UT Spark

All currently affiliated Students, Faculty, and Staff at UT Austin have access to UT Spark. If you see an error message saying "the administrator has restricted certain access," contact the UT Service Desk for assistance.

No. Other users cannot see your conversations unless you export them and share them yourself. On the admin side, prompt information is stored in an encrypted database. UT Spark does not allow data to be used for training AI models. Interactions are logged for security and support purposes only — they are not publicly viewable or reused.

Large language models generate responses probabilistically, not deterministically. Even with the same input, the model follows slightly different reasoning paths each time, and small differences in context influence the outcome. This variability is intentional — it gives the model flexibility and creativity. If you need more consistent results, try being more specific and explicit in your prompt, or use lower-temperature settings if available.

As of August 2025, UT Spark includes GPT-5.1, Claude Sonnet 4.5, and Claude Haiku 4.5 for general chat. Personal Agents are powered by GPT-4o. GPT-5 is also confirmed available. Models are released and updated frequently — always check the top-left corner of the chat window for the current list.

Yes. Personal Agents are available to all users. They are private (visible only to you), lightweight versions of Official Agents, and can be configured with your own custom instructions and knowledge sources (up to 20 files). See the UT Spark: How to Create a Personal Agent guide for step-by-step instructions.

UT Spark supports PDF (.pdf), Excel (.xlsx), Word (.docx), CSV (.csv), and common image formats including JPG, JPEG, GIF, SVG, and BMP. The maximum file size is 16 MB, and there is a 50-page cumulative limit across documents in a session or Personal Agent knowledge base. Personal Agents can store up to 20 knowledge files. Note: these limits and supported formats may change as the platform evolves.

A few common causes: (1) Scanned PDFs or Word docs — switch to Code Interpreter mode after uploading, which uses OCR to read the content. (2) Uppercase file extensions — rename .JPG to .jpg before uploading; uppercase extensions are a known issue. (3) Multiple different file types uploaded to a Personal Agent at once — try uploading in batches by file type.

You can submit feedback through the UT Service Desk (via the UT Spark: How to Get Help guide) or use the thumbs up/down feedback icons within the platform itself. Providing feedback on Official Agent responses is especially valuable — it helps the UT team refine those agents' accuracy and usefulness over time.

No. UT Spark does not allow data to be used for training AI models — external or otherwise. The platform is used solely for inference (generating responses from pre-trained models). All prompt and conversation data is stored in a per-user, access-controlled Azure database and is not shared with third-party vendors or used to improve model training.

The UT Information Security Office's Cloud Services Decision Matrix has not yet been updated to specifically address UT Spark (nebulaONE). Controlled data is generally considered safe based on the current OpenAI model implementation. However, if you're working with HIPAA health data, restricted research data, or other highly sensitive categories, check with the UT Information Security Office before proceeding. When web search is enabled on an agent, your prompts may leave UT's secure environment — avoid sensitive inputs in those cases.

Yes. UT Spark supports speech-to-text input via your microphone. You may see a browser permission pop-up asking you to allow microphone and recording access the first time you use this feature. Allow it to enable voice input.