Will there be a difference in how the answers are generated based on the versions of the software? For example chat GPT 3 and 4?
Based on my experience, there are some differences. For a long time, I didn't subscribe to GPT, but when I did and got GPT-4, I noticed a few key improvements. One notable difference is that GPT-4 gives you access to images and related features. For example, you can upload an image and ask, "Can you tell me what this image is about?" and GPT-4 will provide an answer. This capability is not available with GPT-3.
As for whether GPT-4 provides better answers, I'm not entirely sure. Personally, I don't often need very lengthy responses.
I also use GPT to write small programs. For instance, if I need to check a document for specific information without uploading it, I can ask GPT to create a program for that task. Once, I had about a hundred documents where a customer used a term inconsistently. I asked GPT to write a program that would open all the documents in a folder and identify the errors in the term usage. It accomplished this, allowing me to deliver results without uploading any documents.
It's important to note that GPT isn't just for answering questions. It can be useful in various other ways, such as creating scripts to search through and correct large volumes of documents.
How do you use AI to generate audit reports?
Here are some examples of how to use Chat GPT:
- "Please write me an ISO 13485:2016 Audit Plan"
- "Please create an audit report template in the form of a table for the whole of ISO 13485:2016"
- "Please create a column with the first question for each line"
- "Please write me the evidence that I should be looking for as an auditor"
- "Please put this in a column in the table"
How do you (redesign) processes to ensure the power of AI is used while still keeping end responsibility the human expertise?
- Develop and Implement AI Guidelines
- Develop an AI policy for the company
- Define Clear Roles and Responsibilities
- Identify Key Areas for AI Integration
- Take one area and create a test implementation
- Verify and Validate the Implementation
- Establish Monitoring and Feedback Loops
- Integrate AI with Existing real QMS Processes
- Training and Skill Development
- Maintain Human Expertise
- Iterative Improvement
You've talked about the fact that you have to pre-train Chat GPT before it can give you the right answer. How can we learn pretraining?
Pre-training is similar to how you would train a person, though not in the sense of pre-trained transformation, which is how the tool was originally developed. Instead, it refers to guiding the tool to provide useful feedback in the way you need it.
Consider a document with a table. You might want the table to have specific columns and names across the top. You could instruct the tool, "Please list the hazards typical for an auto-injector." If you notice something missing, you can ask, "What about unpackaging the device?"
Every question you ask during a session contributes to training the device.
If you want to populate your table, you can request that the tool present the answers in specific categories. Typically, it will provide a list, but you can ask it to format the information into a table with the desired columns.
Google Gemini at the start of the year was boasting a 10-million-context window. Is this something I could use now to 'dump' my entire QMS into it and start analyzing it with the tool, and is this something you would recommend?
I would not recommend this approach. I find that Gemini does not provide particularly rich output when handling large volumes of information. Instead, decide on your specific goals and work with one document at a time, setting clear objectives and expectations for the output.
Which AI tools or software are used for QMS?
ChatGPT and Gemini can assist with tasks such as creating tests and checking for compliance gaps in documents. For monitoring updates and changes to regulations, standards, and adverse events, www.hoodin.com is a useful tool. Additionally, www.freyrsolutions.com/artificial-intelligence-ai-regulatory-services offers some tools, but they are supplemented by manual input from Freyr consultants.
Would IFU translations made by AI be accepted by auditors or notified bodies (NBs)? And should we include AI as one of the service providers in our QMS?
No, AI-generated translations are not likely to be accepted by auditors or NBs without validation. Translations need to be validated, and relying solely on AI without a thorough 1:1 validation of the IFU may lead to failures. You should declare how you use AI and include it on your list of service providers, but it should not be subject to validation in the same way as other service providers.
What is the current landscape of AI tools to help in QMS and regulatory submissions?
AI-driven QMS systems leverage large language models to enhance various aspects of quality management, including:
- Document Control: Automating the creation, review, and approval of documents, ensuring compliance, and reducing manual errors
- Training Management: Streamlining the training process by providing personalized learning paths and tracking employee progress.
- CAPA Processes: Identifying root causes of issues and suggesting corrective and preventive actions to improve quality.
- Quality Event Management: Enhancing the tracking and resolution of quality events, such as non-conformances and deviations.
- Audits: Automating audit scheduling, execution, and reporting to ensure thorough and efficient audits.
- Supplier Quality Management: Monitoring supplier performance and ensuring compliance with quality standards.
- Change Control: Managing changes in processes, products, or documents to maintain quality and compliance.
- Risk Management: Assessing and mitigating risks to ensure product and process quality.
AI tools for regulatory submissions leverage large language models to:
- Document Management: Automate the creation, review, and organization of regulatory documents, ensuring accuracy and compliance.
- Compliance Tracking: Monitor and track compliance with regulatory requirements, reducing the risk of non-compliance.
- Regulatory Information Management: Manage and organize regulatory information efficiently, facilitating easy access and retrieval.
- Submission Planning and Tracking: Plan and track the progress of regulatory submissions, ensuring timely and organized submissions.
- Document Authoring and Review: Assist in the authoring and review of regulatory documents, improving the quality and consistency of submissions.
- Regulatory Strategy: Develop and implement regulatory strategies, ensuring alignment with regulatory requirements and business goals.
How would you recommend a company start using AI with their QMS?
First, establish a clear policy. Define where AI can and cannot be used within your QMS.
Second, subscribe to a company account for the AI system. This ensures that everyone uses a centralized, official account rather than private ones. Having a company account helps maintain consistency and reduces the risk of bias. It also ensures that the AI's use aligns with your organization's standards for truth, integrity, and substance.
How do you ensure the confidentiality and security of sensitive data when using AI tools?
You need an information security management system policy, which outlines how digital data should be handled. This policy should cover various aspects such as email management, company information, and data uploads.
ISO 27001 is a relevant standard for information security management systems, but implementing it requires specific training. Your company should also have policies for internet and cloud usage, and employees should be trained on these policies.
For example, a common policy is to prohibit storing company data on private storage. More complex scenarios include consultants who access your network remotely. Even if consultants follow your company's standards, they don't own them, and downloading standards to their own systems can lead to a loss of control and potential violations of agreements.
Therefore, clear policies are essential to define what actions are permissible on the company network and what are not.
