Human-AI Training Parallelization: Empowering Life Sciences Without Replacing Expertise

In this insightful interview, Richard Jaenisch, senior director of Education, Outreach, and Digital Experience at Open BioPharma Research and Training Institute, and David Jaenisch, director of Technology at Prompting Integration and Consulting LLC (PICLLC), discuss their groundbreaking Human-AI Training Parallelization (HATP) framework, a practical approach to integrating generative AI into life sciences workflows without compromising human expertise or regulatory integrity.

At its core, HATP works by parallelizing existing systems rather than replacing them, embedding AI only where it adds value within a given workflow. The framework centers on two key methods: digitally interactive SOPs (DISOPs), which consolidate SOPs and training materials into a single transparent, auditable platform for competency-focused training; and project-based HATP (PBHATP), which captures tribal knowledge and enhances project efficiency through AI assistant development.

Critically, the DISOPs system operates outside of GxP workflows—functioning as a pre-GxP training tool—while being built for full transparency and audit-readiness as organizations approach compliance requirements like 21 CFR Part 11. The platform currently runs on Google's Gemini API, with options for cloud-based client silos or fully local deployment to protect sensitive intellectual property.

Richard and David also address pressing concerns around data privacy and IP ownership. Using Google's API allows organizations to opt out of model training on their data, and existing case law generally supports the position that human-edited, system-maintained content remains owned by the individual or organization that created it.

DISOPs itself is a life sciences-specific customization of PAIR, PICLLC's underlying AI framework, tailored to meet the unique regulatory and operational demands of the industry.

Be on the lookout for Parts 2 and 3 of this interview!

Transcript

Editor's note: This transcript is a lightly edited rendering of the original audio/video content. It may contain errors, informal language, or omissions as spoken in the original recording.

My name is Richard Jaenisch. I am the senior director of Education, Outreach, and Digital Experience at Open BioPharma Research and Training Institute in Carlsbad, California.

I am David Jaenisch. I am the director of technology at Prompting Integration and Consulting LLC, PICLLC for short.

So for the most part, human AI training parallelization will is just effectively parallelyzing whatever system you're already deploying. So generally speaking, it is functionally what you are already doing, but it's enabling AI in the specific components in the workflow where it is pertinent. And so it really depends on whatever AI deployment you're using.

So in that concept, conceptually, for the for that system, it really depends on what you're using. In our case, what we're using is DISOPs, or digitally interactive SOPs, as well as using another piece, of basically Gemini.

But what we are doing though is not in the GxP workflow. So in this case, it's very external because it's all pre-GxP. So in this case, it isn't really hitting that area just because it's mostly on the research side in terms of what we've deployed. We are getting to the point now where we're actually entering 24 CFR Part 11.

However, DISOPs as a program, while its name does imply that it is interacting with the SOPs, it is external. So the idea is that it's taking the SOP, taking the materials that you're working with and putting them all together into one system, and that system is entirely transparent. So it can go through, it can be validated through any method that's already in play. But the thing is, it's generally not as part of your GxP component. So it's more of the training element. So usually if you have a situation where you need to have retraining, where you need to have that, correction, then this is an opportunity to be able to use that.

And then in that sense, you have a completely transparent record of not only the training itself, but how the exact training was deployed. So the benefit of the tool is that it is completely transparent in its entire methodology.

So it is awkward for a generative AI tool to have that, but because of the method of how it breaks down each component and where the actual user specifically has the articulate control of the components, all of that produces a digital record that can then be audited at any time.

In terms of how our product works, we currently work with Gemini on the cloud. It can technically work with any cloud computing instance, but it can also run locally. So if your concern predominantly is about IP theft, locally will obviously make that literally impossible. And if you are running it on the cloud, we create silos for each of our individual customers, so that way they are completely separate from any other client, and there will be no data sharing between them.

If your concern is predominantly about the idea of the model being trained on it, Google explicitly allows you to disable this functionally when you use the API. Certain API use, they allow you to just say, "I'm not training on this." This is also true for their enterprise client. To my knowledge, it's not true for Open AI currently, due to an ongoing lawsuit with The New York Times. I'm not 100% sure about that. Don't quote me on legal.I don't know legal. What this means is you're safe from the perspective of Google getting anything from you or from other companies as well.

And in terms of ownership of IP, that is a complex ongoing legal question in terms of generated content in general. However, Google definitely doesn't own it. Yes. We've decided that. It is either, no one or it is the company, depending on the situation and level of editing. And because these are edited after the fact and they are maintained within a system that's maintained by people and turned into a holistic list of questions, I believe that it would be more likely they'd be owned by the person making them as opposed to being un-copyrightable in that case.

Generally speaking, there have been court cases so far that explicitly state that the human owns the copyright, not the AI tool. So anything generated is generated by an individual, that individual who created the system. And our system is done in such a way that it would not be in that same facet.

But I will say this though, the project-based component of project-based human AI training parallelization on the project's basis, that is owned by the process owner. So whoever the process owner has that IP of whatever the project is and how that's developed, that is a separate contract and a separate system that we generally do at Open BioPharma. And so we have those types of projects that are deployed frequently from time to time. And so in that case, those are deployed and owned by the process owner, which is typically the company we work with. Although occasionally there is a partnership collaboration of some kind.It's not usually a straight partnership. It's usually a situation where IP is developed and there's a lot of discussion around how that works. That being said, that is project-based, and so it's very unique in that way.

The DISOPs component is what David is speaking to. And so those elements, he's talking specifically as PICLLC created PAIR, which is basically the underlying layer of DISOPs. So DISOPs is a customized version of PAIR built specifically for life sciences to handle the unique situations that occur within life sciences. Just so I have some clarification on those pieces.