TLDR: A study analyzed 200,000 Microsoft Bing Copilot conversations to understand how generative AI impacts occupations. It found that information gathering and writing are common user goals, while AI primarily provides information, writes, teaches, and advises. Knowledge work, sales, and administrative support roles show the highest AI applicability, while physical labor roles show the least. The research aligns with previous predictions but highlights that AI’s impact on wages and education is varied, emphasizing that current usage data doesn’t directly predict job automation or augmentation.
A recent study delves into the real-world impact of generative AI on various occupations, offering insights into how people are currently using AI tools like Microsoft Bing Copilot in their work. The research, titled “Working with AI: Measuring the Occupational Implications of Generative AI,” analyzes 200,000 anonymized conversations between users and Microsoft Bing Copilot to understand the activities people seek AI assistance for and the tasks AI itself performs.
The study highlights that the most common work activities for which users seek AI assistance involve gathering information and writing. On the flip side, the AI’s most frequent actions include providing information and assistance, writing, teaching, and advising. Interestingly, the research found a significant distinction between user goals and AI actions, with 40% of conversations showing completely different sets of activities for the user and the AI.
To measure the potential impact on jobs, the researchers developed an “AI applicability score” for each occupation. This score considers how frequently AI is used for an occupation’s work activities, the success rate of those tasks, and the scope of AI’s impact. Occupations with the highest AI applicability scores are primarily in knowledge work, such as computer and mathematical roles, and office and administrative support. Sales occupations also scored high, likely due to their focus on providing and communicating information.
Conversely, occupations involving significant physical components, such as manual labor, operating machinery, or direct physical interaction with people (e.g., nursing assistants, massage therapists), showed the lowest AI applicability scores. This suggests that current generative AI capabilities are less suited for these types of tasks.
The study also examined user satisfaction and task completion rates. It found that activities like writing, editing, and researching information received the most positive feedback and had high completion rates. However, tasks involving data analysis or visual design tended to have lower satisfaction and completion rates, indicating areas where AI’s current capabilities might be less effective.
When comparing their findings to previous predictions about AI’s labor market impact, the researchers found a strong correlation, especially at broader occupational group levels. This suggests that real-world AI usage largely aligns with earlier forecasts regarding which jobs are most likely to be affected by large language models.
Regarding socioeconomic factors, the study found only a weak positive correlation between AI applicability scores and average wages. While occupations requiring a Bachelor’s degree generally showed higher AI applicability than those with lower educational requirements, there was still a wide range of potential impact across all wage and education levels.
Also Read:
- Experimental Studies Confirm Generative AI’s Significant Impact on Workforce Productivity
- Unpacking Public Reactions to ChatGPT: How Jobs and Culture Shape AI Discourse
The authors emphasize that their data represents a snapshot of AI usage and does not predict job automation or augmentation outcomes. They note that new technologies often lead to job restructuring and the emergence of entirely new occupations, rather than simple job displacement. For more details, you can refer to the full research paper: Working with AI: Measuring the Occupational Implications of Generative AI.
