AI Disclosure

The short version: every dictionary entry on Witprinted is written by an AI model. We use a locked prompt, screen every output through a multi-layer safety pipeline, and don't train any AI on your inputs. The entry is intentionally absurd; treat it as comedy, not factual biography.

1. The content is AI-generated

When you type a name and vibe, the dictionary entry is produced by Anthropic's Claude Haiku 4.5 model in response to a fixed system prompt of ours. No human is writing the entry. Every output is screened by automated safety checks before it can be ordered. For the first several hundred orders, every order is also reviewed by a human within 24 hours before being sent to production; after that, an ongoing 5% sample is human-reviewed. Outside that review window, the system runs end-to-end automatically.

2. The prompt is locked

We use a single, version-tracked system prompt that instructs the model to:

• Adopt a dry, sarcastic, encyclopedic tone.
• Generate a fake dictionary entry: headword, pronunciation, part of speech, three definitions, see-also references, and a fake etymology.
• Never reference race, ethnicity, religion, disability, sexual orientation, gender identity, mental health, weight, age, political affiliation, or national origin.
• Stay within "allowed targets" of comedic critique: caffeine habits, procrastination, romantic chaos, work avoidance, social-media habits, hobbies — never the person's identity attributes.
• Treat non-Western names with respect; never mock pronunciation.

You don't pick the prompt. We pick the prompt. The vibe selector you choose just adds a small flavour modifier on top.

3. We don't train on your inputs

• We have zero-retention enabled with Anthropic. Anthropic does not retain your inputs after producing the response and does not use them for model training.
• We do not train models on your name, your vibe selection, your order, or your email.
• We do retain the generated output for 30 days post-delivery (for support and refund handling) and order metadata for 7 years (for tax compliance), per our Privacy Policy.
• For Phase 2 image-based products, the same applies to image AI: zero-retention with the image-generation provider, photo deleted within 24 hours of fulfillment, and your photo never used to train any model.

4. Multi-layer safety screening

Every generation passes through up to five sequential checks before it can become an order:

1. Profanity and slur blocklist on the input.
2. Public-figure name blocklist (politicians, celebrities, athletes, trademarked fictional characters).
3. Generation with the locked, constrained system prompt described above.
4. Post-generation review by a second AI call that scores the output for offence risk; outputs above a threshold are auto-regenerated, and outputs at the highest threshold are hard-blocked and logged.
5. Human review on the first 500 orders and an ongoing 5% sample, with a 24-hour SLA before the order is sent to production.

Hard-blocked attempts are logged with a hashed identifier (no PII) for audit. We don't share these logs with third parties.

5. AI is unpredictable

By definition, generative AI produces different outputs each time. We give you two free regenerations before checkout so you can pick an output you like. Once you pay and the design enters production, the output is final and is what gets printed.

We are not responsible for an output's specific wording when it falls within the bounds set by our prompt and safety pipeline. If you receive your mug and decide on reflection that you don't like the wording you approved at checkout, that is not grounds for a refund — see Exchange Policy.

6. AI cannot replace your judgment

The entry is comedy. It is not biography, profile, or fact. It is not advice. It is intentionally absurd. Don't rely on a Witprinted entry to describe a real person to anyone for any serious purpose.

7. Public figures and celebrities

We do not knowingly generate entries for the names of real public figures. If you order an entry for a name that matches our public-figure blocklist, the order will be refused and refunded. If a name slips through that you believe is a recognisable real public figure, contact us immediately and we will remove it.

8. Phase 2: cartoon portraits from photos

When we launch image-based products in Phase 2, the cartoon portrait will also be AI-generated using a separate locked art-style prompt. We will run every uploaded photo through a celebrity-detection check before generating any image and refuse uploads that match a known public figure. We do not allow customer-supplied prompt overrides on the art style. Phase 2 will have its own dedicated disclosure update before any image product ships.

9. Your right to know what model produced your entry

The model name and version that produced your entry is recorded in our database for the duration of the retention window. You can request this information by emailing hello@witprinted.com with your order ID.

10. Contact

Sawary Trading LLC (operating Witprinted)
Wyoming, United States
hello@witprinted.com