Microsoft hits back at claims it slurps your Word, Excel files to train AI models

Confusion over Connected Experiences settings in 365 apps spark concerns

Updated Microsoft's Connected Experiences option in its productivity suite has been causing consternation amid accusations that the default setting might allow Microsoft to train AI models using customers' Word and Excel documents and other data.

The Windows giant vehemently denies the claims. A spokesperson told The Register: "In Microsoft 365 consumer and commercial applications, Microsoft does not use customer data to train large language models without your permission."

We asked Microsoft what it meant by "permission" and if the permission was opt-in or opt-out, and the IT titan has yet to respond.

Connected Experiences has long been a part of Microsoft Office. Want to do some translation? You're probably using Connected Experiences. Transcribe a recording? Again, Connected Experiences. Do some grammar checking in Word? Connected Experiences will be analyzing your content.

The spokesperson said: "The Connected Services setting is an industry standard setting that enables features that require an internet connection. Connected experiences play a significant role in enhancing productivity by integrating your content with resources available on the web. These features allow applications to provide more intelligent and personalized services."

In recent weeks, users have been looking more deeply at what Microsoft is doing with all this data, and some have worried that it is being used to train the mega-corp's internal AI systems, something Microsoft says it is not.

The suggestion was circulated on social media platforms over the weekend.

A look at a consumer Windows 11 machine running Microsoft 365 2410 showed that the Connected Experiences setting was checked as on by default. But did that mean the customer's content was being used to train an AI? It's unlikely but not outside the realms of possibility.

It is, however, extremely unlikely that content produced by Education and Enterprise users of Microsoft 365 would be collected in this way. After all, security policies are in place to control the Connected Experiences option if that's a concern.

The difficulty folks face is that despite Microsoft's protestations, its privacy statement (as of November 2024) does permit it to do all manner of things with the data it collects. And how does it use that data? "As part of our efforts to improve and develop our products, we may use your data to develop and train our AI models."

In August, Microsoft said it would be using consumer data from Copilot, Bing, and Microsoft Start to train Copilot's generative AI models. At the time, the biz said it would allow customers to opt out and would start displaying the opt-out control in October. It also said it wouldn't be conducting training on consumer data from the European Economic Area.

Could the same apply to documents created by people in the company's flagship productivity suite? It's quite a leap from training a generative AI on what goes into Copilot to making use of Word and Excel documets under the guise of Connected Experiences. The two are very different services.

So, on the one side, Microsoft is clear: It does not use customer data to train models. On the other hand, "we may use your data to develop and train our AI models."

As for what it means by "data" in its privacy statement, the biz writes: "You provide some of this data directly, and we get some of it by collecting data about your interactions, use, and experiences with our products."

The fact that concerns are being raised indicates some users are concerned about Microsoft's AI obsession. The Windows maker must therefore maintain clarity and transparency over what is and is not going to be absorbed into the its models. ®

Updated to add at 1645 UTC, November 27

In response to our question on what Microsoft meant by "permission," if it was an opt-out or opt-in thing, and where a customer would give this permission, the company pondered for over a day before responding: "There are circumstances where enterprise customers may want or consent to our use of their data for foundation model training, such as custom model development requested by the customer."

So now you know.