[Paper Review] Opening the Frey/Osborne Black Box: Which Tasks of a Job are Susceptible to Computerization?

Philipp Brandes, Roger Wattenhofer|arXiv (Cornell University)|Apr 29, 2016

Labor market dynamics and wage inequality18 references10 citations

TL;DR

This paper refines Frey and Osborne's job automation predictions by analyzing individual job tasks rather than whole jobs, using a linear programming model to assign automation probabilities to tasks while preserving the overall job-level probability. The key contribution is a transparent, task-level explanation of automation susceptibility, revealing that routine, low-education, and low-deductive-reasoning jobs are most at risk.

ABSTRACT

In their seminal paper, Frey and Osborne quantified the automation of jobs, by assigning each job in the O*NET database a probability to be automated. In this paper, we refine their results in the following way: Every O*NET job consists of a set of tasks, and these tasks can be related. We use a linear program to assign probabilities to tasks, such that related tasks have a similar probability and the tasks can explain the computerization probability of a job. Analyzing jobs on the level of tasks helps comprehending the results, as experts as well as laymen can more easily criticize and refine what parts of a job are susceptible to computerization.

Motivation & Objective

• To address the opacity of Frey and Osborne's job-level automation probability estimates by analyzing individual tasks.
• To provide a transparent, explainable framework that allows experts and laypeople to critique which parts of a job are automatable.
• To model task-level automation probabilities such that their weighted average matches the job-level probability from Frey and Osborne.
• To investigate correlations between job characteristics—like education level, training, and reasoning skills—and automation susceptibility.
• To identify and correct anomalies in the original Frey and Osborne estimates, such as the 96% automation probability for compensation and benefits managers.

Proposed method

• The method uses a linear program to assign automation probabilities to individual tasks within each job, ensuring that the weighted average of task probabilities equals the job-level probability from Frey and Osborne.
• Task shares (frequency of occurrence in a job) are used as weights in the linear program to balance the contribution of each task to the overall job probability.
• The model enforces similarity in automation probability between related or co-occurring tasks, improving consistency and interpretability.
• The approach leverages the O*NET database, which provides detailed occupational data including task descriptions, required education, training, and reasoning levels.
• The model is evaluated by comparing predicted task probabilities with original job-level probabilities and identifying inconsistencies.
• Statistical correlations are analyzed between job-level automation probability and attributes such as education level, on-the-job training, and deductive reasoning requirements.

Experimental results

Research questions

• RQ1 Which specific tasks within a job are most susceptible to automation, and how can their probabilities be estimated transparently?
• RQ2 How does the frequency of a task (its share in a job) correlate with its automation probability?
• RQ3 To what extent do job characteristics such as required education, training, and reasoning ability predict automation susceptibility?
• RQ4 Are there inconsistencies in the original Frey and Osborne job-level automation probabilities that can be detected and corrected at the task level?
• RQ5 Does the current level of automation in a job correlate with its future automation risk, or is the risk spread across previously unautomated jobs?

Key findings

• There is a strong negative correlation between the level of education required for a job and its probability of automation, with low-education jobs being most at risk.
• Jobs requiring post-doctoral training (level 12) have significantly lower automation probabilities than those requiring only a high school diploma (level 1).
• The amount of on-the-job training required has a weaker effect on automation risk than formal education level, with only a marginal impact on predicted automation probability.
• Tasks that are performed more frequently (higher share in a job) do not show a higher likelihood of automation, indicating no strong correlation between task frequency and automation susceptibility.
• The original Frey and Osborne estimate of 96% automation probability for compensation and benefits managers is flagged as suspicious and likely inaccurate.
• The current level of automation in a job (on a 1–5 scale) shows only a weak correlation (0.23) with future automation risk, suggesting that many previously unautomated jobs are now at risk.