AI Will Replace the Worker. Then Hire the Witness.

The machine makes the decision. A person is kept to answer for it.

I work in law, and a recent paper of mine takes up a narrow question in evidence law: what happens when a machine produces a judgment that no human being has adopted. Working through it, I kept hitting a larger version of the same problem, one that reaches well beyond the courtroom. The legal gap points at an economic one. AI can remove the person who produces an answer while creating demand for someone who has to certify, adopt, reconstruct, or defend it.

The Missing Owner of the Decision

A human expert has traditionally done two things at once. The expert establishes that a method is reliable and adopts the resulting conclusion as their own opinion. Method and judgment arrive together, in one person who stands behind both.

Machine-generated evidence pulls those functions apart. An engineer can establish that a system is reliable, properly validated, and correctly operated, and still decline to say that its conclusion in the particular case is correct. The engineer vouches for the system. The machine produces the judgment. No one necessarily owns it.

My paper argued that when machine-generated evidence expresses a case-specific judgment, the party offering it should have to produce either a qualified person who genuinely adopts that judgment or a record detailed enough to reconstruct how it was produced. That is an evidentiary proposal. But the same gap will open across the economy, wherever production becomes synthetic while responsibility stays human.

The Answerability Economy

Most arguments about AI and employment ask whether a human is still required to perform a task, and increasingly the honest answer is no. AI may read the scan, assess the claim, screen the applicant, price the credit risk, flag the suspect, draft the audit, or determine eligibility for a benefit, and do it faster and more consistently than the professionals who used to.

But productive necessity is only one reason institutions employ people. They also employ people because decisions have to be authorized, certified, explained, appealed, attributed, and defended. None of those functions improves the machine’s output; they make the output usable inside legal and administrative systems that still assign rights and responsibilities through human beings and the organizations they control. The AI does not need those workers, but the institution does.

This is the answerability economy: a layer of white-collar work organized around converting machine output into decisions an institution can stand behind. It is not the same thing as checking whether the machine is right.[1] Verification asks whether an output is correct. Answerability asks who is entitled to adopt it, who must defend it, how it can be contested, and where its consequences land. An output can be perfectly verifiable and still have no authorized owner, and a person can be required to stand behind a judgment whose correctness no one can fully confirm. What the institution needs is not another check on accuracy but a recognized point of attribution.

Its roles are already taking shape. There are people who review machine judgments and assume responsibility for them; auditors who test whether a system performs and catch it failing systematically; staff who preserve the inputs, model versions, and prompts behind an output; specialists who reconstruct how a particular result was produced; officers who hear appeals of automated decisions; investigators who assign responsibility after a failure; professionals who certify systems for specified uses; and expert witnesses who explain and defend machine conclusions in court. These workers occupy the space between machine execution and institutional consequence. The system produces the answer; the human clears it for use.

White-Collar Work Moves Toward Permission

The pattern repeats across nearly every consequential sector. In medicine, AI reads the images and recommends the treatment while a physician reviews and adopts the conclusions that carry the hard cases. In insurance, automated systems resolve routine claims while human reviewers certify the high-consequence denials and handle the contested classifications. In employment, models rank applicants and flag dismissals while someone decides whether those conclusions can be legally defended. In banking, AI underwrites while compliance staff reconstruct adverse decisions and confirm that prohibited factors did not drive the outcome. In government, automated systems calculate benefits, taxes, and enforcement priorities while administrative reviewers keep a channel open for appeal and attribution. In litigation, the machine produces the forensic conclusion while one expert establishes the system and another decides whether to stand behind the result.

The profession does not survive as a single role. Its function splits: execution moves into the machine, authorization stays attached to the institution, and human labor gathers along the boundary between them. White-collar work shifts from production toward permission.

Accuracy Does Not Produce Authority

A system more accurate than any available human can still generate this work, because accuracy is not the same thing as authority. Institutions have to answer the individual case in which the system may have failed. Fairness is the clearest reason, though not the only one; due process, professional licensing, insurance, administrative finality, and the plain need to place liability all generate the same demand. A person denied a job, their liberty, medical coverage, or a public benefit does not experience the model’s overall accuracy rate; they experience one decision, and the law gives them a way to challenge and reconstruct it. That is why automation can produce jobs that look economically unnecessary. The capability to reach the answer already exists. The extra worker exists because consequential authority cannot yet travel directly from a machine output to a legal effect. The job is produced by an institutional constraint, not a productive one.

Modern economies already run large versions of this layer. Compliance officers, auditors, inspectors, claims reviewers, licensing authorities, and administrative judges do not manufacture the underlying goods; they make complex systems governable, legible, and contestable. AI expands this function even as it thins the labor that performs it. Every automated system that approaches consequential authority creates a surrounding demand for certification, provenance, appeal, and the allocation of liability. The more execution moves into machines, the more an institution has to specify what turns an output into an authorized decision.

Legitimacy is labor-intensive.

Real Review, or Review in Name Only

Not every human placed inside an automated process will actually exercise judgment. Organizations will be tempted to build oversight that exists only on paper: an employee receives the machine’s recommendation, clicks approve, and becomes the official decision-maker the institution can point to as proof that a human stayed in control.

Substantive adoption takes more than a signature. It requires understanding the material inputs, evaluating the conclusion, weighing plausible alternatives, and being willing to present the judgment as your own. A worker without the authority, competence, time, or access to the underlying record to reject the machine cannot supply that; they are an interface placed in front of a decision formed elsewhere. The title stays human while the judgment has already moved upstream.

This distinction decides whether answerability work becomes a genuine profession or just a mechanism for pushing liability downward. Some workers will hold real authority to overrule a machine conclusion. Others will absorb responsibility for systems they do not control, which is not oversight at all, only a place to send the blame.

More Job Titles Is Not More Jobs

The answerability economy can multiply categories of white-collar work without producing enough employment to replace what automation removes. One adopting professional can supervise thousands of machine-generated decisions. One audit team can govern systems that replaced whole departments. Appeal officers handle only the small fraction of automated decisions that are ever formally contested. The layer can grow in importance, spending, and authority while employing fewer and fewer people. The titles proliferate while total employment contracts.

This is not the comfortable claim that AI will create as many jobs as it destroys. It is a narrower point about where some of the new work comes from and why institutions keep employing people after machines can perform the underlying tasks. Production can keep growing while the income it used to distribute shrinks, and labor loses its position as the main channel through which income, status, and a recognized place get handed out. Answerability work can slow that displacement, but it will not reverse it.

The Jobs of the Transition

These jobs belong to the interval between two orders. In the old order, humans perform the work, form the judgment, and carry responsibility for the result. In the emerging order, machines perform the work and increasingly form the operative judgments, while responsibility stays attached to human beings and the legal entities they run. A more distant order might recognize machines themselves as holders of authority, assets, duties, and liability. If that arrives, some of this layer could thin. But legal personhood is not a solvent: corporations have held it for centuries and still need officers, auditors, and witnesses to answer for them. Machine personhood would more likely move the answering around than end it.

During the transition, courts will keep demanding witnesses, regulators will keep demanding accountable officers, citizens will keep demanding a way to appeal, insurers will keep demanding someone who can assume liability, and organizations will keep demanding the signature that converts machine output into authorized action. AI will remove the worker who made the decision. Law, administration, and fairness will create another worker whose job is to make that decision answerable. That worker will not operate the machine. That worker will stand between the machine and its consequences.

The machine makes the decision. A person is kept to answer for it.

Notes

[1] The nearest economic account is Christian Catalini, Xiang Hui, and Jane Wu, Some Simple Economics of AGI (Feb. 24, 2026), which argues that as machine execution becomes cheap, the binding constraint shifts to human verification, with rents migrating to validation, provenance, and liability underwriting. Answerability, as used here, is the adjacent but distinct function: not confirming that an output is correct, but owning and defending it once it acquires institutional effect.