This Setting Makes Me Mad

Thanks to a post on Ohana Slack I learned recently of a new setting that Salesforce quietly slipped into the back end and defaulted to ON.

Unless you have been ignoring the news recently, you probably won't be surprised that this new setting relates to AI. Specifically, the setting allows Salesforce to access "customer data" to train AI models.

If you're wondering what "customer data" entails, Salesforce never directly specifies. That, to me, means they define it as “the values in every field on every object, custom or standard, that you have in your org.” In other words, they are going to use all of your data to train AI models.

Ask yourself who owns or will own those models? [Hint: It's not you.]

Setup > Opt Out of Customer Data Access

You can find the setting at Setup > Opt Out of Customer Data Access. When you get there, you'll see that it is ON by default.

If you don't like this (And I don't!), Move the slider to turn it off. You even have to endure an "Are you sure?" screen.

Click again to confirm.

As soon as I learned of this I logged into each of my client orgs and turned it off.

The Clue

Maybe you're wondering why I have made such an unfavorable assumption about how Salesforce has defined "customer data"? Well, if you follow the help link to "Learn More" at the bottom of that setup page, it specifies in which situations this setting is defaulted to ON. It's not defaulted for Government Cloud instances nor if your organization "has decided to opt out, or has previously opted out, of sharing Customer Data with Salesforce."

More Nuance

I was pointed to this help article for a little more explanation of the setting. And I understand that, particularly for predicative AI, most organizations aren't going to have enough of their own data to train models. So I can understand the argument for there being a benefit to training global predictive models. I actually wrote in 2022 about trying to use Einstein prediction builder and the fact that few orgs have enough data to develop predictions on their own.

I will also note that this setting "doesn’t affect Salesforce’s zero-data retention (ZDR) policy with third-party Large Language Models (LLMs)." So [at least for now] this setting is only about training predictive AI and not about training generative AI.

If you read my follow-up article about Einstein Prediction Builder, you may remember that I didn't find it that useful, even using the "pre-built models" from Salesforce. And ask yourself if you've heard anyone else talking about Einstein Prediction Builder in the last several years? The freemium version of Prediction Builder also looks like it's gone away. So is there even any benefit to us if Salesforce works on these models? Or does the benefit mainly accrue to the company that will come around trying to sell them in the future?

A Violation of Trust

Regardless of how hard it might be to build good prediction models, that does not justify taking your organization's private data without affirmative consent and without compensation. Salesforce constantly touts Trust as their "#1 value." Dropping a new privacy-piercing setting into your system and defaulting it to ON feels like a major violation of that trust.

I also just don’t like the idea of handing over my nonprofit clients' data to train a model that is going to make billions that won't benefit nonprofits.