(Always) keeping your data safe
As a user of ChatMDR, ChatIVDR or ChatFDA, you know that one of the key promises of these tools is that they keep your data safe. Your inputs and outputs are not shared with anyone. They are not used for model training.
Why is using data for model training dangerous?
AI models are trained by evaluating how well a model responds to a certain input. This happens millions of times. Answers that correspond closely to the expected output get a higher score than answers that don’t. During training, the model goes through millions or billions of iterations. The model values (“neurons”) that give better outputs are preserved, those that give lower scores are erased and trained again.
In a very simplified example, if the question is What is the most famous quote from Hamlet? then:
- “To be or not to be” would get 10 points
- “To be or not” may get 5 or 6 points
- “I think therefore I am” would get maybe only 1 point (because it is at least a famous quote, attributed to Descartes)
The risk of this is that, just like with a person’s memory, some of the data of the training set is encoded into the model. You won’t find the literal text “To be or not to be” on the server, but the model can be triggered by a prompt and references to Hamlet or Shakespeare, in this case.
For Shakespeare or Wikipedia articles, that’s not so bad. But if you’re a medical device manufacturer and you are working in a niche market where only one or two other companies have the technology you do, you don’t want your data coming out. For example, if you have asked ChatGPT a question based on how your technology works, you don’t want that data to come out when your competitor asks ‘how does XYZ technology typically work?’
ChatMDR, ChatIVDR and ChatFDA safety promise
From day one, we have ensured that that scenario – your data leaking out into someone else’s AI output – can never happen. No matter how unlikely it is, we have completely excluded that event. Our tools will never train on your data. But you knew that already, so why are we bringing it up again?
Our supplier OpenAI has offered money to share prompts and outputs to improve their services. In the most valuable scenario, in eligible organizations that can be worth $75 a day for almost two months – in other words, up to $4,500. Yesterday, ChatMDR management got the email that our organization is one of the eligible companies for this premium.
But we said no. $4,500 or not, we gave you our word. We logged into our dashboard in the OpenAI organization to make sure we weren’t opted into anything. Below, you see the proof that all data sharing is disabled. Your data is and will always be: your data!
