AI Vocabulary
Traceability: do you see the path, or only the conclusion?
Traceability means being able to reconstruct how a result was obtained: what was consulted, what was retained, what was set aside and on what basis. It differs from explanation: one does not ask the system to justify its reasoning, one asks it to preserve what it traversed.
Why explanation is the wrong requirement
A model asked why it answered as it did will produce a reason, plausible and articulate. Nothing guarantees it describes what actually happened: it produces a justification as it would produce any other text.
Requiring an explanation therefore means asking for an account whose fidelity cannot be checked. That is reassuring and of no evidential value.
Traceability bears on verifiable facts: this document was opened, this source consulted, this option encountered. Those elements exist independently of what the system says about them.
The three elements that make traceability useful
The sources actually consulted, not those that might have been. That is what lets verification bear on what grounded the result.
The points at which several positions were defensible, and which was adopted. That is the equivalent of an associate's footnote, the one that tells the reviewer where to look.
What was set aside, and why. Distinguishing not relevant from not found changes everything for whoever signs.
A distinction that changes what you ask of a provider
Many systems now display the sources used, as links beneath the answer. That is useful and covers only a third of the requirement.
What is missing is the rest of the path: sources consulted and then set aside, points where several readings were possible, the moment the system stopped looking. Those elements exist at the time of production and are discarded on delivery, because what is presented is finished work rather than traceable work.
That is a design choice, not a technical impossibility — and therefore a legitimate request to put to a provider.
What it changes for the reviewer
A deliverable without traceability imposes uniform attention: nothing indicates where to concentrate, so everything is read at the same level, which is the opposite of what a review must do.
A traced deliverable restores a map: here is what was checked, here are the points of uncertainty. Review then regains the shape it had with human work, whose hesitations were visible in the writing.
What it does not solve
Traceability makes nothing accurate. It shows where to look, it does not say whether the result is right — that is the reviewer's work, and it remains entire.
It has a cost: it produces volume, and volume nobody consults is a cost without return. It has value only if its format allows a precise piece of information to be found in minutes.
And it does not replace professional traceability, the record of who approved what. The two are complementary and belong to different responsibilities.
Why it matters to a lawyer
Because it is what separates supervising from believing. Signing work whose path cannot be retraced means extending trust to it, which is a different operation from supervision.
The question to put to a provider is therefore precise: what can I see of the path, and not only of the conclusion?