How MaWiki uses AI.

1. What AI does in MaWiki

MaWiki uses one universal AI runtime for chat, document assistance, Today briefings, notes and idea development, Site Builder, Design Studio, reporting, workflows, scheduled assistants, and repository development. MaWiki owns the logical conversation, permission policy, tool access, event history, artifacts, usage and cost record.

The selected provider owns only its native execution segment. Changing model, provider, account, computer, or cloud runner creates an explicit handoff within the same MaWiki conversation. MaWiki does not edit one provider's private session files to make them look like another provider's session.

Read and draft tools can run automatically within the declared surface and your existing permissions. Publishing, sending, pushing code, spending money, or another outward effect asks for approval unless your workspace policy and your explicit session choice allow it. Tenant policy always remains the ceiling.

2. Which models we use

You choose a provider, access method, model, and runner. Availability depends on your workspace policy, subscription, device, and the capabilities required by the current task.

2.1 Cloud execution

Cloud-direct requests go to the configured model API. Hosted coding requests run in a fresh isolated cloud workspace. Cloud execution can incur provider or MaWiki service costs. MaWiki never changes a desktop run to paid cloud execution without your acknowledgement. A workspace can later allow automatic fallback, but the cost warning remains visible.

2.2 Desktop and local execution

MaWiki Desktop can manage Codex, OpenCode, Claude Code, Cursor, and local llama models in isolated MaWiki-owned provider homes. Setup, sign-in state, compatible versions, updates, capabilities, and diagnostics are shown in AI settings. An external binary override is marked unverified.

A local model keeps inference on the device. A desktop subscription CLI follows that provider's own network and account behavior. MaWiki does not copy desktop credentials into the cloud. Moving repository work between desktop and cloud requires a pushed branch or an explicitly reviewed patch checkpoint.

For full detail on the data that flows in each mode, see the AI section of the privacy policy.

3. Known limitations

Generative AI models can be wrong. The output is generated by predicting the most likely next words and is not a verified statement of fact. We cap that risk by feeding the model excerpts from your own documents and asking it to cite them, but you should still treat AI answers as a starting point, not a final source of truth.

• Inaccuracy: the model may misread the source documents, combine unrelated documents, or invent details that are not in the source. Always check the citations.
• Outdated knowledge: cloud models have a training cutoff. Information added to the world after that date is not in the model unless it is in your own documents.
• Style and tone variation: the same question can produce different responses on different days. This is normal for generative models.
• Coverage: the assistant can use only the workspace data, connected sources, tools, and web access you are permitted to use. Web access is separately controlled and off until enabled.

Every AI-generated response in MaWiki is labelled AI-generated and a disclaimer about accuracy is shown under the chat input.

4. Safety measures

We use a combination of model-side and product-side safeguards to reduce the chance of harmful or inappropriate output:

• Trusted providers: cloud AI uses models from established providers (Anthropic, OpenAI) that ship their own content-safety filtering and refuse policy-violating prompts at the model layer.
• One runtime boundary: chat, document, design, coding, and automated assistant work use the same MaWiki session, permission, event, and artifact controls. Provider, model, runner, and paid-cloud changes require an explicit handoff or acknowledgement.
• Workspace-scoped retrieval: every MaWiki tool call is checked against the current tenant, user permissions, product and feature before execution.
• Local-AI provenance: when running locally, MaWiki uses curated open-source models distributed by Hugging Face. We document which model files we recommend and ship the configuration in the desktop app, so you can audit and replace the model yourself if you prefer.
• Per-message reporting: every AI message has a Report button that opens a quick form. Reports go straight to our team for review (see section 5).
• No silent training: MaWiki does not use your conversation content to train models. Cloud providers' training behaviour follows their own published API terms.
• Explicit support capture: anonymous telemetry excludes prompts and responses. Detailed support capture is separate, encrypted, time-limited, and enabled only by you. See the privacy policy.

5. Reporting a problem

If an AI response in MaWiki looks wrong, harmful, offensive, or otherwise problematic, please report it. The fastest way is the Report button next to any AI message inside the app. A short form lets you pick a category (inaccurate, harmful, offensive, or other) and add an optional comment.

Reports are logged against your workspace, reviewed by the Edge Software Pty Ltd team, and used to fix prompts, refine retrieval, and identify wider issues with model behaviour. We do not share report content with other users.

If you cannot use the in-app reporter (for example, the issue prevented you from logging in), email us at ai-reports@mawiki.com with as much detail as you can share.

6. Your responsibilities when using AI in MaWiki

Generative AI is a tool, not a substitute for human judgement. By using AI features in MaWiki you agree that:

• You will not rely on AI output as the sole source for decisions that have legal, financial, medical, or safety consequences.
• You will not use MaWiki AI to generate content that breaches the law in your jurisdiction, infringes someone else's rights, or violates the published terms of the underlying model providers.
• You will respect the privacy of other people whose information you might place in your workspace, including by not uploading content you do not have the right to use.

7. Changes to this policy

We update this policy when we add, change, or remove AI features in MaWiki. The date at the top of the page reflects the most recent change. This page will always show the current policy.

Questions about this policy can be sent to privacy@mawiki.com. For the underlying data-handling rules see the privacy policy.