Balancing Potential and Pitfalls: Leveraging Generative AI for Encapsulating Human Interaction

Here’s a scenario that will likely be familiar: you join an important work meeting, whether it be a pitch, a brainstorm, or a strategy session. You listen intently; the conversation is lively and engaging. You jot a few notes on a notepad or in a word document. A week later, you go to take action on that meeting and when looking at your notes to recall the discussion, what you have in front of you is completely useless. It scratches at the conversation but does little to prompt next steps or accurately reflect what was covered.

Here’s another: you start your day by taking a look at your calendar and find you’re double booked. No problem, you think, I’ll just skip one of the meetings and have the team catch me up. But when speaking to the team later on, you get different bits of information from different people. You take a look at the meeting notes and find similar scrabble as the above example.

In both cases, the content and context of a meeting is immediately lost when the meeting ends. We spend hours every week in meetings, much of it remotely now. To solve the problems our businesses and organizations face, we meet to share ideas, discuss, collaborate and decide how to move forward. Humans are good at this; what we’re not so great at is collecting the output of our meetings in a way that is concise, actionable, and useful.

This is where Generative AI comes in. These models generate new and original content based on a query and a dataset used to train them. If you’ve been on the internet in the last several weeks, it’s likely you’ve encountered the output of such models. Whether it was your acquaintance animated as an astronaut via Stable Diffusion or answers to existential questions by ChatGPT, newsfeeds have been abuzz with experimentation, awe and critique of what Generative AI can do.

And it has the potential to solve the problem of capturing the essence of human interaction, crucially in the case of meetings. With Headroom, we’re using Generative AI to summarize meeting discussions in real-time, generate action items, and provide data and analytics on meeting effectiveness. Then, we build a searchable knowledge base for easy discovery and recall.

Generative AI relies on multiple modalities such as text, video, and audio to make predictions and decisions. As with all artificial intelligence models, the larger the datasets powering them, the more diverse the output can be on a wide range of subjects. The Generative models of today are better than ever because of their capacitance and vast quantity of training data they’re built on.

As these models learn, adapt, and advance, they grow more “human” in their ability to create. The images get more stunning, the answers more accurate, and content more useful. While the applications for Generative AI are seemingly endless, there are also challenges to be considered, namely the challenge of preserving private information. In many cases, the data used to train and test these systems is highly personal, and its misuse can have serious consequences for individuals and society as a whole.

Due to “hallucination”, large language models easily corrupt information, making privacy preserving methodologies especially important. Hallucination refers to a model generating text or images that are not based on any curation of the training data, but rather on patterns it has inferred from the statistics of the training set itself. This can lead to the model generating text that is nonsensical or contains information that is not factually accurate; images that defy nature and physics.

In the case of using Generative AI for remote collaboration, the opportunity and the risk are both present. The content of your meetings is deeply personal: your ideas, personal information, and opinions. You can imagine how that personalization can be invaluable. You get meeting summaries that are specific to your line of business. You receive action items that are assigned to you and phrased in a way you understand. This is possible because the model is trained on your information and knows you.

The risk here is also critical. If that information is also used to train the models of your colleagues and others outside of your organization, it is likely personal information and company-level information could leak and corrupt the Generative output of someone else’s workspace.

To combat this risk at Headroom, we develop a hierarchical model structure to protect privacy:

1. Population Model: This model is trained on generalized internet information available to all people. This gives the model great breadth; it can speak on many subjects in a high-level, generalized way, such as GPT.
2. Organization Model: This model is trained on the population model and then fine-tuned to a given company based on their specific data. The inference from that company-level data is only available to people within that organization. This allows the model to be both broad in scope and deep on organization-specific topics.
3. Personal Model: This model is trained on the population model, tuned on the organization model and personalized on an individual. The inference from that individual’s information is only available to them. Thus the model is broad in scope, deep in organization-level topics and specific in personalized information.

The conventional theory of Generative AI models is that more training data equals better models. This results in models that are a jack of all trades, master of none. Our goal with Headroom is to create models that are jacks of all trades and masters of you. By combining generalized models with highly-specific ones, we are able to do just that, while critically protecting private information. We imagine a world where artificial intelligence, Generative AI in particular, makes the ways we interact and work together more human, creative, productive, and secure than ever before.