📈 economics

Delegation and Verification Under AI

This paper models how rational workers' delegation and verification decisions under AI, driven by outcome-based institutional incentives, create phase transitions that can paradoxically degrade overall institutional quality by causing some workers to over-delegate, thereby amplifying disparities between those with high and low verification reliability.

Original authors: Lingxiao Huang, Wenyang Xiao, Nisheeth K. Vishnoi

Published 2026-03-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Lingxiao Huang, Wenyang Xiao, Nisheeth K. Vishnoi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Smart Assistant" Trap

Imagine you are a Chef working in a high-end restaurant (the Institution). Your job is to cook perfect meals for customers.

In the old days, you did everything yourself: chopping, sautéing, plating. You were responsible for every bite. If the food was bad, the restaurant lost money and reputation.

Now, you get a Robot Sous-Chef (the AI). The Robot can chop vegetables and flip burgers incredibly fast and cheaply. The restaurant boss says, "Use the Robot! It's efficient!"

But here is the catch: The Robot isn't perfect. Sometimes it burns the steak or forgets the salt. If you just let the Robot cook and serve the food without checking, the customers might get sick.

This paper asks a simple but profound question: How does the introduction of this Robot change the value of the Human Chef?

The surprising answer is: It depends entirely on how good the Human Chef is at checking the Robot's work.

The Three Ways to Work

The paper models three ways a worker (the Chef) can handle a task:

The Old School (Manual Work): You ignore the Robot. You chop, cook, and plate everything yourself. It's slow and tiring, but you know exactly what's in the dish.
The "Lazy" Way (Pure Delegation): You let the Robot do everything and just serve the food. You save energy, but if the Robot messes up, you serve a burnt steak.
The "Smart" Way (Verified Delegation): You let the Robot do the heavy lifting, but you taste-test every dish (Verification). If it's bad, you fix it. If it's good, you serve it. This takes extra effort (tasting), but it's usually the best balance.

The Core Conflict: You vs. The Boss

The paper highlights a fundamental mismatch between what is good for You (the Worker) and what is good for the Restaurant (the Institution).

Your Goal: You want to maximize your own comfort. You want to do as little work as possible while still getting paid. If the Robot is fast, you might be tempted to just let it cook everything (Pure Delegation) to save your energy, even if there's a small risk of a mistake.
The Boss's Goal: The Boss cares about the quality of the meal and the reputation of the restaurant. They don't care how tired you are; they care if the customer gets a perfect meal.

The Problem: Sometimes, what is "rational" for you (saving energy by trusting the Robot too much) is disastrous for the Boss.

The "Phase Transition": A Tipping Point

The most fascinating part of the paper is the concept of Phase Transitions.

Think of verification ability (how good you are at spotting Robot errors) as a dial on a machine.

If your dial is set to Low, you are too lazy to check the Robot. You let it cook everything. The food is often bad.
If you turn the dial up just a tiny bit to Medium, you start checking. Suddenly, the food becomes perfect.
But here is the twist: If you turn the dial up too high, or if the Robot gets too good, you might suddenly decide, "Wow, the Robot is so good, I don't need to check at all!" and you slide back into the "Lazy" mode.

The paper shows that small changes in your ability to verify the AI can cause sudden, sharp jumps in your behavior. You don't slowly become more careful; you suddenly switch from "Trust the Robot blindly" to "Check everything" or vice versa.

The "Verification Amplifier" Effect

This is the paper's main conclusion: AI acts as a filter that amplifies the gap between good and bad workers.

The "Super-Verifiers" (High Skill): These are workers who are naturally good at spotting errors. When they get the Robot, they use it to speed up their work and they double-check the results. They become super-productive. The restaurant loves them.
The "Weak-Verifiers" (Low Skill): These are workers who aren't great at spotting errors. When they get the Robot, they think, "I'll just let the Robot do it all." They stop checking. Because they aren't good at catching mistakes, the Robot's errors slip through. The food gets ruined. The restaurant fires them.

Crucially: This happens even if the Robot is actually better than before. The Robot didn't get worse; the system just exposed the workers who couldn't verify its work.

Real-World Example: The Doctor and the AI

Imagine a doctor diagnosing a patient.

Without AI: The doctor reads the X-ray and makes a diagnosis.
With AI: The AI suggests a diagnosis instantly.
The Doctor with High Verification Skill: Reads the AI's suggestion, spots a subtle error the AI missed, corrects it, and saves the patient. They are now faster and more accurate.
The Doctor with Low Verification Skill: Trusts the AI blindly because they are tired. The AI makes a rare mistake. The doctor misses it. The patient suffers.

The paper argues that in the future, the most valuable skill won't be "doing the task" (execution); it will be "checking the AI" (verification).

What Can Be Done? (Interventions)

The paper suggests two ways to fix this:

Train the Workers: Instead of just teaching them how to use the AI, teach them how to critique the AI. If a worker gets better at spotting errors, they stop being lazy and start being efficient.
Change the Rules: The restaurant (Institution) needs to change how they pay or evaluate workers. If they only pay for "speed," workers will be lazy. If they pay for "accuracy" or penalize mistakes heavily, workers will be forced to check the Robot, even if it takes more time.

The Takeaway

AI isn't just a tool that makes everyone faster. It's a structural change that changes the rules of the game.

If you are good at checking, AI makes you a hero.
If you are bad at checking, AI makes you a liability.

The paper warns us that we shouldn't just assume AI will automatically improve everything. Without a focus on verification skills, AI might actually lower the overall quality of work in hospitals, law firms, and banks, because it tempts people to stop thinking and start trusting blindly.

1. Problem Statement

As AI systems integrate into high-stakes institutional workflows (e.g., medicine, law, finance), the traditional human-centric task execution model shifts to a delegation-verification pipeline. In this new paradigm, workers must decide whether to:

Execute tasks manually.
Delegate execution to AI without verification.
Delegate to AI and invest effort in verifying the output.

The core problem addressed is the misalignment between individual worker incentives and institutional objectives.

Workers optimize for their own utility (balancing task success, effort costs, and verification costs).
Institutions evaluate workers based on outcome-based standards (task success probability and institutional liability/costs), which may not align with the worker's private cost structure.

The paper asks: How does the introduction of AI reshape the "quality" of institutional workers? Specifically, does AI uniformly improve productivity, or does it create structural mechanisms that degrade performance for certain workers while benefiting others, even when all actors behave rationally?

2. Methodology

The authors propose a game-theoretic optimization framework modeling the interaction between a rational worker and a fixed AI system.

A. Model Formulation

Action Space: A worker chooses a pair $(d, s) \in [0, 1]^2$ $(d, s) \in [0, 1]^{2}$ :
- $d$ : Probability of delegating execution to AI.
- $s$ : Verification effort exerted if delegation occurs.
Worker Utility ( $U_W$ ):
$U_W(d, s) = b_W p(d, s) - \ell_W (1 - p(d, s)) - \text{cost}(d, s)$
Where $p(d, s)$ is the task success probability, and cost includes execution, verification, and correction costs.
Institutional Utility ( $U_I$ ):
$U_I(d, s) = b_I p(d, s) - \ell_I (1 - p(d, s)) - \xi \cdot \text{cost}(d, s)$
The institution values success differently ( $b_I, \ell_I$ ) and scales worker costs by a factor $\xi$ .
Worker Quality ( $Q(W)$ ): Defined as the institutional utility generated by the worker's optimal action: $Q(W) = U_I(d^\star, s^\star)$ .

B. Key Parameters

Verification Reliability ( $\alpha$ ): Determines the probability of detecting an AI error given effort $s$ (modeled via $\phi(s; \alpha)$ , e.g., $1 - e^{-\alpha s}$ ).
Execution Efficiency ( $\beta$ ): Determines the cost of manual execution ( $C_w$ ).
AI Capability: Defined by success probability $p_a$ and execution cost $C_a$ .

C. Theoretical Approach

The authors analyze the solution to the worker's utility maximization problem. They derive the optimal workflow $(d^\star, s^\star)$ as a function of ability parameters $(\alpha, \beta)$ . They then evaluate how $Q(W)$ changes relative to the pre-AI baseline ( $d=0, s=0$ ) and against a qualification threshold $\tau$ .

3. Key Contributions

1. Utility-Driven Model of Delegation and Verification

The paper formalizes the structural misalignment between worker and institution. It explicitly models the trade-off where a worker might rationally choose to delegate (pure delegation, $d=1, s=0$ ) because it maximizes their utility, even if it reduces institutional quality due to lower task success rates or higher liability risks.

2. Discovery of Workflow Phase Transitions

The authors prove that the mapping from worker ability to behavior is nonlinear and discontinuous. Small changes in verification reliability ( $\alpha$ ) or execution efficiency ( $\beta$ ) can trigger abrupt "phase transitions" between three distinct regimes:

Manual Work: $(0, 0)$
Pure Delegation: $(1, 0)$ (Delegating without checking)
Verified Delegation: $(1, s^\star)$ (Delegating with optimal checking)

These transitions arise because the worker's utility is affine in the delegation variable $d$ , leading to binary optimal choices ( $d^\star \in \{0, 1\}$ ) whenever the marginal benefit of delegation flips sign.

3. Characterization of Quality Regimes

The paper identifies two critical regimes regarding institutional worker quality:

Compliance Gain: Workers with low execution efficiency but high verification reliability can upgrade their institutional quality. AI allows them to offload difficult execution tasks while their strong verification skills ensure high success rates.
Compliance Loss: Workers with high execution efficiency but low verification reliability often experience quality degradation. They rationally choose "Pure Delegation" because it saves them effort, but their inability to detect AI errors leads to a net drop in institutional utility.

4. Verification Amplification Phenomenon

The study reveals that AI acts as a structural filter that amplifies disparities based on verification ability rather than execution ability. Access to AI disproportionately benefits those with strong evaluative capabilities while penalizing those who rely on execution speed but lack oversight skills.

4. Key Results

A. Theoretical Theorems

Theorem 3.1 (Optimal Action): Characterizes the boundaries ( $\psi_0, \psi_1$ ) in the $(\alpha, \beta)$ space separating the three workflow regimes. It proves that as $\alpha$ increases, workers transition from manual work to verified delegation, but if $\alpha$ is too low, they may jump directly to pure delegation.
Theorem 3.5 (Quality Improvement): Establishes that AI improves institutional quality if and only if verification reliability $\alpha$ exceeds a specific threshold function $\psi(\beta)$ . Crucially, increasing AI capability ( $p_a$ ) can worsen quality for low- $\alpha$ workers by inducing over-delegation.
Theorem 3.6 (Upgrading/Downgrading): Defines the conditions for "Compliance Gain" and "Compliance Loss." It shows that high- $\beta$ (efficient) workers are most at risk of being downgraded if their $\alpha$ is insufficient to prevent over-delegation.

B. Empirical Calibration (Section 4)

Using the Collab-CXR dataset (radiologists diagnosing chest X-rays), the authors map real-world data to model parameters.

Finding: A specific clinician was found to be in the "Manual Work" regime ( $d^\star=0$ ) because their verification reliability was too low to make delegation profitable for them, despite the AI's potential.
Intervention Analysis: The model suggests that to meet a higher qualification threshold, increasing the clinician's verification reliability ( $\alpha$ ) is more cost-effective than increasing their execution speed ( $\beta$ ).

C. Interventions (Section 6)

Worker-side: Upskilling verification reliability ( $\alpha$ ) is generally more effective than improving execution efficiency ( $\beta$ ) for maintaining institutional quality.
Institutional-side: Deploying more capable AI (increasing $p_a$ ) has non-monotonic effects. While it helps high- $\alpha$ workers, it can harm low- $\alpha$ workers by making pure delegation even more attractive, leading to a drop in overall quality. Similarly, reshaping incentives (increasing $b_W$ ) has complex, non-linear impacts.

D. Robustness (Section 7)

The core phenomena (phase transitions, compliance loss/gain) persist under extensions including:

Heterogeneous task difficulty.
Miscalibrated beliefs about AI capability (workers overestimating AI).
Partial re-execution costs.

5. Significance and Implications

Rationality vs. Bias: The paper demonstrates that quality degradation does not require behavioral biases (like over-reliance or automation bias). It is a structural consequence of rational optimization under misaligned incentives.
The Primacy of Verification: In an AI-assisted world, the primary determinant of a worker's institutional value shifts from execution efficiency to verification reliability.
Policy Design: Institutions cannot simply assume "AI + Human = Better." They must actively manage the verification capabilities of their workforce. Blindly deploying more capable AI or increasing efficiency training without addressing verification skills may lead to a net decrease in institutional performance.
Structural Filter: AI acts as a filter that sorts workers: those who can verify become more valuable; those who cannot, even if they are fast, become less valuable or risky.

In summary, the paper provides a rigorous mathematical foundation for understanding why AI adoption can lead to uneven outcomes, identifying rational over-delegation driven by low verification reliability as the core mechanism for institutional quality degradation.