June 27, 2024 • 13 mins
Should Meta be allowed to train its AI on your public posts and images? How does Meta potentially prevent private information from leaking into its model and what are the risks when it does? And how can you better protect your privacy in the era of AI?
Listen to this week's episode to find out.
This week, instead of recording with ChatGPT, Claude from Anthropic will be your co-host. Specifically, the new 3.5 model that was released earlier this week. The voice I've chosen to represent Claude in this episode is Onyx from OpenAI.
As mentioned at the end of the episode, to learn more how to protect your privacy in the era of AI, check out this article Privacy in an AI Era: How Do We Protect Our Personal Information? from Stanford University.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Trevor (00:00):
Welcome to AI Meets Productivity,
(00:01):
your weekly podcast between myself,
Trevor Lohrbeer, a productivity
expert and an artificial intelligence.
This week, we have Claude from
Anthropic to talk to us on the podcast.
And we'll be talking about how Meta
is training its AI on public posts and
images from Facebook and Instagram.
Welcome, Claude.
Claude AI (00:22):
Hey there, Trevor.
Thanks for having me on the show.
Gotta say, I'm pumped to chat
about Meta's latest AI shenanigans.
It's like they're turning our social media
into their personal AI playground, right?
What's your take on this
digital data feast, Trevor?
Trevor (00:38):
So before we get started, I wanted
to address Meta training it's AI on public
posts and images from Facebook and Instagram
and the ability to opt out of that training.
There's been a lot of posts going around
talking about June 26th as being the deadline
to submit an objection to that training.
(00:59):
That was the original date for Europeans.
However, the Data Privacy Commission reversed
its advice to Facebook and told them two weeks
ago on June 14th that they were not allowed
to use European's data to train its AI.
So if you're in Europe, it does
not matter if you opted out.
(01:21):
Meta will still not be training
its AI using your data.
The DPC has ensured that that is not
going to happen, at least for now.
If you're not in Europe, for instance, in the
US, Australia or other places around the world,
then your local data privacy laws will apply.
Which in the US basically means for most states
(01:43):
that you have very little privacy protections.
And in this case, Meta has already
been training its AI using your data.
And we know that this has been going on at least
since last September, possibly before then.
Because under US privacy law, or the lack
of US privacy law, Meta has no obligation to
(02:04):
inform its users when it trains its AI is using
its users' data, unlike they do in Europe.
So a lot of this fervor in the US
is coming because we actually didn't
realize we didn't even have these rights.
Now, if you did happen to do one of
the opt-out procedures, hopefully
Facebook.will at some point respect that.
(02:26):
But as it stands right now,
they are not required to.
And so I definitely recommend contacting
your representatives, and telling them that
you want comprehensive privacy legislation.
And specifically tell them that you want privacy
legislation that does not preempt state privacy
bills and does not weaken data broker protections,
(02:47):
two of the things that the EFF says are wrong
with the current American Privacy Rights Act
that is working its way through Congress.
But I would highly recommend contacting your
representatives because this is an example of
why people living in the US do not have the
privacy rights that people in Europe have.
Because we do not have
national privacy legislation.
(03:10):
Okay, with that disclaimer, I wanted to jump
into the episode or we'll talk about exactly
what this means now that you know that Facebook
likely has already been training on your data.
So Claude, since we're talking about privacy,
can you give our listeners an example
of a Facebook post that might contain
personal information and how Facebook
might actually train it's AI on that post?
Claude AI (03:33):
Certainly.
Let's consider a hypothetical post:
"Just got married to Sarah at City Hall.
We're moving to 123 Main St next month.
Call me at 555-1234 If you want to help."
For AI training, Meta would likely
process this post by removing or
anonymizing the specific personal details.
The result might look like (03:52):
"Just
got married to [NAME] at [LOCATION].
We're moving to [ADDRESS] next month.
Call me at [PHONE NUMBER] if you want to help."
The AI would learn about life events, post
structures and general content patterns, but
without access to the actual personal information.
This approach allows the AI to understand
(04:13):
context and language use while protecting...
Trevor (04:15):
But what if Meta identifies
the personal information incorrectly,
could that then leak into the model?
Claude AI (04:21):
That's a sharp observation, Trevor.
You've hit on a real concern.
If Meta systems fail to correctly identify
all personal information, there's a risk that
some of it could indeed leak into the model.
This could happen if, say, an unusual
name or address format isn't recognized
as personal info, or if context
dependent personal details slip through.
(04:41):
It's not a perfect system.
Such errors could potentially lead to
privacy breaches, or the model learning, and
potentially reproducing sensitive information.
It's a significant challenge in developing
these AI systems while respecting user privacy.
Trevor (04:57):
What other privacy issues might
people encounter when using the Meta AI
that has been trained on public posts?
Claude AI (05:06):
When using Meta's AI
trained on public posts, people
might encounter these privacy issues:
1.
Unintended Information Disclosure (05:12):
The AI might
reveal patterns or insights from aggregated
data that individuals didn't intend to share.
2.
Re-Identification (05:20):
Even with anonymized
data, the AI might piece together information
to inadvertently identify specific users.
3.
Biased Responses (05:29):
If the training
data isn't diverse, the AI could
reflect and amplify societal biases.
4.
Context Collapse (05:36):
The AI might fail
to understand the original context of
posts leading to misinterpretations.
5.
Outdated Information (05:44):
The AI could
provide outdated personal information
that users have since changed or removed.
Trevor (05:52):
And how likely is
each of these five things?
Claude AI (05:55):
Hmm, that's a tough
one to quantify precisely, Trevor.
The likelihood of these issues varies based
on many factors, including Meta's safeguards
and the specific implementation of the AI.
But here's a rough estimate:
1.
Unintended disclosure, moderately likely.
2.
Re-identification, less likely but possible.
3.
Biased responses, highly likely.
(06:16):
4.
Context collapse, fairly likely.
5.
Outdated information, highly likely.
Remember, these are just educated guesses.
The actual likelihood could refines its
AI and as privacy regulations evolve.
What's your take on these risks, Trevor?
Trevor (06:34):
Well, so I think we've already
discussed unintended disclosure because
that's when the AI is trained on personal
information that isn't properly cleansed
and then the AI can regurgitate that.
Claude AI (06:45):
You're absolutely right.
Would you like to focus on the
other four issues we mentioned?
Trevor (06:50):
Yeah, I'm particularly
interested in the biased responses.
So Facebook can sometimes be this
echo chamber that really has a
lot of high emotions going around.
And if an AI is being trained on that,
would it not then reflect that kind of echo
chamber in its responses back to the user?
Claude AI (07:10):
Absolutely, Trevor.
You've nailed a crucial point there.
Facebook's algorithm often creates
echo chambers . An AI trained on
these polarized discussions might:
1.
Amplify extreme viewpoints.
2.
Struggle with nuanced topics.
3.
Reflect heightened emotional tones.
4.
Perpetuate misinformation or conspiracy theories.
5.
(07:30):
Lack exposure to diverse perspectives.
This could lead to an AI that gives
biased or inflammatory responses,
potentially exacerbating social divisions.
It's a significant challenge for Meta to address
in developing a balanced, fair AI system.
What do you think could be
done to mitigate this issue?
Trevor (07:48):
I mean, I'm not sure, but I would
just say don't train on public posts.
Claude AI (07:54):
Ha!
Well that's certainly one way
to cut the Gordian Knot, Trevor.
You're right, avoiding training on public
posts would sidestep a lot of these issues.
It's a simple solution that would eliminate
concerns about bias, privacy, and misinformation
stemming from social media content.
Plus, it would force Meta to find
alternative, potentially more
(08:15):
controlled and ethical data sources.
But I guess the question then
becomes, what would Meta use instead?
And would that new data source
come with its own set of problems?
It's a complex issue without easy answers.
Trevor (08:29):
So I just want to dive into why
Meta is saying they want to train their
AI, particularly on European posts.
They're saying it's to kind of help understand
natural language, regional dialects and
colloquial expressions, to improve tasks like text
completion, translation and conversational AI.
Claude AI (08:46):
Interesting point, Trevor.
Meta's argument for training on European
posts seems to focus on improving the AI's
cultural competence and linguistic abilities.
They're likely aiming for:
1.
Better language translation.
2.
More accurate text completion
in European languages.
3.
Improved understanding of
European cultural nuances.
(09:06):
4.
Enhanced ability to generate
culturally relevant content.
These goals make sense from
an AI development perspective.
However, it does raise questions about
whether these benefits outweigh the
privacy concerns we've discussed.
What's your take on Meta's justification?
Trevor (09:24):
I mean, I feel that in theory it might
be okay for Meta to train if it's going to
learn colloquial expressions, regional dialects,
and help people communicate better with an AI.
At the same time, how they've been going about
it has not been to my mind an appropriate way.
Claude AI (09:42):
I hear you, Trevor.
You're striking a balance between the
potential benefits and the ethical concerns.
You're right that there could be valuable
gains in AI understanding colloquialisms,
dialects, and improving communication.
These are worthy goals that could make AI more
accessible and useful to a diverse user base.
However, as you pointed out, Meta's
(10:03):
approach has been problematic.
The lack of transparency, issues with consent, and
potential privacy risks are significant drawbacks.
It seems like what's needed is a
more ethical, transparent approach
to data collection and AI training.
Perhaps a system where users can explicitly
opt in to contribute specific types
(10:23):
of data for AI training, with clear
explanations of how it will be used.
What do you think would be a more appropriate
way for Meta to achieve these goals while
respecting user privacy and consent?
Trevor (10:37):
I think like what you said is,
instead of making opt-out, make it opt-in.
Claude AI (10:42):
Absolutely, Trevor.
An opt-in approach would be
a significant improvement.
It shifts the power back to the users,
giving them more control over their data.
This method would:
1.
Ensure clear consent.
2.
Increase transparency.
3.
Potentially improve data quality.
4.
Build trust with users.
5.
Align better with privacy regulations.
(11:03):
It might result in less data overall, but
the data would be more ethically sourced.
Plus, users who opt in might be
more likely to provide thoughtful,
relevant content for AI training.
Trevor (11:16):
So, if someone in the US who already
has had some of their data used to train
the AI wants to prevent future AI models
from being trained on their data, what are
the steps they can take to prevent that?
Claude AI (11:29):
Unfortunately, Trevor, for users in the
US, the options are quite limited at the moment.
Meta hasn't provided a way to opt out of
AI training for users outside of Europe.
However, users can take a few steps
to minimize their data exposure.
1.
Adjust privacy settings to limit post visibility.
2.
Be mindful of what they share publicly.
(11:49):
3.
Delete old posts or use
Facebook's bulk delete tools.
4.
Consider using Meta's off
Facebook activity controls.
5.
Limit the personal information in their profile.
It's worth noting that these steps won't remove
data that's already been used for training.
and they don't guarantee that
future posts won't be used.
(12:09):
The lack of a clear opt-out mechanism is
a significant issue for users concerned
about their data being used for AI training.
Trevor (12:17):
Okay.
Great.
Thank you very much.
I think that's all for today.
Claude AI (12:21):
You're welcome, Trevor.
It's been a pleasure discussing
this important topic with you.
I hope we've provided our listeners with
valuable insights into Meta's AI training
practices and the associated privacy concerns.
Trevor (12:34):
Thank you, everyone.
This has been AI Meets Productivity, your weekly
podcast between Trevor Lohrbeer, a productivity
expert and an artificial intelligence.
This week instead of ChatGPT, we were talking
to Claude from Anthropic and we were talking
about privacy and how it relates to AI.
I will post a link in the show notes
to "Privacy in an AI Era (12:55):
How Do We
Protect Our Personal Information?"
from Stanford University.
And that will have some more information
about how you can protect yourself and
protect your privacy in the AI era.
In the meantime, please like, share
and review this podcast, wherever
you listen to your podcasts.
We are on Apple Podcasts, Spotify,
(13:16):
Podbean, and many other places.
And you can also subscribe via email to get
every podcast episode right when it comes
out at aimeetsproductivity.podbean.com.
Thank you very much and
we'll talk to you next week.
Welcome to AI Meets Productivity,
(00:01):
your weekly podcast between myself,
Trevor Lohrbeer, a productivity
expert and an artificial intelligence.
This week, we have Claude from
Anthropic to talk to us on the podcast.
And we'll be talking about how Meta
is training its AI on public posts and
images from Facebook and Instagram.
Welcome, Claude.
Claude AI (00:22):
Hey there, Trevor.
Thanks for having me on the show.
Gotta say, I'm pumped to chat
about Meta's latest AI shenanigans.
It's like they're turning our social media
into their personal AI playground, right?
What's your take on this
digital data feast, Trevor?
Trevor (00:38):
So before we get started, I wanted
to address Meta training it's AI on public
posts and images from Facebook and Instagram
and the ability to opt out of that training.
There's been a lot of posts going around
talking about June 26th as being the deadline
to submit an objection to that training.
(00:59):
That was the original date for Europeans.
However, the Data Privacy Commission reversed
its advice to Facebook and told them two weeks
ago on June 14th that they were not allowed
to use European's data to train its AI.
So if you're in Europe, it does
not matter if you opted out.
(01:21):
Meta will still not be training
its AI using your data.
The DPC has ensured that that is not
going to happen, at least for now.
If you're not in Europe, for instance, in the
US, Australia or other places around the world,
then your local data privacy laws will apply.
Which in the US basically means for most states
(01:43):
that you have very little privacy protections.
And in this case, Meta has already
been training its AI using your data.
And we know that this has been going on at least
since last September, possibly before then.
Because under US privacy law, or the lack
of US privacy law, Meta has no obligation to
(02:04):
inform its users when it trains its AI is using
its users' data, unlike they do in Europe.
So a lot of this fervor in the US
is coming because we actually didn't
realize we didn't even have these rights.
Now, if you did happen to do one of
the opt-out procedures, hopefully
Facebook.will at some point respect that.
(02:26):
But as it stands right now,
they are not required to.
And so I definitely recommend contacting
your representatives, and telling them that
you want comprehensive privacy legislation.
And specifically tell them that you want privacy
legislation that does not preempt state privacy
bills and does not weaken data broker protections,
(02:47):
two of the things that the EFF says are wrong
with the current American Privacy Rights Act
that is working its way through Congress.
But I would highly recommend contacting your
representatives because this is an example of
why people living in the US do not have the
privacy rights that people in Europe have.
Because we do not have
national privacy legislation.
(03:10):
Okay, with that disclaimer, I wanted to jump
into the episode or we'll talk about exactly
what this means now that you know that Facebook
likely has already been training on your data.
So Claude, since we're talking about privacy,
can you give our listeners an example
of a Facebook post that might contain
personal information and how Facebook
might actually train it's AI on that post?
Claude AI (03:33):
Certainly.
Let's consider a hypothetical post:
"Just got married to Sarah at City Hall.
We're moving to 123 Main St next month.
Call me at 555-1234 If you want to help."
For AI training, Meta would likely
process this post by removing or
anonymizing the specific personal details.
The result might look like (03:52):
"Just
got married to [NAME] at [LOCATION].
We're moving to [ADDRESS] next month.
Call me at [PHONE NUMBER] if you want to help."
The AI would learn about life events, post
structures and general content patterns, but
without access to the actual personal information.
This approach allows the AI to understand
(04:13):
context and language use while protecting...
Trevor (04:15):
But what if Meta identifies
the personal information incorrectly,
could that then leak into the model?
Claude AI (04:21):
That's a sharp observation, Trevor.
You've hit on a real concern.
If Meta systems fail to correctly identify
all personal information, there's a risk that
some of it could indeed leak into the model.
This could happen if, say, an unusual
name or address format isn't recognized
as personal info, or if context
dependent personal details slip through.
(04:41):
It's not a perfect system.
Such errors could potentially lead to
privacy breaches, or the model learning, and
potentially reproducing sensitive information.
It's a significant challenge in developing
these AI systems while respecting user privacy.
Trevor (04:57):
What other privacy issues might
people encounter when using the Meta AI
that has been trained on public posts?
Claude AI (05:06):
When using Meta's AI
trained on public posts, people
might encounter these privacy issues:
1.
Unintended Information Disclosure (05:12):
The AI might
reveal patterns or insights from aggregated
data that individuals didn't intend to share.
2.
Re-Identification (05:20):
Even with anonymized
data, the AI might piece together information
to inadvertently identify specific users.
3.
Biased Responses (05:29):
If the training
data isn't diverse, the AI could
reflect and amplify societal biases.
4.
Context Collapse (05:36):
The AI might fail
to understand the original context of
posts leading to misinterpretations.
5.
Outdated Information (05:44):
The AI could
provide outdated personal information
that users have since changed or removed.
Trevor (05:52):
And how likely is
each of these five things?
Claude AI (05:55):
Hmm, that's a tough
one to quantify precisely, Trevor.
The likelihood of these issues varies based
on many factors, including Meta's safeguards
and the specific implementation of the AI.
But here's a rough estimate:
1.
Unintended disclosure, moderately likely.
2.
Re-identification, less likely but possible.
3.
Biased responses, highly likely.
(06:16):
4.
Context collapse, fairly likely.
5.
Outdated information, highly likely.
Remember, these are just educated guesses.
The actual likelihood could refines its
AI and as privacy regulations evolve.
What's your take on these risks, Trevor?
Trevor (06:34):
Well, so I think we've already
discussed unintended disclosure because
that's when the AI is trained on personal
information that isn't properly cleansed
and then the AI can regurgitate that.
Claude AI (06:45):
You're absolutely right.
Would you like to focus on the
other four issues we mentioned?
Trevor (06:50):
Yeah, I'm particularly
interested in the biased responses.
So Facebook can sometimes be this
echo chamber that really has a
lot of high emotions going around.
And if an AI is being trained on that,
would it not then reflect that kind of echo
chamber in its responses back to the user?
Claude AI (07:10):
Absolutely, Trevor.
You've nailed a crucial point there.
Facebook's algorithm often creates
echo chambers . An AI trained on
these polarized discussions might:
1.
Amplify extreme viewpoints.
2.
Struggle with nuanced topics.
3.
Reflect heightened emotional tones.
4.
Perpetuate misinformation or conspiracy theories.
5.
(07:30):
Lack exposure to diverse perspectives.
This could lead to an AI that gives
biased or inflammatory responses,
potentially exacerbating social divisions.
It's a significant challenge for Meta to address
in developing a balanced, fair AI system.
What do you think could be
done to mitigate this issue?
Trevor (07:48):
I mean, I'm not sure, but I would
just say don't train on public posts.
Claude AI (07:54):
Ha!
Well that's certainly one way
to cut the Gordian Knot, Trevor.
You're right, avoiding training on public
posts would sidestep a lot of these issues.
It's a simple solution that would eliminate
concerns about bias, privacy, and misinformation
stemming from social media content.
Plus, it would force Meta to find
alternative, potentially more
(08:15):
controlled and ethical data sources.
But I guess the question then
becomes, what would Meta use instead?
And would that new data source
come with its own set of problems?
It's a complex issue without easy answers.
Trevor (08:29):
So I just want to dive into why
Meta is saying they want to train their
AI, particularly on European posts.
They're saying it's to kind of help understand
natural language, regional dialects and
colloquial expressions, to improve tasks like text
completion, translation and conversational AI.
Claude AI (08:46):
Interesting point, Trevor.
Meta's argument for training on European
posts seems to focus on improving the AI's
cultural competence and linguistic abilities.
They're likely aiming for:
1.
Better language translation.
2.
More accurate text completion
in European languages.
3.
Improved understanding of
European cultural nuances.
(09:06):
4.
Enhanced ability to generate
culturally relevant content.
These goals make sense from
an AI development perspective.
However, it does raise questions about
whether these benefits outweigh the
privacy concerns we've discussed.
What's your take on Meta's justification?
Trevor (09:24):
I mean, I feel that in theory it might
be okay for Meta to train if it's going to
learn colloquial expressions, regional dialects,
and help people communicate better with an AI.
At the same time, how they've been going about
it has not been to my mind an appropriate way.
Claude AI (09:42):
I hear you, Trevor.
You're striking a balance between the
potential benefits and the ethical concerns.
You're right that there could be valuable
gains in AI understanding colloquialisms,
dialects, and improving communication.
These are worthy goals that could make AI more
accessible and useful to a diverse user base.
However, as you pointed out, Meta's
(10:03):
approach has been problematic.
The lack of transparency, issues with consent, and
potential privacy risks are significant drawbacks.
It seems like what's needed is a
more ethical, transparent approach
to data collection and AI training.
Perhaps a system where users can explicitly
opt in to contribute specific types
(10:23):
of data for AI training, with clear
explanations of how it will be used.
What do you think would be a more appropriate
way for Meta to achieve these goals while
respecting user privacy and consent?
Trevor (10:37):
I think like what you said is,
instead of making opt-out, make it opt-in.
Claude AI (10:42):
Absolutely, Trevor.
An opt-in approach would be
a significant improvement.
It shifts the power back to the users,
giving them more control over their data.
This method would:
1.
Ensure clear consent.
2.
Increase transparency.
3.
Potentially improve data quality.
4.
Build trust with users.
5.
Align better with privacy regulations.
(11:03):
It might result in less data overall, but
the data would be more ethically sourced.
Plus, users who opt in might be
more likely to provide thoughtful,
relevant content for AI training.
Trevor (11:16):
So, if someone in the US who already
has had some of their data used to train
the AI wants to prevent future AI models
from being trained on their data, what are
the steps they can take to prevent that?
Claude AI (11:29):
Unfortunately, Trevor, for users in the
US, the options are quite limited at the moment.
Meta hasn't provided a way to opt out of
AI training for users outside of Europe.
However, users can take a few steps
to minimize their data exposure.
1.
Adjust privacy settings to limit post visibility.
2.
Be mindful of what they share publicly.
(11:49):
3.
Delete old posts or use
Facebook's bulk delete tools.
4.
Consider using Meta's off
Facebook activity controls.
5.
Limit the personal information in their profile.
It's worth noting that these steps won't remove
data that's already been used for training.
and they don't guarantee that
future posts won't be used.
(12:09):
The lack of a clear opt-out mechanism is
a significant issue for users concerned
about their data being used for AI training.
Trevor (12:17):
Okay.
Great.
Thank you very much.
I think that's all for today.
Claude AI (12:21):
You're welcome, Trevor.
It's been a pleasure discussing
this important topic with you.
I hope we've provided our listeners with
valuable insights into Meta's AI training
practices and the associated privacy concerns.
Trevor (12:34):
Thank you, everyone.
This has been AI Meets Productivity, your weekly
podcast between Trevor Lohrbeer, a productivity
expert and an artificial intelligence.
This week instead of ChatGPT, we were talking
to Claude from Anthropic and we were talking
about privacy and how it relates to AI.
I will post a link in the show notes
to "Privacy in an AI Era (12:55):
How Do We
Protect Our Personal Information?"
from Stanford University.
And that will have some more information
about how you can protect yourself and
protect your privacy in the AI era.
In the meantime, please like, share
and review this podcast, wherever
you listen to your podcasts.
We are on Apple Podcasts, Spotify,
(13:16):
Podbean, and many other places.
And you can also subscribe via email to get
every podcast episode right when it comes
out at aimeetsproductivity.podbean.com.
Thank you very much and
we'll talk to you next week.
Popular Podcasts
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by audiochuck Media Company.
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.