Artificial intelligence (AI) remains a riddle for those of us wondering about its potential impacts on workers.
Big—and often vague—claims have been made about AI’s potential to do things like improve perception (for example, in medical diagnosis from scans), automate judgment (e.g., hate speech detection), and predict social outcomes (for policing or job screenings). Yet for all that, the technology remains a fluid and emergent topic, with no single definition and relatively little real-world examples of adoption to learn from. Most notably, it’s very hard to parse what work AI may take over from humans when there’s no agreement on what, exactly, it can do.
Fortunately, a pathbreaking paper by our Stanford University colleague Michael Webb cuts through many of the subjective claims and assessments. Webb’s method finds the overlap between common job descriptions and new AI-based patents to detail in precise, objective terms what AI can or may soon be capable of in the workplace.
From a pool of 16,400 patents, Webb compiled 8,000 verb-object word pairs using a natural-language processing algorithm. The below table, which appears in Brookings Metro’s report on AI’s potential impacts on people and places, displays the top eight by frequency.
The table is something of a Rosetta Stone for beginning to parse the current and coming capabilities of AI. Since patents are predictions of future commercial relevance, Webb’s verb-object pairs come about as close as we can get to understanding how researchers and companies anticipate AI may be deployed. And since applicants must pay nontrivial fees for filing patents, these words’ predictive value likely exceeds purely subjective expert assessment. In that sense, these 69 verb-object pairs amount to an unusually well-grounded source of insight about what AI does or may soon do.
What do these word pairs say about AI’s capabilities?
The verbs “recognize,” “detect,” and “determine” suggest a wide variety of AI capabilities focused on augmenting human perception. Pairings such as “recognize, face,” “detect, abnormality,” or “identify, illegality” point to a relatively straightforward set of empirical or measurement activities on which AI has been making genuine, rapid progress. On this front, AI now often exceeds human accuracy, whether it be for voice- and facial-recognition or in transportation, medicine, or consumer protection.
Related to this work in perception is a series of “control” activities: “control, process,” “control, emission,” “control, traffic,” etc. These verb-object pairs seem to point to an array of capabilities in which AI will link enhanced perception to automated command-control activities in order to automatically “optimize” performance, say of energy efficiency at a power plant. Some of this perception and control work currently involves human monitoring, but seems relatively routine rather than higher-order or deeply “human.”
With that said, some of the verb-object pairs do appear to entail higher-order, human-centric capabilities. The verbs “determine,” “classify,” and “predict” are especially suggestive. Each of these activities entails a form of categorization of inputs characteristic of human intellect. Pairings such as “determine, relevance” and “determine, risk” or “classify, data” and “classify, pattern” all involve abstract conceptual analysis, as well as—especially—pattern recognition. For that matter, capabilities such as “predict, performance,” “predict, behavior,” or “predict, prognosis” signal that patented AI applications have already begun to mimic high-level human mental processes.
While AI does not yet seem poised for any near-term substitution of all aspects of human cognition, it is pushing toward one particular aspect of intelligence: prediction, which is central to decisionmaking and an essential aspect of all kinds of occupations. Prediction under conditions of uncertainty, as Ajay Agrawal, Joshua S. Gans, and Avi Goldfarb observe, is a widespread and challenging aspect of many information-sector jobs in health, business, management, marketing, and education. It’s no surprise, then, that our recent analysis concluded that high-skill, high-wage occupations would be the most affected by AI—occupations such as financial managers, programmers, market research specialists, and management analysts. Prediction is ubiquitous, but especially central as one moves higher up the white-collar occupational ladder.
In that sense, Michael Webb’s matrix of verbs and objects provides a grounded and precise new set of signals about the future of work. Rather than routine or rote tasks, it is jobs that require higher-order prediction-making that may be most exposed to AI. Or at least, that’s our prediction.