The Reality Gap in AI Exposure

Understanding the future of work requires moving beyond speculative headlines about "robots taking over." To prepare for the Intelligence Age, we must distinguish between "lab-tested potential"—what Large Language Models (LLMs) are technically capable of—and "office-floor reality"—what they are actually doing in the economy today.

1. The Foundation: Theoretical vs. Observed Exposure

The starting point for any labor economist is defining how much a job is "exposed" to technology. However, exposure is not a monolithic metric. There is a critical distinction between a system's theoretical capability and its actual integration into professional workflows.

Theoretical Capability ()	Observed Exposure
Definition: Measures if it is technically possible for an LLM to perform a specific task at least twice as fast as a human.	Definition: Measures actual automated usage in professional settings by combining theoretical capability with real-world traffic data.
Scoring Logic: Based on the Eloundou et al. (2023) system, tasks are scored as 1 (fully possible), 0.5 (requires additional software tools), or 0 (not possible).	Weighting Logic: Derived from the Anthropic Economic Index, where fully automated implementations receive full weight (1.0), while augmentative use receives only half weight (0.5).
Source: Researcher-led assessments of the technical limits of LLM reasoning and processing.	Source: Real-world API implementation patterns and work-related traffic volume in the workforce.

 

Synthesis Insight: Measuring actual usage is a far better predictor of economic disruption than simply listing capabilities on paper. While theoretical capability tells us what might happen in a vacuum, observed exposure provides a lead indicator of structural change, allowing us to identify the most vulnerable roles before displacement manifests in aggregate unemployment data.

While these metrics are highly correlated—97% of actual usage falls within categories deemed theoretically possible—we are currently observing a massive chasm between what is possible and what is practiced.

 

2. Visualizing the Gap: The Capability-Usage Chasm

Data from the Anthropic Economic Index illustrates that AI is currently far from reaching its theoretical limits. Even in highly digital sectors, the technology’s footprint is significantly smaller than its technical potential.

• Computer & Math Occupations: While LLMs have a 94% theoretical penetration rate for tasks in this field, the actual observed coverage is currently only 33%.
• Office & Admin Occupations: These roles have a staggering theoretical feasibility of 90%, yet actual automated usage remains significantly lower, represented as approximately one-third or less of that potential.
• The "Uncovered" Area: Large swaths of the economy remain entirely beyond AI's reach. This includes:
• Physical Agricultural Work: Tasks like pruning trees or operating complex farm machinery.
• Courtroom Presence: The physical representation of clients in a legal setting.
• Manual Precision: Roles requiring complex sensory judgment and physical movement.

Synthesis Insight: This gap exists because technical feasibility does not equal immediate utility. Using the Anthropic Economic Index to track these patterns reveals that adoption is slowed by the "last mile" of implementation: legal constraints (such as the requirement for licensed human signatures for drug refills), specific software interoperability, and the essential nature of human verification steps.

As these institutional and technical hurdles are cleared, the gap will narrow, signaling a transition from experimentation to total integration for high-exposure roles.

 

3. Case Studies: The Most and Least Exposed Occupations

The impact of AI is not distributed evenly. By analyzing observed exposure, we can identify which roles are on the "front line" of automation and which are currently shielded by a "physical guard."

### Profile A: High Exposure (The 'Front Line') These roles involve high volumes of digital-first tasks being actively automated via API implementations.

• Computer Programmers: 75% coverage (the highest observed exposure).
• Financial Analysts: Highly exposed due to high-volume data processing and predictive modeling tasks.
• Customer Service Representatives: Rapidly increasing coverage driven by shifts toward automated first-party API traffic.
• Data Entry Keyers: 67% coverage, as AI can now autonomously read source documents and input data.

### Profile B: Zero Exposure (The 'Physical Guard') Roughly 30% of the workforce has zero observed exposure. These jobs require physical presence or manual dexterity that current LLMs cannot replicate.

• Motorcycle Mechanics and Cooks
• Dressing Room Attendants and Bartenders
• Lifeguards and Dishwashers

Synthesis Insight: The "Physical Guard" represents the current boundary of LLM technology. Until intelligence is paired with affordable, high-dexterity robotics, any job requiring manual interaction with the physical world remains shielded from the immediate displacement risks affecting digital-first, white-collar roles.

 

4. The Human Profile: Who is Actually Exposed?

Synthesizing data from the Current Population Survey allows us to build a profile of the workers currently facing the highest exposure. Unlike the Industrial Revolution, which targeted manual labor, AI exposure is concentrated among a highly educated and highly paid demographic.

• Education & Pay: Workers in the top quartile of exposure are nearly four times more likely to have a graduate degree (17.4% vs 4.5% for unexposed workers) and earn 47% more on average.
• Gender & Ethnicity: Highly exposed workers are 16 percentage points more likely to be female, 11 percentage points more likely to be white, and almost twice as likely to be Asian compared to the unexposed group.
• Projected Growth: The labor market is already pricing in this shift. For the 2024–2034 projection period, Bureau of Labor Statistics (BLS) data shows that for every 10% increase in AI task coverage, projected job growth drops by 0.6 percentage points.

Synthesis Insight: This profile defines the "White-Collar Displacement Risk." Historically, technology augmented the highly educated while replacing the laborer; in the Intelligence Age, the inverse is occurring, creating a unique socioeconomic challenge where the highest earners face the most direct automation pressure.

 

5. Labor Market Outcomes and Policy Responses

While the potential for disruption is high, the current data reflects a market in transition rather than a sudden systemic failure.

Impact vs. Forecast Summary

• Current Reality: There is no systematic increase in aggregate unemployment for highly exposed workers yet. However, we are seeing a drop in the job finding rate (a decrease of approximately 0.5 percentage points per month) for young workers (ages 22-25) in exposed fields, representing a 14% drop in hiring since late 2022.
• The Solution (Industrial Policy): To manage this, we require an industrial policy that ensures the transition to superintelligence remains "People First."

1.      Worker Perspectives: Formal mechanisms for workers to collaborate with management, ensuring AI improves job quality and safety rather than merely intensifying workloads.

2.      Adaptive Safety Nets: Real-time measurement systems that trigger targeted, time-bound, and proportional assistance—such as training vouchers or wage insurance—when specific displacement thresholds are met.

3.      Portable Benefits: Decoupling healthcare and retirement from specific employers, allowing benefits to follow the individual across different jobs or entrepreneurial ventures.

Synthesis Insight: AI's impact is not a sudden "COVID-style" shock that breaks the economy overnight; it is a gradual "Internet-style" restructuring. This window of opportunity allows us to use the democratic process to shape the transition. The goal is to ensure that as safety and capability scale, our social contracts scale alongside them, ensuring the benefits of superintelligence are shared by all.