Accountability migrates from checking the work to vouching for the machine
Licensing and accountability survive the commoditisation of expertise in changed form — the professional stops checking individual outputs and ends up vouching for the system that produced them.
The claim
Accountability and licensing are usually offered as the part of professional work that AI cannot touch. Someone has to be answerable, and a model cannot be struck off. That is correct as far as it goes, and it is commonly taken to mean the professional’s work becomes checking the machine’s output before putting their name to it.
The checking does not survive.
Why it does not
Verification is only sustained while errors are found often enough to justify looking. As the tool’s reliability rises, the rate at which checking is rewarded falls, and attention follows the reward. A reviewer who finds a material error once in a hundred reviews is not performing the same task as one who finds errors in a third of them; they are performing a formality, and they will eventually stop performing it whatever the policy says. This is not a lapse by unusually lazy practitioners. It is the predictable behaviour of competent people working against a very high base rate of correctness.
What remains is attestation. The professional’s name still attaches to the work, the liability still lands on them, but what they are actually putting their name to is the system’s fitness to produce that class of output rather than their own inspection of this instance.
Why this matters rather than being a technicality
The two are not equivalent obligations and the difference is where the governance work now sits. Vouching for a system is a claim about its selection, its configuration, its inputs and its monitored performance over time — things a firm can actually build evidence for. Vouching for an output you no longer read is an unsupported claim wearing the clothes of the old one.
Firms that notice the shift can build the second kind of evidence deliberately. Firms that do not will keep describing their process as review long after the reviewing has stopped, which is a considerably worse position than either.
The durable human contribution described in The relationship is the product and Defensibility lives in what AI can’t access is unaffected by this; both locate accountability as the thing clients pay for without specifying what discharging it involves. This note is an attempt at that specification.