Research map / Large language models / AI and society
A Mathematical Framework for AI-Human Integration in Work
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a mathematical framework modeling jobs, workers, and worker-job fit with decision-level and action-level subskills, deriving phase transitions and merging/productivity-compression results, and illustrating with O*NET and BIG-bench Lite data.
Problem
Existing evaluations conflate reasoning and action skills and lack a formal way to assess how GenAI tools and human workers complement each other at the job level, including when job success changes sharply with ability.
Approach
Decompose each skill into decision-level and action-level subskills with ability profiles under noise; define job success probability via aggregated subskill error rates; prove phase-transition bounds via McDiarmid concentration and show merging workers with complementary subskills yields gains, explaining productivity compression.
Key results
- Under the default setting, increasing decision-level ability a1 by 4.3% (0.492 to 0.513) raises job success P from 0.2 to 0.8 for sigma=0.1.
- Merging complementary workers with a1=0.5 vs 0.4 and a2=0.4 vs 0.5 yields P_12=1 while P_1=P_2=0.4, a gain of 0.6.
- Phase-transition window gamma_1 scales as O(sigma*sqrt(ln(1/theta)/n)), so larger jobs show sharper transitions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase decision-level ability a₁ from 0.492 to 0.513 | Proposed framework · Synthetic linear-profile job, σ = 0.1 | job success probability: 0.2 to 0.8 | Helped |
| Reduce noise σ | Proposed framework · Synthetic linear-profile jobs | sharpness of job success probability transition | Helped |
| Increase decision-level ability a from 0.07 to 0.34 | Proposed framework · Synthetic dependent-subskill job, p = 0 | job success probability: 0.2 to 0.8 | Helped |
| Increase decision-level ability a | Proposed framework · Synthetic dependent-subskill job, p = 0.4 | ability increase for job success probability change: 0.44 | Hurt |
| Combine W₁ decision-level ability with W₂ action-level ability | Complementary workers W₁ and W₂ · Synthetic complementary-worker job | job success probability: P₁₂ = 1 while P₁ = P₂ = 0.4; gain of 0.6 | Helped |
| Increase dependency between subskill errors p | Proposed framework · Synthetic dependent-subskill job | job success probability | Mixed |
| Assign each subskill to the worker with higher expected ability | Workers W₁ and W₂ · Computer Programmer example, a ≤ 0.22 and c ∈ [0.78, 0.82] | job success probability: P_merge = 1 while P₁, P₂ ≤ 0.6; gain of at least 0.4 | Helped |
| Use strongly correlated subskill noise instead of independent noise | Proposed framework · Synthetic job | sensitivity of job success probability to ability | Hurt |
Surprising findings
- Dependence between subskill errors (higher p) widens the transition window, so loosely coupled subskills give sharper gains.
Limitations
- Focuses only on job success probability, ignoring efficiency, time, and cost.
- O*NET is static and survey-based; Big-bench Lite estimates may embed LLM-based modeling biases.
- Subskill division relies on GPT-4o estimates of decision-level degree and task-skill mappings.
Experimental setup
Datasets: O*NET BIG-bench Lite
Benchmarks: BIG-bench Lite O*NET
Metrics: job success probability phase transition width merging gain productivity compression
Baselines: PaLM
Builds on: McDiarmid's inequality O*NET BIG-bench Lite
Closest papers by meaning
- Balancing Humans and Machines: A Study on Integration Scale and Its Impact on Collaborative Performance (AAAI 2024 · similarity 0.87)
- JobBench: Aligning Agent Work With Human Will (NeurIPS 2026 · similarity 0.86)
- Working with AI: Measuring the Applicability of Generative AI to Occupations (NeurIPS 2026 · similarity 0.86)
- When AI Agents Compete for Jobs: Strategic Capabilities and Economic Dynamics of AI Labour Markets (ICML 2026 · similarity 0.85)
- The Generative AI Paradox: “What It Can Create, It May Not Understand” (ICLR 2024 · similarity 0.85)
- Path Dependence under Adaptive AI Delegation (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.