A Tale of Three AIs (Premium)

We’re barely one year into the generative AI era, but we’re finally starting to understand the ways in which this technology can improve our productivity. Part of the hurdle here is our natural tendency to make AI work in familiar ways. But AI is well-suited to certain things and less well-suited to others. And trying to make it work like the tools of the past is a mistake.

With the understanding that my findings aren’t particularly unique, I do have some general observations about how generative AI might be best used. These observations are limited in some ways to my needs—for example, I feel that Copilot, DALL-E, and Gemini are particularly good at creating images from text prompts—and will evolve as the technology evolves and my experience grows.

Generative AI is particularly good for summarizing. This could be a lengthy email, a post-meeting wrap-up, a long article on the web, a PDF, whatever. This isn’t about laziness, though it will be used by the lazy. It’s about making the best use of your time, finding what’s pertinent or the most important, and focusing on that. The implementations vary, but I like the side-by-side approached used by Copilot for Microsoft 365, and how it integrates with core Office apps like Word and Outlook.

Generative AI works best with a limited data set. This is the grounding feature I discussed recently. The more data that AI has, the more inaccurate it will be, and the more it will “hallucinate” and present untruths (or even purely invented information) as fact. Products like Copilot for Microsoft will use grounding to limit AI’s reach to an organization, but individuals will see this in things like custom GPTs that are either provided to us by AI makers or third parties, or things we make ourselves. They will only—and can only--get better in time.

Finally, generative AI works better when you’re more specific. This is the key conceptual hurdle, I think, and the key learning: We’re trained to use traditional search engines like Google’s to get answers to specific questions (“What is the capital of Massachusetts?”), but that’s not an ideal use-case for AI. (Even though almost any AI would handle that particular question; “What happened yesterday in Mexico City?” would be more problematic.)

We’re so trained to be terse in our interactions with these things that it probably doesn’t even occur to most people to behave otherwise when dealing with AI. This is true of image creation, which doesn’t impact too many people. But it’s also true of any generative AI capability. Including those that would benefit a broader range of users. Including the productivity scenarios that really matter.

The earliest theoretical example I used to demonstrate, if vaguely, was how one might use it to perform an unfamiliar task like creating a PowerPoint presentation. In this scenario, I imagined a typical knowledge worker who perhaps spent their days writing text in Wo...