How we score AI exposure

Each role is decomposed into 9–15 representative tasks generated by a language model and guided by O*NET-aligned role context. Tasks are split across three buckets: tasks AI handles end-to-end, tasks where you and AI collaborate, and tasks that stay human. This task-based approach is inspired by prior work — notably Eloundou et al. (2024), "GPTs are GPTs" in Science, and the Felten, Raj & Seamans (2021) AIOE framework — but the specific decomposition here is generated per query and is not sourced directly from those datasets.

Each task is scored on four ordinal dimensions (0–4 scale), using structured prompts with anchored definitions:

• Routinization (R)

  How repeatable and rule-based is the task? Filing the same expense report each week scores high; negotiating a contract scores low.

• Data availability (D)

  Does AI have the inputs it needs — text, code, images, structured data? A task that hinges on tacit context the model can't see scores low.

• Judgment required (J)

  How much human judgment, ethics, or accountability does it take? This is an inverse signal — high judgment lowers automation risk.

• Presence required (P)

  Does it need a body in a room or a real human relationship? Also inverse — physical / interpersonal presence resists automation.

The 0–4 scale is ordinal, not interval. Arithmetic operations applied to these values assume approximate linear spacing for practical modeling purposes.

Each dimension is defined using anchored behavioral criteria to reduce rating variance; however, ratings remain model-estimated rather than independently human-validated.

automation(task) = (R + D) / 8 × (1 − J/4) × (1 − P/4)
augmentation(task) = (D/4) × max(J/4, P/4) × (1 − automation(task))

This functional form assumes:

• R and D contribute linearly and equally to automation potential.
• J and P act as multiplicative suppressors of automation.
• Maximum judgment or presence (a score of 4) reduces automation potential to zero by construction.

These assumptions reflect a modeling choice rather than an empirically calibrated relationship and may be revised as validation data becomes available.

The augmentation formula is heuristic. It assumes augmentation potential rises with data availability and at least one human-dependent dimension (judgment or presence), and falls as tasks become fully automatable.

automation  = (4 + 4) / 8 × (1 − 1/4) × (1 − 0/4)
            = 1.00 × 0.75 × 1.00
            = 0.75   →  75%

augmentation = 4/4 × max(1/4, 0/4) × (1 − 0.75)
            = 1.00 × 0.25 × 0.25
            = 0.06   →  6%

automation  = (1 + 1) / 8 × (1 − 4/4) × (1 − 4/4)
            = 0.25 × 0.00 × 0.00
            = 0.00   →  0%

augmentation = 1/4 × max(4/4, 4/4) × (1 − 0.00)
            = 0.25 × 1.00 × 1.00
            = 0.25   →  25%

Role-level scores are computed as time-weighted averages of task-level scores, where task time shares are estimated by the model and normalized to sum to 100%.

automation_risk = Σ time_share_i × automation_score_i
augmentation_potential = Σ time_share_i × augmentation_score_i

Time allocation is inferred rather than observed and introduces an additional source of uncertainty in role-level scores.

The ± value represents a time-weighted dispersion measure (standard deviation) of task-level automation scores. The displayed dispersion is capped between ±3 and ±18 percentage points for interpretability. A role with consistent task ratings gets a tight band; a role that mixes very automatable and very human work gets a wider one.

Because of this capping, interpret the displayed value as a bounded heterogeneity indicator rather than a raw statistical estimate.

This is not a statistical confidence interval in the repeated-sampling sense — it doesn't quantify uncertainty around a population parameter. It's a dispersion measure that tells you how internally mixed the task profile is. Read it as: "how much do the tasks inside this role disagree with each other?"

Each map is programmatically matched to the closest O*NET-SOC occupation — the U.S. Department of Labor's standard occupation taxonomy — using keyword and similarity-based search via O*NET Web Services. We then confirm the code against O*NET in real time:

• Verified by O*NET — the code exists and either matches O*NET's own top search hit for the job title, or its canonical title overlaps the input meaningfully. The map page surfaces a green pill linking straight to the O*NET occupation summary.
• Best-guess match — the SOC exists but O*NET's top hit for the title is a different occupation. We surface the alternative in the badge tooltip so you can judge for yourself.
• Unverified SOC — the suggested code doesn't exist in O*NET. We fall back to O*NET's top match for the title and show it in the tooltip.

This mapping confirms alignment with an existing occupation code but does not guarantee semantic equivalence between the generated role description and the official O*NET definition.

Note that our time-share weighting is a product choice that goes beyond O*NET's native core-vs-supplemental task convention (which is based on relevance and importance thresholds, not estimated time). It's defensible as a modeling decision, but it is additional structure on top of O*NET, not a feature of O*NET itself.

These scores reflect technical task-level exposure to AI capabilities — not realized adoption, job loss, or labor market outcomes.

Organizational, regulatory, economic, and behavioral constraints are not modeled and may materially affect real-world outcomes.

Methodological foundations:

• Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702). Preprint: arXiv:2303.10130 (we rely on the preprint appendix for rubric-level detail).
• Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What can machines learn, and what does it imply for the workforce? AEA Papers and Proceedings — the Suitability for Machine Learning (SML) framework, the closer methodological ancestor of the GPTs task-exposure rubric.
• Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence (AIOE). Strategic Management Journal. The authors explicitly state AIOE is agnostic about whether AI substitutes for or complements labor — citing it as direct support for an "automation risk" score would overclaim.
• U.S. Department of Labor / ETA, O*NET OnLine (onetonline.org) — task, ability, and SOC taxonomy.

Career-guidance reference (not a methodological foundation):

• U.S. Bureau of Labor Statistics, Occupational Outlook Handbook — used for role descriptions and outlook context, not for the exposure scores themselves.

This system is a structured estimation model with the following limitations:

• Model-generated inputs. Task decomposition, time allocation, and dimension ratings are generated by a language model and may vary with phrasing, model version, or prompt structure.
• Unverified reliability. Test–retest stability, inter-rater agreement, and prompt sensitivity have not yet been formally benchmarked.
• Heuristic functional forms. The automation and augmentation formulas encode modeling assumptions that are not empirically calibrated against observed labor outcomes.
• Ordinal scaling. Dimension scores (0–4) are ordinal but are treated as approximately linear for aggregation.
• Incomplete construct coverage. Factors such as error tolerance, regulatory constraints, and implementation costs are not explicitly modeled.
• O*NET alignment limits. SOC mapping improves standardization but does not guarantee precise occupational equivalence.
• Non-predictive scope. Scores represent technical feasibility, not timelines, adoption rates, or economic displacement.

To improve reliability and validity, future iterations will benchmark:

• Test–retest stability across repeated runs.
• Sensitivity to prompt phrasing and role descriptions.
• Agreement between model-generated and human-rated scores.
• Alignment with O*NET task importance and frequency data.
• Sensitivity of outputs to changes in task weights and dimension scores.

These evaluations will inform calibration and potential revision of scoring functions.