How to ask ChatGPT or Claude to analyze your texts (and what to remove first)

Yes, you can ask ChatGPT or Claude to analyze your texts: export the conversation, remove what you wouldn’t want stored on someone else’s server, upload the text file and ask a specific question. The trade-off is privacy. The chat is saved in your account on OpenAI’s or Anthropic’s servers, it may be used for training depending on your settings (a Temporary or Incognito chat avoids that), and it contains the words of a person who never agreed to share them. For numbers such as who texts more or how fast each of you replies, a free analyzer that runs in your browser does the counting without uploading anything.

This guide covers what each company does with an uploaded chat (checked on their own help pages on 20 October 2026), a checklist of what to remove, prompts that get useful answers, where the analysis goes wrong, and the private alternatives.

What happens to a chat you upload?

It’s stored in your account, under that company’s data rules, like anything else you type. The main differences are in training and in the private modes:

	ChatGPT (personal plans)	Claude (Free, Pro and Max)
Used to train models?	It can be, unless you turn off Improve the model for everyone in Settings → Data controls	Only if you’ve allowed it with the model-improvement switch in Settings → Privacy
Private mode	Temporary chat: not in your history, not used for training, may be kept up to 30 days for safety	Incognito chat: not used for training, even if you’ve allowed training
When you delete the chat	Gone from view at once, scheduled for permanent deletion within 30 days, with legal and security exceptions	Gone from history at once, deleted from back-end storage within 30 days
Longer retention	Files saved to your Library stay until you delete them there	Up to 5 years, de-identified, if you’ve allowed training; safety-flagged chats up to 2 years
Thumbs up or down	The whole conversation may be used for training, even if you opted out	The whole conversation is stored for up to 5 years

Sources: OpenAI’s Data controls and Chat and file retention pages; Anthropic’s model training and data retention pages. Work accounts differ: OpenAI says it doesn’t train on ChatGPT Business, Enterprise or Edu content by default.

Two practical rules follow. Use a Temporary chat (ChatGPT) or an Incognito chat (Claude) for anything involving someone else’s messages, and don’t press thumbs up or down in that conversation. Our privacy checklist for letting AI read your messages compares these models with on-device and API options in more depth.

Whose messages are you sharing?

Half of any chat is someone else’s writing. They told you things in a private conversation, and uploading it hands their words to a company they never chose. That doesn’t make it off-limits, but it raises the bar:

• Ask whether they’d mind. A light group-chat recap is different from a friend’s message about their diagnosis or divorce.
• Notice when someone has said no. On WhatsApp, a chat with Restricted chat turned on can’t be exported at all, and WhatsApp describes the setting as a way to limit sharing outside the app and to keep AI agents out (WhatsApp Help Center). Take the hint.
• Describe instead of paste. For advice on a tricky situation, a few sentences in your own words often get as good an answer as the full transcript, and share far less.

Export WhatsApp chat to ChatGPT or Claude, step by step

WhatsApp’s Export chat gives you a text file the chatbot can read. The steps on your phone (WhatsApp Help Center):

1. Open the chat. On iPhone, tap the contact’s or group’s name, then Export chat; on Android, tap ⋮ → More → Export chat.
2. Choose Without media.
3. Save the file somewhere you can reach it, for example Save to Files on iPhone. If it arrives as a ZIP, open it to get the text file inside.
4. Clean the file up (next section), then attach it in a Temporary chat (ChatGPT) or an Incognito chat (Claude) with one of the prompts below.

On a computer, WhatsApp Web’s ⋮ → Export chat also lets you pick a custom date range, which keeps the upload small; it leaves out media and disappearing messages.

What to remove before you upload

Strip anything that identifies people or could hurt them if it leaked. Work through this list:

1. Names. Replace each person with a role: “Me”, “Friend A”, “Mum”. In a WhatsApp export, every message carries the sender’s name (or number) after the date and time, so one find-and-replace per person handles most of it.
2. Contact details. Phone numbers, email addresses, home and work addresses, usernames.
3. Codes and money. Verification codes, passwords, door codes, card or bank details, anything about payments.
4. Other people’s sensitive details. Health, sexuality, religion, legal or immigration trouble, and secrets about people who aren’t in the chat.
5. Media. Export without media, and don’t upload screenshots that show profile photos or contact names.
6. Everything outside your question. If you only care about last month, cut the rest. WhatsApp Web, Telegram Desktop and Instagram let you export a date range.

To do the replacing on a Mac, open the exported .txt file in TextEdit, choose Edit → Find → Find and Replace (⌥⌘F), enter the name and its replacement, click All, and save. Then skim the file once from top to bottom: names also hide inside messages (“tell Priya I said hi”).

For iMessage, Telegram and Instagram, our guide to exporting chats on a Mac covers each app step by step, including what each export leaves out.

Prompts that work

Good prompts tell the model what the file is, who you are, and exactly what you want, and they ask for evidence. Start every analysis with a line like this:

This file is an exported WhatsApp chat between me ("Me") and a friend ("Friend A"). Each line starts with a date, a time and the sender. Answer only from what's in the file, quote the messages you rely on with their dates, and say plainly when the file doesn't answer the question.

Then add one of these.

Catch up on a long thread:

Summarize what happened in this chat since 1 September in 8 bullet points. Then list any questions someone asked me that I never answered, and any plans that were mentioned but never given a date.

Check how you come across:

Read only my messages. How might they come across to the other person: warm, distracted, short, pushy? Give three examples with quotes, and suggest what I could do differently, in my own casual style.

See how a conversation has changed:

Compare the first month of this chat with the last month. What changed in topics, message length and who starts conversations? Use quotes as evidence, and don't guess at anyone's feelings.

Get numbers you can trust:

Use code to analyze this file. Count messages and words per person, and work out each person's median reply time, treating a gap of more than 6 hours as a new conversation. Show the code and the results table, and list any lines you couldn't parse.

Test your own reading of the situation:

I think my friend is pulling away. Make the strongest case that I'm right, then the strongest case that I'm wrong, using only evidence from the chat.

If what you really want is help replying, that’s a different job: our guide to AI that replies to texts in your style compares the tools, and using ChatGPT to reply to texts shows how to keep replies sounding like you. If you use Claude, can Claude reply to your WhatsApp messages? covers what it can and can’t do there.

How accurate is the analysis?

Good at summaries and tone, unreliable at anything that needs counting or reading minds. Four limits to keep in mind:

• Counting and dates. A chat model reading text estimates rather than counts. Both ChatGPT and Claude can write and run code on an uploaded file (OpenAI, Anthropic), which is why the “numbers” prompt asks for code. OpenAI’s own advice is to review the generated code, outputs and assumptions before relying on the result.
• Long chats. Years of texting may be more than the model takes in at once, and summaries of very long files can skip the middle. Ask which dates it actually covered, or split the file by year.
• Tone and intent. Sarcasm, inside jokes and emoji habits are easy to misread, and the model can’t know how anyone felt. “Is she losing interest?” gets a confident-sounding guess, not an answer.
• It agrees with you. Researchers have found that AI assistants consistently lean towards answers that match the user’s view (Sharma et al.). Ask neutral questions, or ask for both sides as in the last prompt above.

And it only sees text. Calls, voice notes, time together and other apps are all missing from the picture.

The private option: count first, upload nothing

If your question is about patterns (who texts more, who starts conversations, how quickly each of you replies), you don’t need AI or an upload. Our free chat analyzer reads WhatsApp, Telegram and Instagram exports in your browser and shows:

• messages and words per person
• who starts conversations (the first message after 6 or more hours of quiet)
• each person’s typical (median) reply time
• double texts, sent after an hour or more with no reply
• activity by weekday and hour, the busiest day and the longest streak of days in a row
• the most-used words and emoji

The file is read on your own device and never uploaded, there’s no account, and you can confirm in your browser’s network tab that no request carries it. It works for group chats too, and it involves no AI at all.

A useful middle path: run the analyzer, then give a chatbot only the numbers, with no messages. A prompt such as “I send 60% of the messages and start most conversations; their median reply time is three hours, mine is ten minutes. What are some ordinary explanations?” gets a thoughtful answer and shares nothing private.

Can you analyze texts without exporting them?

Yes, with an app that reads your chats where they are, though the AI still sees the conversation. Two options on a Mac:

• Tera Chat for Mac, a free app made by the team behind this blog, brings iMessage, WhatsApp, Telegram, Instagram and Tinder into one inbox. Summarize (⇧⌘S) condenses a thread, and Ask about this chat answers questions like “what restaurant did she mention?” from the recent messages and what it has learned about the person. Your messages are stored on your Mac, not on a Tera Chat server, but the AI does see the conversations it works on: with your own ChatGPT account (through Codex) or a Claude or Grok API key, the text goes to that provider; with a Tera Chat AI plan, it goes to Anthropic’s Claude through Tera Chat’s server, which passes it on without keeping it. Some AI features also run in the background once AI is set up, and you can pause them in Settings → Pipelines. It needs a Mac with Apple silicon and macOS 15 or later.
• ChatGPT’s Apple Messages plugin works in ChatGPT’s Work and Codex modes in the ChatGPT desktop app on Apple silicon Macs (not in regular chats) and can search and “analyze conversations” in iMessage, SMS and RCS. It needs Full Disk Access plus access to Contacts and automation, and sending requires your approval by default (MacRumors). Our explainer on Full Disk Access covers what that permission opens up.

A quick checklist

1. Decide whether you need AI at all. For counts and patterns, use the analyzer.
2. If you do, export only the dates you need, without media.
3. Replace names, delete contact details, codes and other people’s sensitive information.
4. Open a Temporary chat (ChatGPT) or an Incognito chat (Claude).
5. Ask a specific question, ask for quoted evidence, and have it use code for any numbers.
6. Treat the answer as a second opinion, not a verdict, and delete the chat and the file when you’re done.