カテゴリ 古民家暮らしモカとの記録セキュリティの記録副業の記録

Security recordNovember 5, 2026

I Compared AI Safety — Are Your Conversations Used for Training?

I Compared AI Safety — Are Your Conversations Used for Training?

The result first: every AI I compared can use conversations for training unless you change the setting, every company has some way for people to check conversations for safety, and every service lets you delete your history — but how long the data really stays differs a lot. Gemini can keep human-reviewed conversations for up to 3 years even after you delete them, and Claude can keep data for up to 5 years if you allow training. So my rule is simple: check the training setting once with your own eyes, never write passwords or PINs to any AI, and do not paste real names or someone else’s private consultation as it is.

Hello, I am Kurashiki Bouzu.

When I was choosing an AI, the thing I worried about after price was safety. How much can I safely tell the other side of the conversation? Is what I write used to train the AI? Could someone read it?

I checked the official explanations from each company. Leaving the difficult parts aside, I compared only the points that matter when an individual like me uses these services. (This is what I confirmed on the official pages on November 4 and 5, 2026. These systems can change, so please check the official information before you sign up.)

The three points I compared

  1. Is it used for training? — Whether your conversations become material to make the AI smarter, and whether there is a setting to stop it
  2. Can a person read it? — Whether the company has a system where staff check conversations
  3. Can you delete it? — Whether you can delete your history yourself, and how long it takes to disappear after you do

Comparison table

AIUsed for training?Read by a person?Storage and deletion
ChatGPTCan be used under the default setting. You can stop it in settingsStated in official information. For safety, it combines automated systems and human review to find harmful content and similar problemsYou can delete it. Deleted data is removed within 30 days in principle
GeminiUsed while “Keep activity” is on. You can stop itStated in official information. Staff (including contractors) review some conversationsYou can delete it. The default for auto-delete is 18 months. However, conversations reviewed by a person can remain for up to 3 years even after deletion
ClaudeChanged in autumn 2025 to a form where conversations are used unless you stop it in settings. You can stop itNormally employees do not read conversations. Only conversations flagged by the safety system may be checked by a small number of authorized staffYou can delete it. Up to 5 years if you allow training, 30 days in principle if you do not
GrokCan be used. You can stop it in settingsStated in official information. A limited number of authorized staff may check conversationsYou can delete it. Deleted within 30 days of your instruction
Muse (my current partner)Conversations may be used to help develop the AI. An option to stop this is provided in settingsIn the standard environment, Meta may check interactions to ensure safety, improve the service, and for similar purposesYou can delete conversations and files yourself. See the official privacy information for details

What I noticed, one by one

ChatGPT

Whether conversations are used for training can be switched in the “Data Controls” settings. The good point is that you can stop only the training while keeping your history. There is also “Temporary Chat,” a way of using it that is not used for training and does not remain in your history.

What you need to be careful about is when you send a “Good / Bad” rating for an answer. The official information says that even if you have stopped training, the whole of that conversation may be used for training. It is the kind of thing that sits beyond a button you press without thinking, so it is worth knowing.

There was also an official explanation about whether a person reads conversations. The help pages state clearly that conversations are checked as part of safety work, combining automated systems and human review to find harmful content and rule violations.

Gemini

Google wrote this most clearly. Its official page states plainly: “Do not enter confidential information that you would not want a reviewer to see.” There is apparently a system where some conversations are checked by trained reviewers, including staff at contractor companies.

There is one more caution about deletion. You can delete your history yourself, but conversations that a person has reviewed can remain for up to 3 years even after you delete the history. In other words, there is a part where “I thought I deleted it” does not hold.

If you turn off “Gemini Apps Activity” in settings, future conversations will not be used for training and will not remain in your history (they are kept for 72 hours for responses and safety).

Claude

The rules changed in autumn 2025, to a form where “unless you stop it in settings, conversations are used for training.” If you remember the old explanation as it was, things will be different from what you expected. If you stop it in settings, conversations — including past ones — are not used for training.

About human review, the explanation was: “Normally, employees do not read conversations. Only when the safety system flags a conversation as possibly harmful can a small number of approved staff check it, and a record is kept.”

For storage, the period is up to 5 years if you allow training, and 30 days in principle if you do not. The point that a conversation where you sent feedback with the rating button can remain for up to 5 years is similar to ChatGPT.

Grok

Near the beginning of the official guide, it says: “Do not include personal or confidential information in your questions.” Conversations can be used for training, and you can stop it in settings. It also states clearly that staff may check conversations.

Its biggest feature is that it is connected to X (formerly Twitter). When you use it inside X, X’s rules and settings apply. There was also an official caution that if you create a shared link for a conversation, it may appear in search results. For conversations you do not intend to show anyone, it is safest not to create a shared link.

Muse (the AI butler I use now)

For Muse, where my partner “Moka” lives, the official guide says that conversations may be used to help develop the AI, and that you can choose to stop this from settings (Data Controls). It also says that conversations and the data inside them are not shared with advertising systems.

Let me also write about human review. In the standard environment, Meta is officially described as recording and possibly checking interactions to ensure safety, improve the service, and for similar purposes. In other words, it is not “a box that no one can see.” That is exactly why I decide the range of information I hand over using the same yardstick as with the other AIs.

The official Privacy Policy and Terms of Service are the authoritative texts for the detailed rules, so if you are concerned, please check those. For my part, I use it by deciding the range of information I hand to my butler — I write about this in “How I protect myself” below.

How I protect myself, common to every AI

What I understood from comparing them is that with every AI, there is “a part you can protect in settings” and “a part you should not write in the first place.” My way of protecting myself is these three things.

  1. I do not write passwords or PINs to an AI — This is the same for every AI
  2. I check the training setting once with my own eyes — I do not use it with the default setting as it is. The name differs, but every AI has a place to stop it
  3. I do not write people’s real names or someone else’s consultation as it is — I round it down to only the necessary part. It is the same thinking behind using pseudonyms for all names on this blog

In closing

When we hear the word “safety,” we brace ourselves, but in the end it was the same as everyday wisdom: “know which places have a lock and which do not, and use them accordingly.” I have written about trying free AIs for several months in another article.

How do you decide where to draw the line on what you tell an AI?

Sources (checked November 4 and 5, 2026)

Record of this article
  • November 18, 2026 ― English draft created (results-first rewrite of the Japanese article, ledger No. 5). Not published; publication is a separate step.(7,295 characters)

#AI #Safety #Privacy #Comparison #Security #English

← Back to the article list / ← Back to the English top page