There was a time when a number could not get into a decision without first surviving a few questions.
An analyst put a figure in front of you, and before it went any further, you asked where it came from, why they believed it, and what they would say if someone challenged it. This was all part of the job; the grilling was quality control that cost you nothing but the asking.
That habit has quietly fallen away. That figure comes from a machine now, and the instinct to question a person does not transfer to it. AI hands you the figure with more composure than any analyst, and the reflex to push back softens in the face of something so fluent and confident. The irony is that the questions matter more now, not less. A human faces the consequences of citing wrong facts, and thus often puts in the work to avoid them. This doesn’t apply to AI. It will state an invented number in precisely the tone it uses for a sound one, apologise without embarrassment when challenged, and carry on exactly as confidently as before. The one voice in the room that should be questioned hardest is now the one that invites questioning least.
The good news is that you do not need to check every single source yourself to hold an AI report to account. You need to know what to demand, and to demand it consistently enough that the work arrives ready. Here are five questions:
1. Where did this come from, and does the source actually say it?
The simplest question, and the one that separates a figure with a provenance from a figure that merely sounds authoritative. A claim that cannot name where it came from has not been researched; it has been asserted, and an unqualified assertion in a confident sentence is exactly the thing you are trying to catch. The more important half of the question is whether the source genuinely supports the point, because the common failure is not just a broken or made-up link but a real, respectable source attached to a claim it never actually made. A citation that exists is a long way from a citation that holds, and this is a gap consistently overlooked.
2. How sure are we of this claim, and how was that determined?
A claim deserves transparent trust validation you can interrogate, rather than a tone you take at face value. The useful version of trust is not a feeling but a number you can take apart, one built from how good the sources are, whether they corroborate each other, and whether anything contradicts them. When the work is written down and can be recomputed, trustworthiness stops being an impression and becomes something you can audit, query, and, where necessary, dispute.
3. When I read a claim, can I see how much to trust it?
Trustworthiness is only useful if it travels with the sentence it belongs to. A score that sits in an appendix or a footnote is a score nobody consults, and the claim continues to be read as though it were settled. The thing to ask for is the evidence sitting next to the finding itself, so that when you read a strong claim and a shaky one, you can tell which is which without leaving the line you are on.
4. Do any of these sources disagree, and what was done about it?
A claim is only as good as what stands behind it, and the most revealing thing behind it is often a disagreement. When two credible sources conflict, an AI tends to resolve it quietly, averaging or choosing without telling you there was ever a dispute. That buried conflict is usually where the real story sits, and a report worth trusting keeps both sides and flags the tension rather than smoothing it into one tidy answer. Honest uncertainty is worth more than a false consensus.
5. If I read this again next quarter, how much would still hold?
The question that treats research as a living thing rather than a document you finish and file. Some of what you are reading will not survive three months, particularly in a fast-moving field, and the useful instinct is to know which parts are most likely to move and to plan to look again rather than to assume the report will keep indefinitely.
We thought these five questions were worth asking of every claim, so we built a tool that holds every claim to all five and called it SPARK, a Self-verifying, Portable, Agentic Researcher Kit. The point is simple: don’t just take the answer AI gives you, get the receipt. SPARK is what hands you that receipt, the source, the score, the conflicts and the changes, on every claim, so you are not the one left asking. It began as a fix for our own research and grew from there. Read more about it here.
