AI Transparency Notice
Partner replies, AI edits, generated images, vector clean-up, transcription and other generated results may be wrong or synthetic. Scribble identifies AI at the point of interaction and keeps a human in control of insertion, saving and publication.
1. What Scribble’s AI is
Scribble is an interface to a model provider chosen by the user. RQAI does not train a general-purpose model on your notes. You supply a provider credential, select a model and initiate each request. Depending on the feature, the request can include your prompt, active conversation, selected text, relevant note content, writing-voice profile, image or temporary audio recording. The current app sends the request from your device to the selected external service rather than through an RQAI prompt proxy.
2. Providers, routing and roles
Supported configurations may include Anthropic, Google Gemini, OpenAI, DeepSeek, Groq, OpenRouter and custom compatible endpoints. OpenRouter is an intermediary and routes to a model provider, so more than one external organisation may receive the request. DeepSeek publishes PRC processing/storage information. See the Provider & Data Routing Notice.
Local Whisper performs transcription on-device after downloading a model. Browser Web Speech is controlled by the browser/OS. External providers determine model behaviour, retention, safety filters, geographic processing and improvement/training controls.
EU AI Act roles must be assessed feature by feature. A model company may be a general-purpose model provider; RQAI may have provider obligations for a Scribble-branded AI system/interface; and a professional user or organisation may be a deployer. Personal, non-professional use may be treated differently. This notice is transparent about the current assessment and is not a regulatory certification.
3. EU AI Act transparency controls
- Direct interaction: Partner and AI entry points identify AI from the beginning of the interaction.
- Human control: output is a proposal. It is not silently inserted or published; the user decides what to keep and publish.
- Synthetic content: generation is labelled in the interface and users must preserve/add disclosures required by law or provider/platform terms.
- Machine-readable marking: provider/model capabilities vary and ordinary text can lose provenance when copied. RQAI is conducting a feature-level Article 50 assessment and does not claim that every present export carries a durable interoperable mark. Do not remove provider provenance metadata or rely on its presence without checking the exported format.
- Accessibility: notices use visible text rather than colour alone.
Relevant Article 50 transparency obligations apply from 2 August 2026. Current European Commission guidance should be used when assigning provider/deployer duties and technical marking requirements.
4. Public-interest text, deepfakes and disclosure
Professional deployers may need to disclose AI-generated/manipulated public-interest text and deepfake image, audio or video. Human editorial review and responsibility can affect some text-publication duties; artistic or satirical context can affect presentation but does not permit deception. DeepSeek’s current terms separately require users who disseminate its output to verify it and clearly identify AI generation.
Scribble flipbooks and drawings are user-directed creative tools. If AI clean-up/generation could mislead a reasonable viewer about authenticity, disclose it. Do not create non-consensual intimate material, fraud, deceptive impersonation or fabricated evidence.
5. Known limitations
- Models can hallucinate facts, citations, quotes, code and events.
- Output can reflect bias or omit relevant perspectives.
- Safety filters can fail or over-block.
- Aliases, model names, capabilities, routing and terms can change.
- Generated code/interactives can contain errors; Scribble sanitises and sandboxes previews, but exported material still requires testing.
- AI cannot determine copyright ownership, confidentiality, defamation, consent or legal compliance for you.
6. Prohibited and high-impact use
Do not use Scribble AI for prohibited manipulation, exploitation, social scoring, unlawful biometric categorisation/identification or sensitive-trait inference. Do not use it as the sole or decisive basis for recruitment, employment, education admission, credit, insurance, benefits, policing, migration, justice, healthcare, emergency or other legal/similarly significant decisions. Scribble is a writing tool, not a high-risk decision system.
7. Data protection, eligibility and AI literacy
Before sending personal data, minimise/remove identifiers, understand provider retention/training and establish a lawful basis. External providers have their own account eligibility/guardian rules; Scribble’s local note-taking core does not verify age. Organisations should screen for a DPIA, complete applicable transfer assessments and train staff on limitations, privacy, security, bias, provenance and human review.
8. Reporting
Report harmful, misleading or incorrectly labelled AI behaviour to hello@rqai.co.uk with “Scribble AI report” in the subject. Do not include provider keys, confidential notes or unnecessary personal data.