Are we all just unpaid AI trainers?

A couple weeks ago, Meta announced it’s installing new software on employees' computers that tracks their keystrokes and mouse movements. The information collected will help train its artificial intelligence. And that’s apparently sparked some backlash within the company over privacy concerns.

That got us thinking about all of the other ways that people and businesses are training AI – some ways more apparent than others.

For a survey of this landscape, “Marketplace Morning Report” host Sabri Ben-Achour spoke with Panagiotis Ipeirotis, a professor of Technology and Business at NYU Stern. The following is an edited transcript of their conversation.

Sabri Ben-Achour: Is this just something freaky that Meta cooked up, or is this part of some larger trend on how we train our AIs?

Panagiotis Ipeirotis: Fifteen years back, I was working with oDesk, which is now called Upwork, which was allowing people to hire contractors anywhere in the world. And to establish trust, one of the things that they were doing was they were recording the screen of the employees and the keystrokes and the mouse movements, etc. So, this thing exists for almost 20 years.

Ben-Achour: What's new about it, I guess, is the fact that it's done for purposes of AI training. So, I just wonder if there are ever new ways that we are training AI and may not know about it.

Ipeirotis: Well, this goes back — 20 years back. Do you remember all these CAPTCHAs?

Ben-Achour: Yes, like, “click here to prove you're not a robot”?

Ipeirotis: Yeah, and you need to click on a bicycle, on a bus, recognizing different objects. You are providing labels to a confused AI system that tries to figure out: Is there really a bus here? And so on. So, all these reCAPTCHA is providing directly training data. But also, when people participate on Reddit — people ask questions, there is a lot of discussion back there, people vote on the answers. This is very useful and extremely often leveraged by the LLMs when they are trying to get knowledge about obscure topics.

Ben-Achour: Are AI firms constantly looking for new and creative ways to train their models?

Ipeirotis: 100%. There is a general policy question: Who gets the benefit of this data? For example, you go on Reddit, you participate in a discussion, and Reddit afterwards sells this data to LLM companies: to OpenAI, Anthropic, Google, etc. It's actually one of the biggest sources of revenue. You, as a Reddit participant, don't get paid anything. The LLM companies, they train their models. They get value out of that. So, the only person that doesn't get value out of it is the people who contribute the data.

Ben-Achour: Are there any guardrails or laws that would regulate the privacy aspect of this? Or, the people kind of “unwittingly training models” part of this?

Ipeirotis: Companies typically do have the right terms of service that gives them the ability to mine the data, train on the data, and so on. You do have the right to opt out often, but this is literally just deleting your data from someone's database. This doesn't work well when we're talking about trained models. We don't really have much of a framework around such scenarios.