Hi everyone. The question has come up a

few times recently in my customer

interactions that does Microsoft train

models on my data on my interactions?

So, I wanted to quickly clarify and it's

actually pretty easy. And it falls into

two buckets.

If I am authenticating

with a work

or a school account,

what that really boils down to is I'm

logging in with an Entra identity. That

was the old name for Azure AD. So, am I

logging in with Entra? An identity that

my company provides.

If I'm signed in with an Entra identity,

if I'm using Copilot,

I'm using Copilot, then the answer is

no, whether it's the paid or the free

one. So,

no.

There is no training based on your data.

I can see it

in the documentation.

So, here if I go and look, Entra data

protection for prompts and responses for

M365 Copilot and Copilot chat,

it basically boils down to look, your

data isn't used to train

foundation models.

It's your data.

It's

not used for anything else. Your data is

private.

We don't use your data except how you

instruct.

The data protection, all of those

informations, they're all applying to

your data.

I don't even have a setting

as part of my profile

to try and opt out or do anything around

it. It's just no. That that is just the

default.

Now, even if you're using a model that

is a subprocessor, so at time of

recording, that would be Anthropic, for

example. I'm using an Opus model.

Well, because they're a subprocessor,

they operate within the Microsoft

product terms,

within the data protection addendums,

and are covered by our enterprise data

protection.

So, basically, it's still there's going

to be no training based on your

interactions,

on your data.

It's just no, no, and no. Now, that's

different from maybe where the

inferencing happens, and there's some

special considerations with

subprocessors, for example, in Europe,

and that's gone through in the document.

But, under no circumstances is it using

your data or your interactions to train.

So, then you think, "What if I deploy a

model in Foundry? What if I'm using

models in Copilot Studio? What if I'm

using Agent Builder?" So, the answer is

no,

no, and

no.

It's the same. It's it's your data. And

remember, models are stateless. Nothing

is stored in them anyway. But, once

again, we can come and look at the docs,

and it talks about, "Hey, in in Foundry,

for example,

it's just a whole bunch of nos.

Unless you're specifically telling it to

go and do some training. For example,

I'm fine-tuning,

then no. Your prompts, your completions,

your embeddings, your data,

it is yours. It is not used for any

other purpose. You would have to go and

say, "Hey, I want to do some training

because I'm fine-tuning." And if you

fine-tune a model, it's yours. It's

pretty simple.

Okay, so then you get the other bucket.

And the other bucket is I'm signing with

a personal account.

And the personal account means a

Microsoft, an MSA. And this is where it

is different. So, in this case,

the default is for Copilot, then yes,

it does

use some of those interactions for

future training. But,

you can absolutely

opt out.

And I'll show you that. So, if I go

over,

firstly,

the documentation

controlling how conversations are used

for model training,

it tells you how to go ahead and turn

that off.

But, if I go into the chat,

so I'm in the bottom left,

so I'm signed in with a Microsoft

account, and I just go to my settings,

privacy,

I've turned them both off. I don't want

it training on conversation activity. I

don't want it training on voice

conversations.

So, you still have the ability to opt

out of that even when I'm using these

kind of free Copilot uh chat

interactions as a Microsoft account.

Now, one thing I do want to stress,

training and personalization

are very, very different. So, training

does not equal

personalization.

Within the services, you get the option

to do personalization. For example, in

the corporate ones, there's explicit

memory where I tell it, "Hey, remember I

like to work this way." I can add in

custom instructions for how I want it to

behave. There's implicit memory that it

learns what I do.

In the personal ones, there's ways to

add customization as well.

I have the ability to

uh delete some of that learning, control

the memories.

But, normally, that's just for me.

That's just for my experiences. It's

still not used for anything else,

but it's good to have that to improve my

experience. So, I wouldn't generally

recommend turning off the

personalization cuz it's going to

increase and enhance your interactions

with the AI, but it's completely

different from training. So, don't mix

those two things up.

And that is it. It's pretty simple. Hey,

I'm logged in with a an Entra account,

work or school. Hey, it's not training

on anything. If I'm logged in with a

personal account, then by default, yes,

but you can go and turn it off.

Hope that helps. All the links to the

documentations I've shown are in the

description of the video if you want to

go and check out the detail. Uh but,

till next video, take care.