AI Disclosure

This page covers the apps that use language models — currently Cloister AI, Jesus — WWJD and Best Dad — and satisfies the transparency obligations in Article 50 of the EU AI Act.

Where an app is not listed here, it contains no AI. Each app’s own privacy policy states what it does.

AI interaction

Where one of our apps generates a response, that response is produced by a language model, not by a person. Each app states this in its interface.

None of our apps simulate a real, identifiable individual. Jesus — WWJD generates reflective text in a scriptural register and states in the app that it is “an AI reflection grounded in Scripture — not the literal words of Jesus, and not infallible”. A setting switches it to third-person teaching.

We generate no synthetic audio or video of any real person.

Training

Nothing you write in any of our apps is used to train any AI model. Prompts, documents and generated answers are processed on your device and are not transmitted. There is no inference endpoint, prompt log or dataset.

We also do not:

• fine-tune models on user data;
• send prompts to a cloud AI provider — we have no relationship with OpenAI, Anthropic, Google or any other model API for runtime inference;
• retain generated output anywhere off your device;
• use your data to improve, evaluate or benchmark anything.

The sole exception is the text of a generated answer in Jesus — WWJD’s optional Voice mode, sent to a text-to-speech provider to be read aloud. Your question is not sent. The provider does not train on it. The feature is off unless you enable it.

Where inference happens

App	Inference	Runs offline
Cloister AI	On device — Apple Foundation Models, or an open-weights model executed locally	Yes
Jesus — WWJD	On device — Apple Foundation Models or a bundled open-weights model	Yes, except optional Voice mode
Best Dad	On device — Apple Foundation Models	Yes

Cloister AI ships with a working model and allows larger ones to be downloaded. Downloads are user-initiated and the licence and size are shown first. Downloading a model transmits a file request, not your data.

Limitations

• Output may be inaccurate. Language models produce fluent text regardless of whether it is correct. Verify anything you rely on.
• A document may be misread. Cloister AI cites the source passage for each answer so it can be checked.
• Models that run on a phone are smaller than cloud models and perform less well on difficult problems.
• Models have no knowledge of events after training. Cloister AI’s optional web search is off by default.
• There is no profile of you, no history across devices, and no learning from corrections beyond what an app stores locally.

Not professional advice

• Best Dad is general parenting guidance. It is not a doctor, midwife or nurse, does not diagnose, and must not be used to decide whether to seek medical care.
• Cloister AI is not a lawyer, accountant or clinician. Its redaction features are a tool, not a guarantee. Verify the output.
• Jesus — WWJD is an aid to reflection. It is not doctrine, pastoral counselling or spiritual direction, does not replace the Sacrament of Confession, and is not a crisis service.

None of our apps are emergency services. In Australia, emergencies are 000 and Lifeline is 13 11 14.

Human oversight

Every output is presented for you to accept, edit or discard. Nothing is acted on automatically and no decision about you is made by a model.

Where an app can take an action from generated output — Cloister AI creating a reminder or calendar entry from a meeting — the action requires your confirmation.

Automated decision-making

Our apps make no decisions about you that produce legal or similarly significant effects. There is no profiling, scoring, ranking or eligibility determination, and nothing engaging Article 22 of the GDPR.

Personalisation is computed on your device from what you told the app and is content selection, not a decision about you.

EU AI Act

The EU AI Act classifies systems by risk. Our assessment:

• Not prohibited practices (Article 5). No subliminal manipulation, no exploitation of vulnerability, no social scoring, no emotion inference in workplaces or education, no biometric categorisation, no untargeted facial image scraping.
• Not high-risk (Annex III). Our apps are not used for biometric identification, critical infrastructure, education access, employment decisions, essential services eligibility, law enforcement, migration, or administration of justice.
• Subject to Article 50 transparency, as AI systems that interact directly with people. This page and the in-app notices satisfy that obligation.
• General-purpose AI models. We do not develop or place models on the market. We integrate models developed by others — Apple, and the publishers of the open-weights models below — as a downstream deployer.

Best Dad and Cloister AI’s health-adjacent features are not medical devices. They provide general information and do not diagnose, treat, monitor or predict disease, and fall outside the EU Medical Device Regulation and the Australian therapeutic goods framework.

Australia

Australia has no AI-specific statute in force. The federal government has issued a Voluntary AI Safety Standard and consulted on mandatory guardrails for high-risk settings; neither imposes obligations on us.

The general law applies regardless, including the Australian Consumer Law prohibition on misleading conduct.

Models and licences

Apple Foundation Models

Apple’s on-device model, part of iOS, governed by Apple’s terms. Used by all three apps.

Cloister AI and Jesus — WWJD do not use Apple’s Private Cloud Compute for inference.

Open-weights models

Model	Licence	Used by
Qwen3 0.6B	Apache License 2.0	Cloister AI (bundled), Jesus — WWJD (bundled)
Qwen3 4B	Apache License 2.0	Cloister AI (optional download)
Gemma 3 4B	Gemma Terms of Use	Cloister AI (optional download)
Llama 3.2 3B	Llama 3.2 Community License	Cloister AI (optional download)

Full licence text for bundled models is included in the app and shown under Settings, Model licences. Downloadable models display their licence before you download them.

We use these models as published and have not fine-tuned them on user data.

Other components

• MLX — Apple’s array framework, used to execute open-weights models on device. MIT License.
• Text-to-speech — an ONNX speech model in Cloister AI, running on device. Jesus — WWJD’s optional Voice mode uses a third-party service, described above and on the Subprocessors page.
• Scripture text in Jesus — WWJD is the Douay–Rheims (Challoner), in the public domain. Copyrighted translations and the Catechism are not included.

Copyright

The models we use were trained by their publishers under their own terms.

Output from a generative model may resemble existing work. You are responsible for ensuring that anything you publish from an app’s output does not infringe another person’s rights.

Reporting a problem

Report output that is harmful, dangerously wrong or inconsistent with this disclosure to contact@leapstudio.dev.

Inference happens on your device and we have no record of it. Include the wording, what was asked, and which app and model.