No one asked before training AI on your photos, your posts and your voice. Here is how to switch off what you can in about fifteen minutes, and the truth about what no setting can undo.
It's the silence consent that most companies are eager for us to accept.
The root problem policy researchers are flagging in other AI contexts right now systems built so the burden of protection sits entirely on the person with the least power in the relationship.
Opt-out-by-default isn't just a dark UX pattern, it's a legal default, and the EU's slow move toward opt-in standards is really the only lever that fixes it at the source.
Everything else, fifteen minutes of toggles, registries, artist tools, is damage control applied after the architecture has already decided who has to do the work.
Genuinely useful as a checklist regardless. But it's a good reminder that "go find the switch" will always lose to "the switch should never have defaulted to on" as a long-term fix. We need to change that.
Indeed Dr. Sam, as users, citizens and as ourselves, we have, in my opinion, all the necessary tools and means to "vote" on how things should be. And the simples one, and perhaps the most difficult for some, is where we decide to spend our money and also our time. Simple decision, that remind us what is really important to us.
Thank you for these. Have switched off one that I found but not all of them yet. Thanks for the detailed breakdown on how. Reminds me that I have probably put the most private info uploaded into perplexity but haven’t checked what their privacy settings are yet. For whatever reason i felt less concerned about my data being with them as it’s a different kind of company but it still is what it is. Hope I can find out what the best measures are there. If you think of updating the post with anything else, maybe consider perplexity for other users. Especially all the folks who took their free 12 months pro account promotion they ran in partnesship with venmo & paypal and then started using it without updating the settings.
You are so welcome Kevin. Please do leave a comment if you work out how to do this in Perplexity. As that would be very helpful for many readers. I am sure. 🙏
Perplexity privacy has an opt out setting for prompts, BUT that is overridden if the user uses "public" or "share" in their processes. So, secure behaviour is another knowledge/skill for "AI literacy".
I have an anthropological theory to explain individual choices on gen ai, if anyone has similar hunches. Only tested on small cohort of online executive Masters students so far.
A few days ago, Substack showed me a post very, very specifically targeted to a specific phrase I used in a Google Meet with my therapist that morning. It was *very* specific. Not something I have ever searched, Googled, posted about, talked about with family. It was not on the map ANYWHERE outside of that conversation prior to that morning. It's not even adjacent to any of my usual interests in any way. At all. And I would testify to that in a court of law under oath and more importantly, I would remember if it had ever come up before that.
I'm really, really curious about how and when I EVER consented to Google listening to my therapy sessions much less sharing that data with any social media platform. And that was the second time something like that happened in two weeks.
The other was a very weird content crossover from YouTube to Substack but in a very, very indirect way. That I also never consented to.
I'd be terrified of this capability if it surprised me in any way. It's just the tip of the surveillance iceberg unfortunately. The worst part is that both times it was useful information but still intrusive in a very uncanny valley kind of way. I felt disgust like my skin was crawling and had a feeling of being violated in some inexpressible way.
What stood out to me is that this isn't just a conversation about privacy. It's a conversation about how we define meaningful consent.
There's an important difference between information being technically available and people realistically understanding what they've agreed to. As AI becomes more integrated into everyday products, designing for genuine understanding may become just as important as designing for compliance.
Sam, thank you for this. I went through each item. I am happy to report that I had already blocked tracking and training on all of my accounts. Frankly, I am very proud of myself! 😆 I thought surely I had missed something. The main point I'd like to make is the difference between privacy in Europe and America. It's night and day. Americans truly don't see and many don't even care about companies hijacking their entire lives. We should look to GDPR for oversight and regulations. The worst of it is in our public schools.
Well done Traci! And I agree entirely. I also think that even Europeans need to better understand why data sovereignty is important and the fact that if a platform is free, the cost is probably your data.
Thank you! You can also opt out of substack using your writing to train AI. I’ve turned a lot of these off a while ago, but I’ll have to revisit just to double check.
Great practical guide. The toggles are worth the fifteen minutes, but the part you're honest about at the end is the part I keep thinking about: the protection was never the setting, it was the price of reading you.
I published a piece yesterday on the same thread from a different angle: what happens when AI drops the cost of reading everyone to zero, age verification hands over the last identifier, and the log of how you think sits on a server with no legal privilege. The toggles protect tomorrow. The architecture is already built.
Sam, I literally love your work, you and two other in particular just give me so much inspiration! The consent you have to chase through buried settings is exhaustion dressed up as choice" is the line means everything to me. I went and checked the Claude toggle right after reading — I had no Scooby-Doo that the policy flipped in August 2025, and given how much personal context I run through Claude day to day, that's not a small thing to have missed.
What strikes me reading this next to your sycophancy piece is that they're the same shape of problem from two different angles. Sycophancy is the AI quietly giving you less than it should. This is the AI quietly taking more than it should. Both rely on the same thing: you not noticing, because nothing announces itself. The fix in both cases is the same discipline — go and look, on purpose, at the thing that's designed to be skipped past.
Appreciate you doing the unglamorous work of listing the actual menu paths. I'd have given up after Google LOL!
These are helpful to stanch the bleeding, but there's no recovery of blood already leeched.
We also can't call AI out for doing something nefariously novel: to some degree, this is behavior tracking, (now real time) personalization, aggregation of transactional, location, and social media activity - on steroids.
Still, if the adoption rate of such defenses is high, some of the data AIs have already collected will age out of relevancy. And maybe the desperation for current data drives change in the default behavior of data collectors.
I had already made alterations to most of these, but not to social media. Very helpful.
This makes me nervous: “Treat the toggle as a real step rather than a cast-iron guarantee, because how fully these opt-outs hold has been questioned.”
The other thing I do is delete the thread at the end of each day and start a new thread the next day (when it’s something important).
This seems to work because I definitely have to deal with the hassle of interacting with Claude or ChatGPT as if it is new to whatever I was talking with it about the day before.
But I’m wondering, based on your quote, if changing the settings and deleting the thread is still not enough?
I worry that it might not be Kerry. I have had instances of Claude chats bleeding over into other ones that I thought were in 'Incognito mode'. Sadly the truth is that we should start from a position of complete distrust with these companies.
It's the silence consent that most companies are eager for us to accept.
The root problem policy researchers are flagging in other AI contexts right now systems built so the burden of protection sits entirely on the person with the least power in the relationship.
Opt-out-by-default isn't just a dark UX pattern, it's a legal default, and the EU's slow move toward opt-in standards is really the only lever that fixes it at the source.
Everything else, fifteen minutes of toggles, registries, artist tools, is damage control applied after the architecture has already decided who has to do the work.
Genuinely useful as a checklist regardless. But it's a good reminder that "go find the switch" will always lose to "the switch should never have defaulted to on" as a long-term fix. We need to change that.
The hardest of agrees Diamantino. Hopefully enough people start voting with their switches and then their wallets for these companies to take note...
Indeed Dr. Sam, as users, citizens and as ourselves, we have, in my opinion, all the necessary tools and means to "vote" on how things should be. And the simples one, and perhaps the most difficult for some, is where we decide to spend our money and also our time. Simple decision, that remind us what is really important to us.
Thank you Sam! I turned it off in Claude as soon as you posted that note back then ❤️
Wahoo! Honestly I would never have knowm if not for @ToxSec 🙏
🙏🔥 super glad to hear it hah 😁
Not all superheroes wear capes. Some of them wear beanies too.
He always saves the day, doesn't he 😊
😛💚
Thank you for these. Have switched off one that I found but not all of them yet. Thanks for the detailed breakdown on how. Reminds me that I have probably put the most private info uploaded into perplexity but haven’t checked what their privacy settings are yet. For whatever reason i felt less concerned about my data being with them as it’s a different kind of company but it still is what it is. Hope I can find out what the best measures are there. If you think of updating the post with anything else, maybe consider perplexity for other users. Especially all the folks who took their free 12 months pro account promotion they ran in partnesship with venmo & paypal and then started using it without updating the settings.
You are so welcome Kevin. Please do leave a comment if you work out how to do this in Perplexity. As that would be very helpful for many readers. I am sure. 🙏
Perplexity privacy has an opt out setting for prompts, BUT that is overridden if the user uses "public" or "share" in their processes. So, secure behaviour is another knowledge/skill for "AI literacy".
Thanks Richard. This is very helpful. 🙏
I have an anthropological theory to explain individual choices on gen ai, if anyone has similar hunches. Only tested on small cohort of online executive Masters students so far.
What is it Richard?
Grid-group cultural theory
Thank you. Going down a rabbit hole now. 🙏
Not too far. I have a draft conceptual paper. But I'd rather present empirically.
The students I tested it on saw their position in the 2×2 before I explained.
Love this as an educational tool Richard. 🙏
A few days ago, Substack showed me a post very, very specifically targeted to a specific phrase I used in a Google Meet with my therapist that morning. It was *very* specific. Not something I have ever searched, Googled, posted about, talked about with family. It was not on the map ANYWHERE outside of that conversation prior to that morning. It's not even adjacent to any of my usual interests in any way. At all. And I would testify to that in a court of law under oath and more importantly, I would remember if it had ever come up before that.
I'm really, really curious about how and when I EVER consented to Google listening to my therapy sessions much less sharing that data with any social media platform. And that was the second time something like that happened in two weeks.
The other was a very weird content crossover from YouTube to Substack but in a very, very indirect way. That I also never consented to.
I'd be terrified of this capability if it surprised me in any way. It's just the tip of the surveillance iceberg unfortunately. The worst part is that both times it was useful information but still intrusive in a very uncanny valley kind of way. I felt disgust like my skin was crawling and had a feeling of being violated in some inexpressible way.
What stood out to me is that this isn't just a conversation about privacy. It's a conversation about how we define meaningful consent.
There's an important difference between information being technically available and people realistically understanding what they've agreed to. As AI becomes more integrated into everyday products, designing for genuine understanding may become just as important as designing for compliance.
Sam, thank you for this. I went through each item. I am happy to report that I had already blocked tracking and training on all of my accounts. Frankly, I am very proud of myself! 😆 I thought surely I had missed something. The main point I'd like to make is the difference between privacy in Europe and America. It's night and day. Americans truly don't see and many don't even care about companies hijacking their entire lives. We should look to GDPR for oversight and regulations. The worst of it is in our public schools.
Well done Traci! And I agree entirely. I also think that even Europeans need to better understand why data sovereignty is important and the fact that if a platform is free, the cost is probably your data.
Thanks Sam! This is very useful!🩷🦩
You are very welcome Pinkie! 🙏
Useful guide, and also a fairly grim picture of modern consent.
Apparently saying no now requires spare time, patience and the instincts of someone escaping an online maze.
Exactly this, Daniel. The word Byzantine doesn't even come close to covering this.
And this is why I hate AI and these tech thieves! They are manufacturing our destruction to make more money for themselves.
Absolutely, Kathy, and making it so hard for us to regain our data for ourselves
Thank you! You can also opt out of substack using your writing to train AI. I’ve turned a lot of these off a while ago, but I’ll have to revisit just to double check.
Thanks Sundeep, that is really helpful and absolutely something I should have included at a meta level.
Great practical guide. The toggles are worth the fifteen minutes, but the part you're honest about at the end is the part I keep thinking about: the protection was never the setting, it was the price of reading you.
I published a piece yesterday on the same thread from a different angle: what happens when AI drops the cost of reading everyone to zero, age verification hands over the last identifier, and the log of how you think sits on a server with no legal privilege. The toggles protect tomorrow. The architecture is already built.
https://techtrenches.dev/p/the-cost-of-reading-everyone-just
Thanks, Denis. Looking forward to reading this.
Sam, I literally love your work, you and two other in particular just give me so much inspiration! The consent you have to chase through buried settings is exhaustion dressed up as choice" is the line means everything to me. I went and checked the Claude toggle right after reading — I had no Scooby-Doo that the policy flipped in August 2025, and given how much personal context I run through Claude day to day, that's not a small thing to have missed.
What strikes me reading this next to your sycophancy piece is that they're the same shape of problem from two different angles. Sycophancy is the AI quietly giving you less than it should. This is the AI quietly taking more than it should. Both rely on the same thing: you not noticing, because nothing announces itself. The fix in both cases is the same discipline — go and look, on purpose, at the thing that's designed to be skipped past.
Appreciate you doing the unglamorous work of listing the actual menu paths. I'd have given up after Google LOL!
You are so welcome Lee! And very happy to be doing the unglamorous work. 🙏
Haha
These are helpful to stanch the bleeding, but there's no recovery of blood already leeched.
We also can't call AI out for doing something nefariously novel: to some degree, this is behavior tracking, (now real time) personalization, aggregation of transactional, location, and social media activity - on steroids.
Still, if the adoption rate of such defenses is high, some of the data AIs have already collected will age out of relevancy. And maybe the desperation for current data drives change in the default behavior of data collectors.
Very true. Which is why we must be ever vigilant with the data we do decide to share. 🙏
I had already made alterations to most of these, but not to social media. Very helpful.
This makes me nervous: “Treat the toggle as a real step rather than a cast-iron guarantee, because how fully these opt-outs hold has been questioned.”
The other thing I do is delete the thread at the end of each day and start a new thread the next day (when it’s something important).
This seems to work because I definitely have to deal with the hassle of interacting with Claude or ChatGPT as if it is new to whatever I was talking with it about the day before.
But I’m wondering, based on your quote, if changing the settings and deleting the thread is still not enough?
I worry that it might not be Kerry. I have had instances of Claude chats bleeding over into other ones that I thought were in 'Incognito mode'. Sadly the truth is that we should start from a position of complete distrust with these companies.
Very handy info, thanks 👍🌻
You are very welcome Alice. 🙏