Effective date: 6 June 2026 Last updated: 8 October 2026
This page explains where Overthink This uses AI, what is sent to the AI providers, what safety measures are in place, and how you can report an answer. It is also our disclosure under Google Play's AI-Generated Content policy.
1. Where AI is used in the App
| Surface | Model lane | Purpose |
|---|---|---|
| Spiral generation (verdict + outcomes + action steps + in/out-of-control) | Open models served by Groq first (Google Gemini Flash models instead, while we have Gemini switched on); if that is unavailable, a model hosted by NVIDIA; if that is unavailable too, an open model hosted by Hugging Face | Generate the editorial spiral output from user-submitted text |
| Spiral name suggestion | same | One-line title for the spiral |
| Text-check ("would this text be received well?") | same | Predict three plausible responses with probabilities |
| Compatibility check (Pro) | same | Score two people across seven axes |
| Bible Mode | same | Anchor the verdict in a real Bible verse |
| Oracle Mode | same, with a per-card voice prompt | Stylistic rendering of the spiral in the voice of a fictional archetype card |
| Safety classifier (perpetrator / victim / safe) | same, with a deterministic keyword pre-filter | Refuse harmful prompts and surface crisis resources |
All AI calls are made server-to-server from our edge functions. The mobile app does not ship any model weights or run on-device inference, and it holds no AI provider keys.
What is sent. Only the text you typed for that request and — only if you turned on "Let Otto use my details" under Profile → About you — the optional details you entered there (name to be called, age, gender, relationship status, hobbies). Your email address, account ID and other spirals are never sent.
2. What the App does NOT use AI for
- Personalised advertising. We do not advertise in the App.
- Profiling for legal or financial decisions.
- Biometric identification, voice cloning, or deepfake generation.
3. AI output is fiction, not advice
Every AI surface is labelled as a "Take", a "Verdict", a "Map of the loop", or similar editorial framing. The App does not present AI output as fact, professional advice, or a real statement of a real person.
When no provider returns a usable answer (rate limit, malformed answer, outage) the App says so plainly: it shows the attempt count, tries up to three times, and then tells you the AI is busy. It never substitutes a pre-written answer and presents it as generated. You can cancel a run at any time; a cancelled run is not saved.
4. Safety measures
| Measure | Implementation |
|---|---|
| Refuses to produce CSAM, threats, hate speech, instructions for violence | Pre-prompt classifier + post-output filter; if either flags the content, the spiral is refused or regenerated |
| Crisis intervention | Victim-classified inputs trigger an interstitial with real region-specific helpline numbers (988, Samaritans, Lifeline, etc.) |
| Anti-abuse | 3-strike system, account bans and cooldowns, each recorded in a moderation log |
| Real-person defamation prevention | Post-output filter rejects libel patterns, fake-quote shapes, conspiracy patterns, and doxxing-shaped strings before saving the spiral |
| Religious-content honesty (Bible Mode) | Prompt explicitly forbids invented or paraphrased verses |
| Operator review | The moderation log gives the operator a complete audit trail of strikes and bans |
| Logging and metrics | We log when the help screen is shown so we can monitor crisis use patterns |
| Rate limiting | Server-side limits cap AI calls at 30 per hour and 120 per day per user, with an overall daily ceiling for the whole service; sign-in attempts are limited per IP address |
| Security logging | Rejected, malformed or suspicious requests are recorded in a security log (no written content) and serious events alert the operator |
| Input validation | Server-side caps on body size (256 KB), field length (200 / 2,000 / 8,000 chars), and malformed-JSON rejection make prompt-stuffing attacks impractical |
| Banned-user gate | Banned accounts cannot reach the AI surfaces at all — a banned account can only view its status, sign out, or delete itself |
5. How users can report harmful AI output
Three channels:
- In-app: at the bottom of any spiral, text-check or compatibility result, tap Report this output.
- Email: fizzythemusician@gmail.com — please include the spiral ID (visible at the top of the result page) and a short description.
- Appeals: if a moderation decision was triggered by an incorrect classification, see Community Guidelines § Appeals.
We review reports within 7 days. Confirmed harmful output is removed from the user's archive, and the underlying prompt + AI response are added to our internal evaluation set to prevent recurrence.
6. Training and data use
We do not train any AI model on user content, and we do not sell or license it for that purpose.
The AI providers process each prompt to return an answer, under their own API terms:
- Groq is the provider used for most requests.
- Google Gemini is used only while we have it switched on. On the tier of the Gemini API we use, Google may keep prompts and answers and use them to improve its products.
- NVIDIA and Hugging Face are fallbacks, used only when the providers above cannot answer.
Because of this, please do not put names, addresses or other identifying details of yourself or other people into a spiral. We never attach your name, email address or account to a prompt.
Aggregated, de-identified usage statistics (counts, error rates, fallback rates) are reviewed internally to improve the App.
7. Inputs and outputs we don't generate
- We don't generate images.
- We don't generate audio or voice clones.
- We don't generate video.
- We don't generate executable code intended to be run by the user.
If any of these surfaces is added in the future, this disclosure will be updated and submitted to Google as a material change to the App.
8. Provider attribution
The AI providers used by the App are:
- Groq, Inc. — open models (such as GPT-OSS and Llama), accessed via the GroqCloud API.
- Google LLC — Gemini models, accessed via the Generative Language API (only while switched on).
- NVIDIA Corporation — models served through the NVIDIA API catalog (fallback).
- Hugging Face, Inc. — open-source models served via Hugging Face Inference Providers (second fallback).
All three operate under their own terms and acceptable-use policies, which we comply with.
9. Contact
- Report harmful output: fizzythemusician@gmail.com
- Moderation appeal: fizzythemusician@gmail.com
- General privacy: fizzythemusician@gmail.com