Regulatory Open Forum

 View Only
  • 1.  ChatGPT translation of regulatory docs - FDA

    This message was posted by a user wishing to remain anonymous
    Posted 21-Apr-2023 10:36
    This message was posted by a user wishing to remain anonymous

    Hi all, 
    so far we have been using certified companies (ISO 17100) to translate clinical EU documents to be submitted to the FDA. Translations included investigator brochures, local competent authorities approval letters, user manuals etc. We have been quite disappointed by the extremely high cost and the poor results (since of course the translator is not a subject matter expert) and we are considering now to change strategy (and save money) and start using ChatGPT to have a first translation and have internal employees (fluent in English) to review the translation. Is this approach acceptable for FDA or there is an expectation of certified translations?
    Should we include a statement saying that the translation is done internally?
    Thanks for your help/



  • 2.  RE: ChatGPT translation of regulatory docs - FDA

    Posted 01-May-2023 14:42

    This is a great question and hopefully others can chime in.



    ------------------------------
    Ajit Basrur
    Worcester MA
    United States
    ------------------------------



  • 3.  RE: ChatGPT translation of regulatory docs - FDA

    Posted 02-May-2023 00:45

    I do not have much experience with the agency's generative AI-generated language translation, I suspect neither do many others.

    From what we have observed, the translation by GPT3.5+ is fairly accurate in a variety of languages. Ultimately, you own the translation, so the final quality is the responsibility of the sponsor, not the service.

    Without getting into specifics, it may be prudent to use GPT API rather than ChatGPT to prevent your data from entering the public domain.



    ------------------------------
    Pravin Jadhav
    CEO
    Lansdale PA
    United States
    ------------------------------



  • 4.  RE: ChatGPT translation of regulatory docs - FDA

    This message was posted by a user wishing to remain anonymous
    Posted 03-May-2023 07:53
    This message was posted by a user wishing to remain anonymous

    For translations, FDA expects the name, address, and a brief statement of the qualifications of the person who made the translations. A translation of literature or other material in a foreign language is to be accompanied by copies of the original publication and an English translation that is verified to be complete and accurate.




  • 5.  RE: ChatGPT translation of regulatory docs - FDA

    Posted 06-May-2023 12:14
    Hi Anon-
    Can you provide a reference where it states the expectations of FDA for these translations?
    Typically, I also recommend using certified translation companies, however I am getting pushback on costs, etc. to support global registrations of our FDA product.

    Thanks-
    Allison

    Sent from my iPhone




  • 6.  RE: ChatGPT translation of regulatory docs - FDA

    This message was posted by a user wishing to remain anonymous
    Posted 10-May-2023 08:33
    This message was posted by a user wishing to remain anonymous

    Hi Allison, 
    This is in 21CFR10.20 Subpart B (c)(2).




  • 7.  RE: ChatGPT translation of regulatory docs - FDA

    Posted 11-May-2023 01:18

    Using ChatGPT for regulatory submissions sounds like a potential nightmare.  In the context of medical devices, the expectation is that the manufacturer evaluates suppliers, contractors and consultants "on the basis of their ability to meet specified requirements, including quality requirements" and to "define the type and extent of control to be exercised over the product, services, suppliers, contractors, and consultants, based on the evaluation results" (21 CFR 820.50).  At the very least, proper evaluation of a LLM like ChatGPT would likely prove difficult simply due to the nature of the model and the training data used in its creation.  Add to that the potential for a model to evolve beyond the initially evaluated state or for errors to be introduced due to the nature of the information being translated (not to mention potential bias in the training data) and you may end up with a struggle using such a service. 

    As to whether or not one may include a statement stating that the translation is done internally, that does not seem to be a completely honest statement if an externally developed LLM is used in the process. 



    ------------------------------
    Christopher Erwin
    Scottsdale AZ
    United States
    ------------------------------