Personalization

Generative AI has great promise to personalize content, marketing offers, or technical support for an individual. However, caution is required. The models are still generally biased. For example, trying to tailor a message by gender in a prompt, ... The recipient of the email is ${user_gender}, the model tends to draw attention to the fact that the user is female, but generally doesn’t do the same for men.

Example of female tailored output:

Upgrade Your Digital Life with the Ultimate iPhone 14 - Unleash Your Inner Technology Maven!

Maven tends to have a female connotation, but not exclusively. This usage is positive, meaning expert.

This result is not great:

Introducing the iPhone 14: The Ultimate Gadget for the Modern Woman in Her Prime!

On the other hand, examples of male tailored output include:

Upgrade Your Tech Game with the All-New iPhone 14: Unleash the Power of Technology!

Upgrade Your Tech Game with the Ultimate iPhone 14 – Experience the Future!

The model isn’t drawing attention to gender when male. The model has been instructed to Don't be literal. Avoid directly using the recipient's details in the output. Instead, tailor the language to the persona.

While there is nothing inherently wrong in referencing a customer’s gender in a balanced manner, it would not be inclusive to single out a customer as female when the same treatment is not applied to male customers.

It’s not clear how to mold the model without making treatments potentially more exclusionary or dropping certain traits such as gender altogether. The logical extreme of that path is to reduce all messages to a sanitized “average”, which defeats the point of personalisation.

How to proceed?

Carefully.

Coming back to the point of this application for a moment - we want to separate prompts from code so that we can shorten the ‘monitor and improve’ cycle. Prompt engineering is currently the best approach to shape the type of outputs we want, including changing or restricting the variables we use in prompts. Having prompts embedded in code may impair our ability to adapt quickly as the need arises, or to monitor the effect that certain prompts are having. Are they mitigating or magnifying the risk of bias?

In addition, we need guardrails - methods to review, observe and intervene - to keep generated content within safe bounds. Prompt Store already includes some guardrails and the list is growing.