– The reason we have achieved this is that we have developed and invested in our own platform, GPT UiO, and that we have had a good collaboration with NTNU. We are actually the first in Norway to offer a solution for conversational AI with up to red-level data. We are very proud of this, and it is our competent and interdisciplinary development teams that have made it possible, says section manager Pål Fugelli in the IT department.
Now the functionality in GPT UiO can also be used on so-called "red data," that is, confidential information that requires extra protection. The color classification, and the basis for it, is described in UiO's classification guide.
What is new?
It has long been possible to choose which language model GPT UiO should use when you give it a task. Now two of the models are approved for data up to the red level. All models are labeled according to the traffic light principle we know from the Classification Guide and Storage Guide. This makes it easy to see which models you can use for which types of data. The number of models approved for the different classes will vary over time.
– It is good that we focus on having high local expertise. Then we do not become completely dependent on American cloud services. Currently, the two models approved for red data run on machines physically located at NTNU. Later, we plan for some red models to run at UiO. New models are constantly coming, and some models are also removed from GPT UiO. Running the models locally increases our digital independence and resilience, explains chief engineer Oddmund Møgedal, who works on the development of GPT UiO.
Be careful!
Although it is now possible to input red data into GPT UiO, there are still some things you must be careful about. – Remember that the purpose and legal basis under GDPR must be concretely assessed for all processing and storage of personal data, and that UiO's routines for processing personal data must be followed. You must choose the correct language model yourself based on the data classification. The traffic light principle solution has been user-tested and makes it easy to see which models can be used. However, it is your responsibility to select the correct model, Møgedal reminds.
Fugelli adds: – And we must remember that red classification is still a limitation. There is also black data (strictly confidential), and that cannot be entered into any of the models in GPT UiO. We are looking into the possibility of running language models within TSD (Services for sensitive data). Such models could potentially be used with black data, but such a solution is currently only in the planning phase.
Want to know more?
- Read about GPT UiO
- UiO's guide for classification of data and information (Classification guide)
- If you are unsure about the collection, storage, or processing of red data, contact your nearest manager.