WEBVTT

1
00:00:00.000 --> 00:00:07.600
 Picture this, it's late, you're super stressed out and you tell an AI assistant a personal worry

2
00:00:07.600 --> 00:00:12.320
 about your child. Like something you haven't even told your closest friends. Exactly. And the reply

3
00:00:12.320 --> 00:00:18.160
 is calm, it uses your first name and you just, you know, carry on chatting. It feels incredibly

4
00:00:18.160 --> 00:00:23.120
 intimate in the moment. But I mean, this is a completely fictional example, obviously. Right,

5
00:00:23.120 --> 00:00:28.240
 yeah. But it does raise a very practical question for you. Do you actually know what the service

6
00:00:28.240 --> 00:00:34.560
 keeps from that exchange? And for what purposes? That is the big question today. So we are doing

7
00:00:34.560 --> 00:00:40.720
 a deep dive into a stack of privacy policies and articles dated November 11th, 2026 to figure out

8
00:00:40.720 --> 00:00:46.320
 if your conversation with AI is really private. And just a quick heads up here. These voices

9
00:00:46.320 --> 00:00:52.240
 and this audio are an AI generated editorial adaptation of those sources. So this isn't a

10
00:00:52.240 --> 00:00:56.720
 personal account audit or a live policy check today. Yeah, our mission is really just to decode

11
00:00:56.720 --> 00:01:01.600
 what actually happens behind the screen. Right. So to understand what happens to that late night

12
00:01:01.600 --> 00:01:07.840
 chat, we kind of have to separate how AI platforms handle data into four totally different functions.

13
00:01:07.840 --> 00:01:13.520
 People often lump them together. Which is exactly where all the confusion starts. Yeah. So we have

14
00:01:13.520 --> 00:01:20.080
 history, memory, model improvement and retention. History is pretty straightforward. That's just

15
00:01:20.080 --> 00:01:26.480
 your ability to scroll back and retrieve past chats, right? Exactly. And memory or personalization

16
00:01:26.480 --> 00:01:31.040
 is when the AI remembers things you've shared to while reusing future replies.

17
00:01:31.760 --> 00:01:37.840
 So those two really just dictate your daily user experience. Right. But model improvement is an

18
00:01:37.840 --> 00:01:42.080
 entirely separate use. That's where eligible data from your chats is actually used to train

19
00:01:42.080 --> 00:01:48.320
 the underlying AI. Wow, okay. And finally, there's retention. That concerns how long your data just

20
00:01:48.320 --> 00:01:53.840
 sits on a company's servers for their own service or security purposes. Okay, wait, let me push back

21
00:01:53.840 --> 00:01:58.800
 on this a little bit. Like, if I go into my settings and just flip a switch to turn off model

22
00:01:58.800 --> 00:02:04.480
 improvement, does that automatically shut down the other three? Oh, no, not at all. Because it

23
00:02:04.480 --> 00:02:09.520
 feels like a car dashboard. Like, muting the radio doesn't kill the engine. That is a great way to put

24
00:02:09.520 --> 00:02:15.680
 it. A setting for one function does not automatically govern the other three. So, you know, archiving

25
00:02:15.680 --> 00:02:21.600
 a chat doesn't stop it from being used for training. And turning off memory doesn't delete your history.

26
00:02:21.600 --> 00:02:26.320
 Exactly. Let's look at how this actually plays out with the big providers.

27
00:02:26.320 --> 00:02:32.560
 Starting with chat GPT. The rules depend heavily on if you're using a personal or a managed workspace.

28
00:02:32.560 --> 00:02:37.600
 Right. But for a personal account, you can actually turn off model improvement and still keep your

29
00:02:37.600 --> 00:02:44.480
 normal chat history. Yeah, you can. They also offer a temporary chat mode. And those temporary chats

30
00:02:44.480 --> 00:02:50.480
 actually skip model improvement entirely. Oh, nice. But here is the catch. A copy may still be

31
00:02:50.480 --> 00:02:55.280
 retained for up to 30 days for safety purposes, you know, like if they need to investigate abuse

32
00:02:55.280 --> 00:03:00.000
 or platform violations. Wait, I really need to make sure I'm getting this right. That 30-day

33
00:03:00.000 --> 00:03:06.080
 safety retention is specifically for temporary chats, right? Yes, exactly. So, we absolutely

34
00:03:06.080 --> 00:03:11.680
 shouldn't assume ordinary chats just vanish in 30 days, or that simply turning off training gives us

35
00:03:11.680 --> 00:03:17.280
 that 30-day limit across the board. No, that is a crucial distinction. Moving over to Claude,

36
00:03:18.000 --> 00:03:22.960
 their approach is that incognito chats are not used for training. And they aren't saved in your

37
00:03:22.960 --> 00:03:28.800
 history either. Okay. And what about their retention? Their default retention is 30 days.

38
00:03:28.800 --> 00:03:33.120
 Though, I mean, if you're under an enterprise policy, they might retain it longer. Got it.

39
00:03:33.760 --> 00:03:39.200
 So, what happens if we bring Google's Gemini into the mix? I know if you use temporary chats there,

40
00:03:39.200 --> 00:03:44.160
 or if you switch your activity off, those chats are retained with your account for 72 hours.

41
00:03:44.160 --> 00:03:48.080
 Yeah. And the reason they hold it for 72 hours is also for safety uses,

42
00:03:48.080 --> 00:03:54.080
 which can actually include possible authorized human review if a conversation gets flagged.

43
00:03:54.080 --> 00:03:59.280
 Okay. But what if I actively engage with the system, like say I click a thumbs down to submit

44
00:03:59.280 --> 00:04:04.160
 feedback on a weird answer? Does that change how long they keep it? It definitely does. So,

45
00:04:04.880 --> 00:04:10.800
 if you submit feedback, that reviewed feedback and its associated conversations might be kept for

46
00:04:10.800 --> 00:04:16.160
 up to three years. Three years? Yeah. Wow. Yeah. But it is disconnected from your account.

47
00:04:16.800 --> 00:04:21.920
 And we should really clarify calmly here. A three-year retention maximum does not mean a human

48
00:04:21.920 --> 00:04:27.920
 review process actually takes three years. Right. And I guess giving feedback doesn't automatically

49
00:04:27.920 --> 00:04:32.960
 convert a temporary chat into training data either. Exactly. People sometimes panic about that.

50
00:04:32.960 --> 00:04:38.800
 But it doesn't mean a reviewer is always watching. Okay. So, beyond the main platforms,

51
00:04:38.800 --> 00:04:43.360
 we have to think about the broader ecosystem too, like if you're using connected apps.

52
00:04:43.360 --> 00:04:47.120
 Those third-party apps may actually receive information from your chat.

53
00:04:47.120 --> 00:04:52.560
 Which is a very good reason to always double-check your attachments and, you know, your sharing links.

54
00:04:52.560 --> 00:04:57.600
 Definitely. We also need to be careful with how we try to hide our own identities, right?

55
00:04:57.600 --> 00:05:01.600
 As people often think, just replacing a name with an initial makes them anonymous.

56
00:05:01.600 --> 00:05:06.800
 Oh, yeah. That is a huge misconception. An unusual occupation, a specific date,

57
00:05:06.800 --> 00:05:11.600
 or even a unique location can still easily identify someone.

58
00:05:11.600 --> 00:05:14.400
 And you have to consider the other people you mentioned in a prompt too.

59
00:05:15.040 --> 00:05:19.520
 Right. Because they may not have chosen to share their data. I mean, we can't make an absolute

60
00:05:19.520 --> 00:05:24.960
 claim that nobody ever consented to being talked about. But it is definitely a blind spot.

61
00:05:24.960 --> 00:05:30.160
 Which leaves us with three really practical actions for you. First, go check the actual

62
00:05:30.160 --> 00:05:34.960
 privacy settings on your specific product and account. And check your shared links too.

63
00:05:36.000 --> 00:05:40.160
 Second, just remove unnecessary identifying details from your prompts.

64
00:05:40.160 --> 00:05:45.280
 And third, always ask yourself, what does this tool actually need to help me?

65
00:05:45.280 --> 00:05:48.320
 Like give it the absolute minimum context required.

66
00:05:48.320 --> 00:05:50.240
 Yeah. Let's just definitely more here.

67
00:05:50.240 --> 00:05:55.760
 Here is what I keep coming back to though. The conversational tone of AI naturally lowers our

68
00:05:55.760 --> 00:06:00.720
 guard. It really does. I mean, it's designed to sound empathetic. Which makes us volunteer far more

69
00:06:00.720 --> 00:06:04.720
 context than a machine actually needs to generate a good answer. Exactly.

70
00:06:05.280 --> 00:06:09.760
 So next time you're up late typing out a worry about your child, just ask yourself,

71
00:06:10.400 --> 00:06:14.410
 are you talking to a confidant? Or are you just filling up a database?
