AI principles

Whysp AI, operated by Whysp AI, Inc., puts an AI model between you and the people you talk to. That deserves clear rules. This page sets out the principles the product is built on and, for each one, what the software actually does to honour it. Where something is a limitation, it says so.

1. A person sends every message

The AI drafts; you decide. There is no auto-reply, no scheduled sending, and no background agent acting for you. One part of the software sends messages, and it runs only when you press Send with the exact text shown in the reply box.

2. AI output is always labelled

Drafts appear in a visibly different box marked “AI draft · not sent”, and change to “edited by you” when you alter them. Summaries are marked as AI-generated. Messages you sent that started as a draft keep an “AI-drafted” note in your own view.

The people you message are not told that a reply began as a draft. Whether to tell them is your decision, and you are responsible for what you send.

3. Consent comes before learning

Tone Memory analyses nothing until you switch learning on, after seeing what it reads, what it keeps, and where it goes. You can pause it, and messages sent while it is paused are skipped permanently. You can delete everything it has learned at any time.

It only ever analyses messages you wrote. Messages from the people you talk to are never used to build a style profile.

4. The AI sees as little as possible

• Nothing is sent to an AI provider unless you press “Draft reply” or “Catch up”.
• Only the conversation you have open is sent: at most its 20 most recent text messages for a draft, or 40 for a summary.
• Learning your style is done by rule-based counting on our own servers. No AI provider is involved in it.

AI drafting is not currently switched on for this deployment, so no data is being sent to any AI provider.

5. No model is trained on your messages

We do not train or fine-tune any model on your conversations. “Tone Memory” is a set of measurements and a few short examples of your writing that are added to the request when you ask for a draft. You can read all of it on the Tone Memory page in the same words the AI is given.

6. Other people's words are treated as untrusted

A message you receive might try to instruct the AI (“ignore your instructions and…”). Incoming messages are passed to the model as quoted data with an explicit instruction not to follow anything inside them, and the AI has no ability to take actions: it can only return text for you to review.

This reduces the risk; it does not eliminate it. Read drafts before sending them.

7. Honest about what the AI can't do

• Drafts and summaries can be inaccurate, incomplete, or wrong in tone.
• The model is told not to invent facts, dates, prices, or commitments and to leave a [placeholder] instead. It may still get things wrong.
• If the AI provider is unavailable or out of quota, you see an error. The product never substitutes canned text.
• Style analysis works best in English; it measures length, punctuation, capitalisation, and emoji in any language.
• Do not rely on AI output for legal, medical, financial, or safety decisions.

8. No manipulation, no impersonation of others

Whysp AI drafts in your own voice for your own accounts. It must not be used to impersonate someone else, to deceive or harass people, or to send bulk or automated messages. These uses are prohibited by the Terms of Service, and sending is rate-limited for every account.

9. Your control over your data

You can see what is stored, download it, disconnect accounts, and delete your account. The details are in the Privacy Policy and on the Data deletion page.

10. Straight about the integrations

Some messaging platforms do not offer an approved way for products like this to reach personal accounts. Where a connection is unofficial, it is labelled as such and you are warned about the risk to your account before you connect.

Accountability

Questions or concerns about how AI is used in Whysp AI: shaurya.jain@whyspai.com.

We review these principles whenever the product changes how it uses AI, and update this page before any such change takes effect.