Should magicians tell AI their secrets?
The short answer: not the methods. Here's what the big AI tools do with what you type, where the real risk sits, and how to use them without handing over your material.
Table of contents
AI is good at a lot of what a working magician needs: patter drafts, show orders, contracts, and emails to bookers. It's also the first notebook in history that may learn from what you write in it. For most people that's a privacy footnote. For a craft built on secrets, it's the whole question.
Why AI is bad at magic, for now.
Magic methods aren't generally available to the public. They live in books, lectures, and conversations between magicians. Large language models like ChatGPT, Gemini, and Claude learn from what they can read, so they have very little to pull from. Ask one how a marketed effect works and you mostly get confident guesses.
That gap is the magician's code, working at scale.
What happens to what you type.
Each of the big AI tools publishes its own rules. As of October 2026:
- ChatGPT. OpenAI says ChatGPT improves by training on the conversations people have with it unless you opt out, using a setting called Improve the model for everyone. Temporary Chats are not used for training. OpenAI's data policy.
- Gemini. Google says your conversations help improve its services, including its machine-learning models, and that human reviewers read some of them. Reviewed chats are kept for up to 3 years, even if you delete your activity. Google itself tells you not to enter confidential information. Gemini Apps Privacy Hub.
- Claude. Anthropic asks you to choose. If you allow it, your chats can be used to improve Claude and are kept for up to 5 years. If you don't, they're kept for 30 days. Anthropic's consumer terms update.
- Siri. Apple says your audio is not stored unless you opt in to Improve Siri and Dictation. Your request history, with transcripts, may be kept for up to 2 years, and transcripts may be used to fine-tune its models. Apple may review a subset of them. Apple's Siri and Dictation privacy page.
- DeepSeek. DeepSeek says it uses your data to train and improve its models, and keeps it for as long as you have an account. Its policy gives you a right to opt out of training but names no setting. DeepSeek's privacy policy.
- Grok. SpaceXAI says it may use your prompts and Grok's replies to train its models, and you control that with the Improve the model setting. Private Chat is not used for training. A limited number of its people may review conversations. SpaceXAI's consumer FAQ.
- Perplexity. Perplexity says it uses data to improve its products, including its AI models, and keeps personal data only as long as it needs to. It names no training setting in its notice. Perplexity's privacy notice.
- Mistral. On Mistral Vibe, your input and output are used to train its models by default unless you opt out, and a thumbs up or down with a comment is always used. Enterprise customers are opted out by default. Mistral's help center.
OpenAI and Anthropic both say their business and API products aren't used for training by default. The free and personal plans, the ones most magicians use, are the ones that can learn from you.
As of October 2026
Trains on your chats by default
Yes
Unless you opt out.
The setting that turns it off
Improve the model for everyone
Temporary Chats are not used for training.
Human review
Not stated here
Retention
Not stated here
Trains on your chats by default
Yes
Conversations help improve its services, including its machine-learning models.
The setting that turns it off
Not stated here
Human review
Yes
Human reviewers read some conversations.
Retention
Up to 3 years
For reviewed chats, even if you delete your activity.
Trains on your chats by default
Only if you allow it
Anthropic asks you to choose.
The setting that turns it off
Your choice
Say no and chats are kept for 30 days.
Human review
Not stated here
Retention
Up to 5 years
If you allow training.
Trains on your chats by default
Transcripts may be
Apple says transcripts may be used to fine-tune its models.
The setting that turns it off
Improve Siri and Dictation
Unless you opt in, your audio is not stored.
Human review
Some
Apple may review a subset of transcripts.
Retention
Up to 2 years
Request history, and longer for reviewed transcripts.
Trains on your chats by default
Yes
DeepSeek says it uses your data to train and improve its models.
The setting that turns it off
Not named
The policy gives a right to opt out but names no setting.
Human review
Not stated here
Retention
While you have an account
And longer for a legitimate business interest.
Trains on your chats by default
May
SpaceXAI says it may train on your prompts and Grok's replies.
The setting that turns it off
Improve the model
Private Chat is not used for training.
Human review
Some
A limited number of authorized personnel may review conversations.
Retention
Until you delete
Deleted data takes up to 30 days to clear.
Trains on your chats by default
May
Perplexity says it uses data to improve its AI models.
The setting that turns it off
Not stated here
Human review
Not stated here
Retention
As long as needed
No fixed period is given.
Trains on your chats by default
Yes
Input and output are used by default unless you opt out.
The setting that turns it off
Opt out
Thumbs up or down feedback is always used.
Human review
Not stated here
Retention
Not stated here
Google's own page tells you not to enter confidential information.
Not stated here means this guide has no answer from them on it. That is not the same as a no.
Settings change, and training isn't the only leak.
Those are today's settings. Policies change, defaults change, and a switch you set once doesn't follow you to a new account or a new app. Anthropic's current rule arrived in 2025 as an update to its terms.
Training also isn't the only way a chat gets out. In July 2025, thousands of shared ChatGPT conversations turned up in Google search results, because a sharing option made them discoverable. OpenAI called it a short-lived experiment and removed it. Search Engine Journal has the details.
And what a model learns, it may repeat to whoever asks next.
A working rule.
Tell AI about the show. Never about the method.
- Patter and scripting, described by what the audience sees.
- Show orders, pacing, and set lists.
- Marketing copy, booking emails, contracts, and invoices.
- Learning theory: practice schedules, memory, and stage fright.
Keep out how an effect is done, how a prop is built, and anything from a book, a lecture, or a download you paid for. That last one is someone else's secret as well as yours.
The test is simple. If you wouldn't post it on a public forum, don't type it into a chatbot.
The test
If you wouldn't post it on a public forum, don't type it into a chatbot.
Tell AI about the show.
- Patter and scripting, described by what the audience sees.
- Show orders, pacing, and set lists.
- Marketing copy, booking emails, contracts, and invoices.
- Learning theory: practice schedules, memory, and stage fright.
Never about the method.
- How an effect is done.
- How a prop is built.
- Anything from a book, a lecture, or a download you paid for.
That last one is someone else's secret as well as yours.
Where your secrets should live.
Methods, handling notes, and rehearsal footage belong somewhere that can't read them. Paper works. So does software that encrypts on your device before anything is uploaded, so the company running it has nothing to read, nothing to train on, and nothing to leak.
That's how Trickifi is built. Your repertoire is end-to-end encrypted, so no one, including Trickifi staff, can access it, and none of it is used to train AI. See how Trickifi keeps accounts secure.
Where a note can sit
Methods, handling notes, and rehearsal footage belong somewhere that can't read them.
A chat window
What you write
- Methods
- Handling notes
- Rehearsal footage
Your own words
What the server holds
- Methods
- Handling notes
- Rehearsal footage
The same words
It arrives readable. It may be used to improve a model, and some conversations are read by people.
An encrypted vault
What you write
- Methods
- Handling notes
- Rehearsal footage
Your own words
What the server holds
Ciphertext
It is encrypted before anything is uploaded, so the company running it has nothing to read, nothing to train on, and nothing to leak.
Paper works for the same reason. It has never read anything either.