About AI review of quality documents
Dejan Kosutic at Advisera Expert Solutions wrote the example about the review of documents by an AI compliance app (see „Compliance implementation projects“, „Review of documents“ https://advisera.com/articles/compliance-future-with-ai).
Review of documents. […] For example, if you have written a Quality Policy, but omitted a commitment to continual improvement, the app will suggest that you include this commitment because it is required by ISO 9001.How would an AI theoretically review a quality document?
Well, when we analyze how a very popular AI is functioning, we will see that it won´t take long when it will happen. Andrej Karpathy, AI researcher and founder member of OpenAI explained on a Microsoft conference how GPT-2 gets trained, let´s say how GPT learns (https://build.microsoft.com/en-US/sessions/db3f4859-cd30-4445-a0cd-553c3304f8e2).
At minute 2:46 he is explaining Tokenization on a text sample, you can try it for yourself by using the https://platform.openai.com/tokenizer an paste the text of ISO 9001:2015, 5.2 Policy, 5.2.1 Establishing the quality policy in this tool.
Top management shall establish, implement and maintain a quality policy that:
a) is appropriate to the purpose and context of the organization and supports its strategic direction;
b) provides a framework for setting quality objectives;
c) includes a commitment to satisfy applicable requirements;
d) includes a commitment to continual improvement of the quality management system.
At minute 5:00 he explains the a) Pretraining of GPT-3, while following him in the video please acknowledge that is information is from the calendar year 2020.
At minute 11:40 you´ll understand how OpenAI made it working with this chat or answering machine, b) supervised finetuning.
All you need to do is to tip in the ISO 9015:9001 standard text and a dozen quality policy documents in the AI, called a) pretraining (minute 5:00, Andrej´s presentation). Believe me you´ll find a dozen free samples online. Then hire ISO 9001 consultants for a few hours to b) supervised finetune (minute 11:40), and you are ready to go. Feasible job, really.
The challenge remains, even with AI, effectiveness.
I believe all this powerful AI tools that come up next time will help us clear paper, sorry documented information. But no one should forget, it´s about adaption in business. I can´t see any spot on the horizon on that without breaking the law in personal surveillance. We have a lot to do here, example to make it impossible that I´m able to reverse engineer AI model to a personal behavior of human individuals. But this is exactly what we do in generative AI, we want AI to tell us the core of its recommendations (ex. source of fact check). When we want to use AI as a kind of control and recommendation for the effectiveness of an QMS we need to the absolute opposite.