You are the product!

Are you happy for your video calls, emails and messages to train AI?

In March 2023, Zoom changed their online Terms of Service (note enterprise customers have different contracts) with Section 10 aiming to better explain how they use and who owns the various forms of content across their platform.

Just think about that for a minute. Who do you think owns the conversations you are having over Zoom or any other video conference platform? If you are like most people, you would say that the platform is providing a service but the content IP should remain with the user.

But should it be different if you are paying for a service, versus using a free service that is monetised through data or advertising?

Now more than ever, if it is free - you are the product!

So how do you feel about Google using Gmail to train Bard? This tweet from Kate Crawford went viral as it showed Bard claiming that it was trained on Gmail content. Google officially responded and denied the claim, stating that Bard can hallucinate and make mistakes.

And if you ask Bard today how it is trained it will say that it is trained on the public web, the world factbook, social media, forums, online conversations and of course google search. However, it also notes that there are other specific public and private data sets which are confidential.

The reality is that Gmail, Google Search, YouTube are all free services, you are being targeted with ads based on this content and that ad targeting has long been powered by AI – it’s just you couldn’t talk to that AI! You might say that being targeted by AI is one thing, but having something predict what we are thinking is another – let’s just pause to consider how does Gmail provide those helpful autocomplete sentences? Blake Lemoine, a former Google employee responded to the tweet from Kate Crawford claiming that LaMDA (the LLM behind Bard) is also what drives autocomplete and autoreply in Gmail, claiming that de-personalised Gmail content is being used as AI training data.

It is no secret that our public content is being used to train AI, the challenge comes around the transparency of knowing what private content is being used to train AI. There has been a lot of public debate around Elon Musk buying Twitter for the purpose of training AI and Meta launching Threads for the same reason.

But interestingly when Zoom casually updated their Terms of Service (TOS) in March 2023, no-one really noticed – until in May, when they announced a collaboration with Anthropic, an artificial intelligence company that conducts research and develops AI products such as Claude. That’s when people started putting two-and-two together and after receiving “customer feedback” (aka substantial backlash), Zoom further updated their Terms of Service in August 2023 to state that...

“Zoom does not use any of your audio, video, chat, screen sharing, attachments, or other communications like customer content (such as poll results, whiteboard, and reactions) to train Zoom’s or third-party artificial intelligence models.”

Zoom’s integration of Claude into their platform offers an interesting case study in separating the use of LLMs without training the foundation model. Zoom IQ is the resulting service that can provide meeting summaries and compose team chats.

The Future For Now team has been using AI in meetings for some time now, it acts as an assistant - taking notes and providing an action list, but we have always asked permission of the meeting attendees.

It is great to see that Zoom are applying this best practice, giving users control by defaulting this new functionality to an “off” position and providing notifications to all in the room when it is enabled.

With the rise of privateGPTs we are going to see more companies creating a responsible framework for use of and training of LLMs.

Image Source: Zoom https://blog.zoom.us/zooms-term-service-ai/