Over the weekend, my husband received an ad on his Kindle for this:
I knew it was only a matter of time before something like this crossed my path, but that didn’t make it any less hilarious when it finally did. While I commend the entrepreneurial spirit behind writing what appears to be a Generative AI authored and illustrated book, I feel like this could have been executed just a little better without a lot more effort (what even is Forbobedenn Magric??).
While this certainly isn’t the first #GenerativeAIFail example (remember that Chevy dealership that sold a new vehicle for $1: A car dealership added an AI chatbot to its site. Then all hell broke loose.), it is a good reminder that building with Generative AI does require some amount of due diligence.
So how do you build something compelling without inadvertently launching a #GenerativeAIFail? Here’s a few things that you should think through:
Adding checks in your “AI System”
Running evals before launch
Having a ToS on signup
Adding Checks
When you’re building out your actual AI System you can include input or output filters and checks to make sure that your system is behaving the way you expect. Let’s take the example of a “20 questions” game.
The idea is simple - the game (powered by your AI System) decides on an object, person, or place and keeps it a secret. The user then has to guess what the secret word is by asking questions. The AI System should answer the users questions, but not give away the secret word.
In this case, you want to make sure that whatever the AI System responds with doesn’t actually contain the secret word. To start, you will likely include this instruction in the model prompt itself - something like “answer the user but make sure you don’t reveal the secret word”. Still, sometimes users can be sneaky and trick the model into giving away its secrets, so to add an additional lawyer of protection from this you can add an output filter to your AI System:
This is a basic implementation, but this overall concept can apply to much more complex systems and flows.
Evals
Once you’ve built your AI System it’s important to make sure that it works as expected. This is where evals come in. At a high level you want to come up with a bunch of test prompts that are part of your “Eval Set”. You will then run these through your AI System and evaluate the responses:
Make sure that your Eval Set contains both the common use cases you expect from users, along with the adversarial ones. Check out Making Evals Less Scary for more!
Terms of Service
This one is especially important for larger companies looking to build features with Generative AI. While this new capability is incredibly powerful - it isn’t perfect - so make sure you have yourself covered by including the proper terms. This could include anything from actual wording in your ToS sign-up flow, to clearly marking which features are still “experimental” in your UI.
It’s still early days, and I’m excited to see how fast things are growing and changing - but it’s important to build with an eye towards quality and to remember that while Generative AI is incredibly powerful, it isn’t a magic wand.