SW Sadeepa Wijesekara On The TWIML AI Podcast

“In the agreements, the frontier model providers say they are not training with that data. That may be true. But before data gets to training, there are multiple layers it goes through. All the prompts are getting logged, and they may be erasing that too. And then there are some evaluations they run. So they are not training the model, but evaluation might capture that scenario, and then the evaluation is what kicks off and says, this should be the answer. That's the level of detail that's not disclosed.”

The TWIML AI Podcast · AI Builders · November 2026

“In the agreements, the frontier model providers say they are not training with that data. That may be true. But before data gets to training, there are multiple layers it goes through. All the prompts are getting logged, and they may be erasing that too. And then there are some evaluations they run. So they are not training the model, but evaluation might capture that scenario, and then the evaluation is what kicks off and says, this should be the answer. That's the level of detail that's not disclosed.” — Sadeepa Wijesekara, The TWIML AI Podcast

Context

Sam raised recent accusations that a frontier lab trained on a researcher's data and published first. Sadeepa says contract promises not to train on customer data leave open what happens in logging and evaluation. It is her speculation about what providers don't disclose, not a claim about a specific case.

Transcript

The TWIML AI Podcast Around 41:58 into the episode
Sadeepa Wijesekara

…a data gravity issue? Do you want to send your sensitive data? And these days, data and your internal processes is your IP. And once it goes to the cloud, how sure you are that it is protected? So I think those are the risk questions that you should be asking, not necessarily the cost questions.

Sam Charrington

…side. But then there's this whole other set of factors around privacy and IP leakage and control and like seeding your competitive advantage to a, for many organizations, to a potential future competitor. A lot of things that we heard in the days of cloud, right? That's right. They're analogous.

Sadeepa Wijesekara

And I look at them, they are kind of real in a way, because when you send a prompt out to a GPU, that prompt is visible in clear text. Administrator can just do some commands and dump the memory and see what's going on. So that is a risk if you're willing to take that because the prompts has real, could be a real business scenario that you might want to protect. So if you're ready for that, then you can go to Frontier and it's not a concern for you.

Sam Charrington

Yeah, it kind of strikes me that the closest analogy to this is like the, I don't know what we want to call it, the math gate stuff with OpenAI. Like, you know, there's a point in time where, you know, the, you know, the frontier model provider is going to steal my data as like of this big…

Sadeepa Wijesekara

Yeah. And I think, and they may, I mean, in the agreements, right, with the frontier model providers, they said they are not training with that data. That may be true. But before a model gets to that data or for training, there's multiple layers it goes through, right? So there's definitely going…

Sam Charrington

you're not training on my data, but are you making synthetic data based on my data that you are training on?

Speaker names come from our own diarization. The position is estimated from where the line sits in the episode.

More from The TWIML AI Podcast