Terms of Use KI.Toolbox
Scope and purpose
The KI.Toolbox is a service of the Scientific Computing Center (SCC) at the Karlsruhe Institute of Technology (KIT) for the use of various generative AI models (artificial intelligence). The service is currently aimed exclusively at KIT employees. A later opening for other KIT members (including students, guests and partners) is planned.
The provisions of the KIT ICT regulations on permitted use, user obligations, operator authorizations, measures in the event of violations, logging and liability apply. These terms of use supplement the ICT regulations exclusively with AI-specific requirements (data categories/risk classes, choice of model locally vs. Azure, use of checks, export controls, data protection features).
Access, authentication and accounts
Access is via KIT account via OpenID Connect/SSO. Two-factor authentication is currently not required.
Roles/rights: Standard users and administrators. Group/project roles with special rights are technically possible and can be assigned by the SCC in special cases.
Service description, models and providers
Open source models hosted locally at the SCC vary depending on availability and version. The decommissioning of local models is announced at least 4 weeks in advance and disseminated via SCC message channels.
For the OpenAI models obtained via Azure, the training use of customer data is disabled; processing within EU regions (Germany, Sweden).
The service is operated as a "best effort" - high availability and performance are aimed for, but without an SLA or availability guarantee.
Data categories, protection classes and model selection
Personal data (Art. 4 No. 1 GDPR) may only be processed via the models hosted locally at KIT. Processing of personal data via the Azure/OpenAI models is prohibited.
Special categories of personal data (Art. 9 GDPR) are generally not permitted in the KI.Toolbox (very high risk).
Risk classification of the KIT: decisive is https://www.isb.kit.edu/134.php.
- High-risk use: only with locally hosted models
- Very high risk use: prohibited
The lawful processing, selection of a suitable model and necessary protective measures are the responsibility of the user.
It is prohibited to enter export-controlled, sanction-relevant or other security-restricted information into the KI.Toolbox.
Logging, storage, deletion and backups
Chat histories, uploaded files and AI-generated outputs are stored in Open-WebUI on a user-related basis for one year; users can delete these independently at any time. When the account is deactivated (e.g. by leaving the KIT), chat histories are deleted after 24 hours at the latest.
System backups with standard rotation/retention periods may result in deleted data being retained in backups for a limited period of time; operational access only takes place for recovery and security purposes.
For capacity and billing purposes, when an AI request is sent, the metadata of the request (user ID, time, selected model, number of input/output tokens) is stored on SCC servers and is not merged with other data. This data is stored for one year. The content of the requests and the answers are not stored for this purpose. It is not possible for users to delete this metadata themselves.
Training, fine-tuning and evaluation
User content (prompts, uploads, outputs) is treated as strictly confidential by the SCC and is not used for training, fine-tuning, evaluation or research purposes.
Training use by the third-party provider is deactivated for the models provided via Azure.
Copyright and sovereignty of results
Users are responsible for compliance with copyright, license, trademark, personality and confidentiality rights and other third-party rights. The relevant KIT regulations apply.
Generated content may be inaccurate, incomplete or legally problematic. Before further use, an appropriate technical and legal review by the user is required.
Users retain the rights to their own entries. Users acquire rights of use to generated content to the extent permitted by law; the rights of third parties remain unaffected. License conditions of individual open source models may also have to be observed.
Permitted use
Use in examination contexts is only permitted with the express permission of the responsible examination supervisor and within the framework of the applicable examination regulations.
The creation and use of manipulative AI content (e.g. deepfakes) without identification/authorization is prohibited. The model must be selected strictly in accordance with the 'Data categories/risk classes' section.
Fair use
The offer is free of charge for KIT employees; there are currently no fixed quotas/rate limits. The SCC reserves the right to introduce consumption-based billing for the use of the KI.Toolbox in the future; all users of the KI.Toolbox will be informed of this at an early stage.
Users are obliged to use resources fairly in order to ensure availability for all.
Responsibility of the users
There is no content or legal review of inputs or outputs. Generated content may be incorrect or misleading and must not be used without reflection. The professional review and responsibility for further use lies with the user.
Data protection
The privacy policy for the KI.Toolbox is authoritative.
Governance, compliance and documentation
The provisions of the KIT ICT regulations on permitted use, user obligations, operator authorizations, measures in the event of violations, logging and liability apply. Furthermore, the KIT AI guidelines(https://www.kit.edu/downloads/KI-Leitlinien-de.pdf) and the KIT risk classification(https://www.isb.kit.edu/134.php) must be observed.
Automatic archiving of official results does not take place. Required archiving and documentation must be ensured by the users (e.g. export of chat histories).