How are reused the data collected by Trados CoPilot?

Hello everyone,

Here the more detailed version of my question:

Does the data collected by Trados CoPilot's LLMs (GPT-3.5 and GPT-4) when working with it on a project are stored on a dedicated server linked to our RWS account ?
Or are they sent  to OpenAI servers to train their LLMs ?

I ask this because I attended the webinar on June 25th about Trados Studio 2024 and its new features, and at some point someone asked: "How does the use of A.I. affect confidentiality issues?"

To this question, Daniel Brockmann answered we should ask our customers if they're okay with us using A.I. to be more productive, like in any typical translation contract's part about confidential content. If the customers are not okay, we just have to disable the option. Up to us to decide in cases with no real confidential matters from what I understood.

But he didn't specify if it could be a problem in the sense that the data collected by Trados CoPilot could reappear in its future suggestions on other projects (or just be stored somewhere less protected against hack, like our PC), or if it could be a problem in the sense that the data collected could be used by OpenAI to train its LLMs.

So : is Trados CoPilot directly connected to OpenAI servers?

I really need to know.

Thanks in advance for your help.


Best regards,
Céd'.

emoji
Parents
  •  

    I am not an RWS employee, just a user like you. I assume you are referring to the "AI Assistant" in Trados Studio.

    First of all, Trados Studio is a software that runs on your machine. You can watch any data it sends anywhere using free programs like Fiddler (Classic) or, if you intend to dig seriously into this, WireShark.

    "AI Assistant" is a component of Trados Studio 2024 (there are plugin precursors for 2022) that sends data from Studio to OpenAI or Azure (whatever you define in the configuration), receives data from these services and displays them in Studio. This data transfer happens from your machine to OpenAI/Azure. RWS is NOT the man in the middle. You can see this in Fiddler. You can see the entire data package that goes out, and the return package that is returned to your machine. There is no black box here. OpenAI and Azure themselves are black boxes, but they both state that they do not use data from their APIs for training purposes. (You should make sure this still applies if it's crucial for you.) Do they keep their word? Your guess is as good as mine.

    If you are working with highly sensitive data, chances are your customer won't want this data to be transmitted anywhere. But chances are then that they won't send you Trados packages by email, which is notoriously not-secure. However, this is just about Trados Studio.

    There is a checkbox about taking part in some customer feedback programme, and I honestly don't know which kind of data this gathers. I would be interested to know.

    Settings dialog box with Customer Experience tab selected, showing options for the Trados Copilot Customer Experience Improvement Program. 'Yes' option is selected to join and help improve the AI Assistant.

    Hope that helps a bit. If you are suspicious, consider this: RWS has a reputation to lose. The gain they would have from sneakily harvesting data is very limited, and the potential damage would be colossal. But do install Fiddler and see for yourself. You can see all data transfers, like to MT providers etc.

    Daniel

    emoji


    Generated Image Alt-Text
    [edited by: RWS Community AI at 10:03 AM (GMT 1) on 21 Aug 2024]
  • Hello Daniel,

    Thanks a lot for your answer Daniel, it really helps!

    Indeed, I referred to the AI Assistant.

    Thanks for these softwares you recommended me too, I didn't know them.

    I would like to use your answer for an internship report where I conduct research on the use of big data in the field of translation.

    Would it be all right if I quote you?



    Best regards,

    Céd'

    emoji
  •  

    Well, all contributions to this forum are public, and you have the right to quote. I mean, thanks for asking, but you actually don't need my permission - you just need to reference your source.

    Daniel

    emoji
Reply Children