Technical Summary: The Scaling Paradox in Human–AI Collaboration

Problem Statement

The discovery of Scaling Laws in artificial intelligence (Kaplan et al. 2020; Hestness et al. 2017) establishes that AI capabilities improve predictably as model size, data, and compute increase. However, in operational settings, AI rarely functions in isolation; it operates within human–AI collaborative systems where human workers must allocate effort to review, validate, or correct AI outputs.

The central problem addressed is whether the empirical gains from AI scaling translate into improved performance for the joint human–AI system. While intuition suggests that more capable AI should allow humans to process more tasks, recent empirical evidence (e.g., Becker et al. 2025; METR) indicates mixed or even negative productivity outcomes. The paper posits that human misperception of AI capabilities is a critical, often overlooked mechanism that can break the scaling effect, leading to a "scaling paradox" where increased AI scale reduces overall system performance and firm profit.

Methodology

The authors develop a parsimonious analytical model of human–AI collaboration to study endogenous effort allocation under scaling laws.

1. The Human–AI Workflow

The model formalizes a three-step sequential process:

1. Setup: A human worker initiates a project with a fixed time cost $t_0$.
2. AI Execution: The AI attempts the task with a success probability $1 - e^{-\alpha s}$, where $s$ is the AI scale (logarithmic measure of resources) and $\alpha$ is the scaling factor (task-AI alignment).
3. Human Review: The worker allocates time $t$ to review and correct the AI output. If the AI fails, the human corrects it with probability $1 - e^{-t}$.

The joint success probability is $p(t; \alpha, s) = 1 - e^{-(\alpha s + t)}$. The worker has a fixed capacity $C$ and seeks to maximize the expected number of successful projects.

2. Misperception Mechanism

The model introduces a behavioral parameter where the worker perceives the scaling factor as $\hat{\alpha}$, which may differ from the true $\alpha$.

• Over-perception ($\hat{\alpha} > \alpha$): Workers overestimate AI capability, leading to aggressive effort reduction.
• Under-perception ($\hat{\alpha} < \alpha$): Workers underestimate AI capability, leading to excessive effort allocation.

3. Firm Perspective and Misalignment

The firm maximizes profit, defined as the net reward from successful projects minus the deployment cost $c \cdot s$. Crucially, the firm bears the AI cost while the worker does not, creating a structural firm-worker misalignment regarding optimal effort levels. The model analyzes how misperception exacerbates this misalignment.

4. Policy Interventions

The study evaluates two operational levers:

• Cost Internalization (CI): Shifting a fraction $\rho$ of the AI deployment cost to the worker to alter effort incentives.
• Perception Alignment: Interventions (e.g., training, transparent reporting) to align $\hat{\alpha}$ with $\alpha$.

Key Contributions

1. Integration of Scaling Laws into Operations Management: The paper embeds the scaling law as a primitive in an analytical operations framework, moving beyond empirical ML studies to examine how scaling affects the entire production system when human effort responds endogenously.
2. Identification of the Scaling Paradox: The authors identify human misperception as a novel behavioral mechanism where scaling fails. Specifically, over-perception can cause the joint system's performance to become non-monotone with respect to AI scale, creating a region where larger AI systems yield worse outcomes.
3. Asymmetry of Misperception Risks: The paper distinguishes the asymmetric effects of over- vs. under-perception. Over-perception reverses scaling benefits and amplifies firm-worker misalignment. Under-perception preserves monotonic improvement (albeit at a slower rate) and can, counter-intuitively, partially mitigate firm-worker misalignment by inducing workers to exert more effort.
4. Policy Design for Distortion Mitigation: The study characterizes how cost internalization and perception alignment should be tailored to the direction of misperception and the economics of deployment (e.g., cost levels).

Key Results

1. Perfect Collaboration ($\hat{\alpha} = \alpha$)

When humans accurately perceive AI capability, the scaling law holds. As scale $s$ increases:

• Human effort per project ($t^*$) decreases non-increasingly.
• Project throughput increases.
• Total system reward increases.
  The system transitions from a "Human-Only" regime to a "Human-in-the-Loop" (HIL) regime, and eventually to an "Artificial General Intelligence" (AGI) regime where human effort beyond setup is zero.

2. Over-Perception ($\hat{\alpha} > \alpha$)

• Effort Distortion: Workers reduce effort too aggressively ($\hat{t} < t^*$).
• The Scaling Paradox: In the HIL regime, if the degree of over-perception is sufficiently high, the realized reward becomes non-monotone in $s$. Initially, scaling improves performance, but beyond a threshold, the reduction in human effort outweighs the AI's technical gains, causing total reward to decline.
• Firm Impact: This paradox is amplified at the firm level. As reward declines, the firm's profit drops even sharper because the firm continues to pay the increasing deployment cost $c \cdot s$ for a shrinking number of successful projects. Over-perception widens the gap between worker and firm-optimal effort.

3. Under-Perception ($\hat{\alpha} < \alpha$)

• Effort Distortion: Workers over-invest effort ($\hat{t} > t^*$).
• Attenuated Scaling: The scaling law holds (reward increases with $s$), but gains are slower than in the perfect collaboration case because workers inefficiently allocate too much time per project, reducing throughput.
• Firm Impact: Under-perception can partially align worker behavior with firm interests. Since the firm prefers higher effort than the worker (to reduce costly AI deployments), the worker's "over-effort" under under-perception can inadvertently reduce the firm-worker misalignment, sometimes improving firm profit relative to the perfect collaboration baseline.

4. Policy Interventions

• Cost Internalization:
  • Low AI Cost: Shifting costs to workers (high $\rho$) generally increases firm profit.
  • High AI Cost: Firms should absorb more costs (low $\rho$) to prevent workers from abandoning AI tools.
  • Under Misperception: Under over-perception, firms can shift more costs to workers (who are already over-confident). Under under-perception, firms must bear more costs to encourage adoption.
• Perception Alignment:
  • Over-Perception: Alignment is unambiguously beneficial for both workers and firms, preventing the scaling paradox.
  • Under-Perception: Alignment is not always beneficial for the firm. Correcting mild under-perception can reduce firm profit by pulling workers away from the firm-preferred high-effort level. Aggressive alignment is only beneficial if under-perception is severe.

Significance and Claims

The paper argues that AI scaling is not merely a technological challenge but a behavioral and operational one. The authors claim that:

1. Automatic Value is a Myth: Larger AI systems do not automatically lead to better operational outcomes; the value depends critically on how increased capabilities shape human beliefs and collaborative efforts.
2. Management of the Interface: Organizations may benefit more from managing the human–AI interface (via perception alignment and incentive design) than simply investing in larger, more expensive models.
3. Risk of Over-Perception: The "scaling paradox" suggests that current organizational trends of aggressively scaling AI while reducing human oversight—driven by over-optimism—may lead to deteriorating performance and amplified profit losses.
4. Asymmetric Interventions: Operational policies must be asymmetric; what works for over-perception (e.g., cost shifting, aggressive alignment) may be counterproductive for under-perception.

The study concludes that realizing the value of deployed AI requires firms to adapt human beliefs and incentive structures to the specific capabilities of the AI system, cautioning against the view that technological progress alone guarantees organizational performance.