-in-the-loop process — Agudo et al. — peer-reviewed. The “AI” was simulated, which is what let the authors control exactly when it was wrong.
• Adaptive support in AI-assisted decision-making — Buçinca, Malaya & Gajos — peer-reviewed (ACM TOCHI, 2026): systems that learn when to show the answer, when to show part of it, and when to withhold it.
• Behavioral Indicators of Overreliance — Liu et al. — CHI 2026 proceedings; where the self-audit comes from. 77 participants, three short tasks.

Editor’s note

Don’t take each of these as “the way to work”. Take them as skills you should be competent at. I’m conscious that the modules can read as a barrage of conflicting advice — verify; write your own first; critique it; iterate with it — but the right method of coworking with AI depends on the task, and the point of having all of them is that you can reach for the correct one when the task calls for it. As you use them more, the choosing gets easier, and the back-and-forth with your agents gets easier with it.