Now that most everyone’s played around with ChatGPT, the focus has turned to tuning it to get better outputs. “Prompt engineering” is the buzzword du jour, though while drafting this post, a search for “prompt engineer” turns up zero relevant job postings. Hmm.

I can imagine prompt responsibility being added to our existing responsibilities as departments vie to bend AI outputs to their missions. Security is going to want to lock things down. Marketing is going to want output to be “on brand.” Sales is going to want to, well, sell more stuff. We’re certainly in some danger of just shipping our org structures when it comes to crafting AI policies.

As an AI language model …

And it’s easy to imagine all these policies creating significant friction in the user experience, as we’re seeing today as OpenAI tunes ChatGPT to be more consumer friendly. Prompt output has been changing day to day, and users are left confused and unsure of how to achieve the results they seek. Experts one day, and noobs the next.

What’s not clear yet is how you track changes in AI performance. How do you gauge the effect of a newly released policy? How do you know if users are more productive? These are questions we still need to answer, I think. There’s a lot of ways to lose control of your customer relationships and reputation.

On the bright side, designers and researchers are potentially well-positioned to navigate these problems, facilitating cross-functional discussions and collecting user feedback and insights. Ensuring that your AI models are contributing positively to your business is going to take some effort to get right.