AI Policy

For teachers: policy by group or student

Open Groups, choose a group and expand AI policy on the right. Under Policy for, select Whole group or one student. You can block AI access, enable Closed AI mode, block internet, images, code execution or connectors, and limit context messages or interactions. Options you do not set continue to use the existing school policy.

A student policy is added to the group policy and can only make it stricter. A teacher can therefore adapt the learning environment without weakening safeguards set by the school or organisation. Use Remove own policy when the exception is no longer needed.

Use the settings for a clear teaching purpose: turn off internet when students must reason from supplied sources, limit interactions for short feedback rounds, or use Closed AI mode to reduce distractions. Explain the choices to students and review individual exceptions regularly; they are a support measure, not an automatic judgement or punishment.

The separate AI access panel can open a supervised window of 30, 45 or 60 minutes when the group is normally blocked. Extra restrictions can be applied during that window. Organisation and school blocks, access hours and student-specific restrictions remain in force, so temporary access cannot bypass the school's safety baseline.

Through AI Policy you manage rules that are enforced server-side for every AI request. The policy is separate from daily quotas, access times, and the Features tile.

Levels and Exceptions

The Environment tab contains the default policy. Under Schools and groups you only add deviations:

1. add a school exception;
2. if necessary, load the groups within that school;
3. if necessary, load the learners within a group;
4. configure and save only the desired exception.

Schools, groups, and learners not added inherit the parent policy. When resolving policy, the order is environment → school → group → learner. A stricter block at a higher level cannot be lifted at a lower level.

Availability

Choose Always on or Off. Access by days and hours is managed separately via Access Times. Both controls apply: if AI policy is off, access times do not grant access.

Limits

• Max. previous messages in context limits how much conversation history is sent to the model.
• Max. context tokens further limits the estimated token size. The newest messages are retained; you will be notified when older context is omitted.
• Max. output tokens limits the maximum length of a response.
• Max. interactions per session stops new turns after the number is reached.
• Max. session duration stops new turns after the set time has elapsed.

Leave a field empty to inherit the value from the parent level. Below each field, the management environment shows the effective value and the source. The lowest positive limit from environment, school, group, and learner is binding. Regenerations, document generation, and tool usage also count as interactions and use the same limits. The desktop runtime fetches this policy server-side before each local generation.

Theme Warnings

A theme warning checks prompts and responses for specified keywords. With Keywords and contextual AI analysis, the analysis can also recognize synonyms and provide contextual relevance, reason, and a safe summary. Choose a theme, detection method, keywords, severity, and action:

• Signal creates an alert but allows the request to proceed.
• Block stops the request.
• Escalate creates an alert and sends it via the configured in-app and/or email channels to the designated recipients.

Use Add Preset to quickly add an editable base for self-harm, violence, bullying, or substance use. Always review the keywords, severity, action, summary instruction, and recipients for your own educational context. A preset only becomes active after you save the policy.

A free model output never blocks independently: only a deterministic keyword match can enforce the Block action. The contextual analysis can provide an additional signal or explain a possible false positive. For each theme, provide a neutral summary instruction, set a retention period, and configure escalation recipients only within the allowed school scope.

When a response is subject to block policy, streaming is temporarily delayed. The full response is first checked and only then saved and made visible. A blocked response is not saved in the conversation or version history.

See Following up on Theme Warnings for filters, status transitions, ownership, and privacy.

Closed Educational Context

Closed educational context is a strict baseline profile for judicial or otherwise highly controlled educational settings. It blocks by default voice chat, internet and image searches, image and document generation, code generation and execution, MCP/connectors, and assistance with hacking or bypassing.

Lower levels cannot lift these blocks. You can also enter an additional policy instruction; this is added server-side alongside the regular system prompt. Availability, access times, limits, and alerts are set separately.

In AI-School, closed mode also removes the following features from the interface and blocks them on the server:

• voice chat and transcription;
• file uploads and File Management;
• assistants;
• support requests.

Text chats and their history remain available. Once a chat message is completed in closed mode, a learner can no longer overwrite the prompt, response, or saved versions. Regenerate therefore creates a new linked chat message with a version audit label; the original message remains unchanged. If closed mode is later disabled, messages previously created in closed mode remain immutable.

After closed mode is disabled, the normal options become visible again. The server rechecks the policy for realtime connections, transcription, uploads, and support requests, among other actions; hiding a menu option is therefore never the only security measure.

Checking

After a change, start a new chat with an account at the relevant level. Check both allowed behavior and the expected block, and also review the alert overview for theme warnings.