Check which tasks AI does well before you trust it
Updated: 18 hours ago
AI is very good at some tasks. At other tasks, AI is wrong, and the errors are not easy to see.
The study. Researchers from Harvard, MIT, Wharton and other schools worked with Boston Consulting Group. In the study, 758 consultants did realistic tasks with or without GPT-4.
The finding. On tasks that AI does well, consultants with AI did 12.2% more tasks. They did the tasks 25.1% faster. The quality of their work was more than 40% higher. On a task that AI does badly, consultants with AI were 19 percentage points less likely to find the correct answer.
Why it matters. The limit between the tasks that AI does well and the tasks that AI does badly is not clear. The authors call this limit a "jagged frontier", which is an irregular border. People who trust AI on every task can make more errors on the difficult tasks.
The takeaway. Test AI on each type of task in your team. Use AI fully on the tasks that it does well. Check its work on the tasks that it does badly.
Want to read more?
Subscribe to plaintalkbusiness.com to keep reading this exclusive post.
