Product

Verification Must Not Be an Option

Leaving verification to the user’s discretion is silently transferring to them a responsibility that the whole culture of the craft refuses to rest on one person.

In most current legal AI tools, verification is an option left to the user. The tool produces, and it falls to whoever receives the output to decide whether to verify, how, to what degree. The implicit promise is reassuring, the user stays in control, but at the scale of an organization, this way of presenting things is a silent transfer of responsibility. And it is this transfer, more than any technical defect, that poses the problem.

Why the option does not hold

Because a user, in practice, does not verify systematically. Not out of negligence, but out of cognitive load. Under time pressure, under the flow of matters, under the accumulation of small decisions, verification slides toward the implicit, toward the partial, toward “it looks correct.” It is human, it is universal, and that is precisely why no serious legal organization grounds its reliability on individual vigilance alone. To count on everyone always verifying is to count on a constancy the nature of the work makes impossible.

Verification left to each person’s discretion always ends up giving way under the load. It is not negligence, it is fatigue.

There is a cruel asymmetry between the cost and the benefit of verification, which explains why it gives way. Verifying costs a definite, immediate effort, every time; not verifying costs nothing most of the time, because most of the time the output is correct. The user who skips a verification is rewarded in ninety-nine cases, and punished in the hundredth, which comes rarely and late. This incentive structure mechanically pushes toward slackening, regardless of each person’s professional conscience. You cannot ask a dispositif to hold against the incentive it creates itself.

One must see that this is not a problem better willingness would solve. You can train users, remind them of the importance of verifying, sensitize them to the risks: none of this holds against the reality of a busy day. Vigilance is a resource that depletes, and a dispositif that depends on it is a dispositif that will give way, not always, but often enough that one cannot rely on it. The problem is not the quality of the people; it is that they are asked for a constancy no one can hold.

What the craft’s culture has always known

The whole culture of the craft is built, for a long time, on the opposite idea: verification must be procedural, integrated, redundant, independent of each individual’s will to do well. The second look, the review by a senior, the validation by a partner, the cross-check between colleagues are not marks of distrust toward people; they are dispositifs that shelter reliability from individual failures, by never resting it on one person. This culture is the fruit of long experience of what can go wrong when a single look suffices.

This culture is poorly honored by current AI tools, which restore in practice what the craft had patiently set aside, namely a reliability suspended on the vigilance of one. By leaving verification optional, they return legal work to a stage prior to all its procedural guarantees, the one where the quality of the deliverable depended on an individual’s care and nothing else. It is a regression no one decided, but that the design of the tools imposes silently.

The craft spent a century no longer depending on a single look. Current AI tools bring it back.

This regression is all the more paradoxical in that AI is precisely the tool that produces faster, hence more to verify. By increasing the volume of outputs while leaving verification optional, current tools worsen the problem they claim to solve: they pour onto the user a mass of production that would need checking, at the very moment they remove the procedural guarantees that made that checking tenable. The more AI produces, the more optional verification becomes untenable, and the more urgent it becomes to make it structural.

At MAX, this observation translates into an explicit principle: verification must not be an option, it must be structural. The verification steps are not proposed to the user, they are triggered by the architecture, at defined moments, on defined criteria, and cannot be bypassed without leaving a trace. This translates into concrete mechanisms: automatic verification loops on sensitive outputs, human validation points triggered by the nature of the matter and not by the user’s will, cross-checks between sources, between steps, between versions, systematic trace of what was verified, by whom, how, and of what was not.

The result of this choice is that at any moment, the organization can document the level of verification applied to any deliverable, without having to hope that everyone did their part conscientiously. Verification ceases to be a promise to become an observed fact, inscribed in the architecture. It is a difference of nature with tools that merely recommend caution: recommending caution shifts the burden onto the individual; inscribing it in the dispositif removes it from their shoulders.

Recommending caution shifts the burden onto the individual. Inscribing it in the dispositif removes it from them.

This inscription in the dispositif also changes the nature of what an organization can promise its clients. As long as verification is optional, a firm can only guarantee its good intention: it hopes its teams verify, but it cannot prove it. The moment verification is structural and traced, it can affirm, with proof, the level of control applied to each deliverable. Verification ceases to be a matter of trust in people to become a demonstrable property of the organization, which is exactly what a demanding client expects, and what the craft’s culture has always sought to guarantee.

There is, in this stance, a form of respect for the user that may seem counterintuitive. Making verification structural is not distrusting the professional, it is relieving them of a burden no one should carry alone. It is recognizing that, even excellent, even conscientious, they work under constraint of time and attention, and that the role of a good infrastructure is to protect them from their own moments of fatigue, not to exploit them. Structural verification takes nothing from their competence; it surrounds it with what makes it tenable over time.

This is how this principle aligns AI use with the real professional culture, in which verification is a value and not a chore. It makes serene adoption possible on sensitive subjects, because responsibility does not rest on the isolated user’s vigilance alone. The systems that will survive deployment will not be those that ask to verify better; they will be those that have made verification impossible to remove.

The systems that will survive deployment will not ask to verify better. They will make verification impossible to remove.