When AI Makes Proving Ability More Expensive

A market for demonstrating ability

A polished cover letter can help an employer evaluate an applicant because producing it takes effort. But its meaning changes when AI makes polished writing easier to imitate. Vaccari’s preprint, posted on November 2, 2026, asks how providers’ choices of technology and prices affect that process.

The model contains people with two fixed ability levels, an evaluator, and AI providers. People know their own ability; the evaluator sees their submitted signals but cannot observe subscriptions, AI input, or production methods. Providers know the population’s ability distribution, not each customer’s type.

People can produce an observable “artifact,” such as an application, and harder-to-substitute evidence, such as a supervised work trial. Institutional capacity limits how much of the second signal is available. Providers choose the editing effort their tools require from each ability group and, in the baseline model, how much production they can assist. They then charge membership and usage fees.

Why easier production can mean more work

The central mechanism is imitation. AI reduces production effort for both groups, but the model assumes that lower-ability people save more effort per unit of assistance. They can therefore find a subscription worthwhile at a less demanding artifact than higher-ability people do.

Over some output ranges, improving an artifact becomes cheaper for an assisted lower-ability person than for a higher-ability person still working manually. A slightly better submission then need not credibly demonstrate higher ability. Recognition may require a much larger improvement.

Vaccari shows that profitable AI provision can consequently reduce welfare even when ability remains fully revealed. Assistance saves resources on each unit produced, but the additional signaling costs needed to deter imitation and the technology’s development costs together can outweigh those production savings.

The model also permits “pooling”: both groups submit the same signals and receive the same assessment. In partial pooling, all higher-ability people and some lower-ability people share a signal, while the remaining lower-ability people identify themselves through the basic submission.

Providers can profit from potential imitators

At a profitable optimal monopoly offer that fully reveals ability, higher-ability people purchase assistance. Lower-ability people do not buy it in the resulting outcome, but their ability to imitate still constrains what stronger applicants must produce.

Improving assistance for lower-ability people can therefore increase higher-ability customers’ purchases. In an extension with two established providers choosing improvements before prices, Vaccari constructs an equilibrium where one improves assistance for both groups and earns positive net profit, while its rival retains its baseline technology.

Both the improved and baseline outcomes reveal ability perfectly, yet the improvements lower welfare relative to retaining the existing tools. Importantly, that example still uses fewer total signaling and improvement resources than manual signaling. Its loss is relative to existing AI, not to having no paid assistance.

Prices matter beyond the final bill. A membership fee affects when adoption becomes worthwhile; a usage price affects the cost of producing a stronger signal. The model allows profitable below-cost usage, including free usage financed by membership revenue.

Competition and verification are not automatic fixes

For two or more providers with identical, already-developed technologies, the paper establishes a pricing equilibrium with no membership fee and usage priced at compute cost. Ability is fully revealed and providers earn zero operating profit.

That does not guarantee cheaper signaling overall. When monopoly provision of the same technology also separates ability groups, competitive pricing requires more signaling resources because imitation becomes cheaper. Development costs can also sustain a sole provider: an additional entrant would lose money under the specified competitive pricing outcome. This is an equilibrium construction, not a claim that every market has that outcome.

Hard evidence offers another route to recognition. However, more available verification can change whom providers serve. In constructed examples, a higher externally imposed minimum capacity shifts provision from assisting higher-ability people in separation to assisting lower-ability people in pooling, worsening assessment accuracy. This is not a voluntary expansion knowingly chosen to worsen accuracy.

What the findings establish

These are conditional theoretical results. Ability is fixed, individual AI adoption is hidden, and artifact quality has no direct productive benefit in the welfare calculation. The analysis restricts signaling to finite-support equilibria and uses additional selection rules. Subscription at exact cost indifference matters especially for profitable shared outcomes.

Within those assumptions, the lesson is that productivity, information, and welfare are different outcomes. Accurate assessments can require wasteful effort, while less informative signaling can sometimes save enough resources to improve welfare. Evaluating AI assistance therefore requires accounting for how prices and technology change what people must produce—not merely how cheaply they can produce it.