Skip to main content
Talking Paper

AI Use & Limitations

Last updated: May 2026 (P-17 GDPR Art. 8 age floor disclosure) · Version: 3.9

NOT A MEDICAL DEVICE

Talking Paper is a journaling tool. It is not a medical device, not therapy, and not a substitute for professional mental-health care. The AI responses you see are language-model output, not clinical advice, diagnosis, or treatment.

Models in use

We route prompts through OpenRouter, which fronts multiple underlying model providers. As of May 2026 (P-17 GDPR Art. 8 age floor disclosure) the project routes the following models:

  • openai/gpt-4o-mini
  • text-embedding-3-large
  • cohere/rerank-v3.5

The active set may change as OpenRouter rotates routing. The canonical mapping of model → underlying provider is in our Subprocessor list.

Safety mitigations

We apply several guardrails before, during, and after each AI dispatch:

  • Tier-based AI request rate limits (free / pro / teams).
  • Per-user monthly token-cost cap with a soft-cap email at 80% (H-4).
  • Hard input-length cap on journal entries to prevent cost amplification.
  • Crisis-resource fallback gate (H-1): when our crisis classifier trips, the assistant defers to the crisis-line directory below instead of generating a clinical-sounding reply.
  • Streaming responses cancel upstream on client disconnect (F-5) so we do not bill abandoned generations.

If you are in crisis

Model limitations

Large language models can produce confident-sounding output that is wrong, biased, or unsafe:

  • Hallucination: the model may invent facts that sound plausible.
  • Bias: training data reflects systematic biases that can surface in responses, particularly around mental-health framing, gender, and culture.
  • No clinical history: the assistant cannot see your medical history, diagnoses, or medication. Its replies are not personalised to clinical context.
  • No continuity of care: the assistant is not a therapist and does not maintain a treatment plan across sessions in the way a human clinician would.
  • English-language bias: performance is strongest in English; nuance and idiom in other languages may be missed.

Prompt retention

Your journal text is sent to OpenRouter (and through it to the underlying model provider) only for the duration of generating a response. Our agreement with OpenRouter (commercial tier) provides for retention of inference inputs for up to 30 days for abuse-monitoring purposes; OpenRouter does not use your prompts to train models. The underlying model providers' retention is governed by their own published policies, linked from our Subprocessor list.

We use Langfuse for observability. Langfuse capture is opt-in only (H-2): when enabled, journal text is masked and truncated to 4 000 characters before it reaches Langfuse. You can opt out at any time from your privacy settings.

User feedback

If an AI response is inaccurate, unsafe, or distressing, please tell us. The fastest channel is our Contact form. We review every report and use it to tune prompts, classifiers, and the crisis-fallback gate. Where appropriate, we will also delete the offending generation from our logs.

Changes to this page

Material changes are versioned alongside our other legal documents. The current version is 3.9; see also our Privacy Policy and Terms of Service.