AI Company Reaches into Chat Box: OpenAI Logs Keystrokes, Google Mines Uploaded Content to Train AI

According to Dynamic Beating monitoring, Axios has started investigating what user data AI companies are actually collecting.

The findings revealed that just chat logs are no longer sufficient. Now they also want to know what you click on your computer, what photos you send, what content you view, and then use this data for memory, personalization, and advertising purposes.

OpenAI has added Computer History. After Mac users opt in, ChatGPT and Codex can record clicks, inputs, and switches within specified apps and websites to gradually piece together a long-term working memory of yours. The feature is off by default, does not capture screenshots or recordings. OpenAI states that temporary operation records will not be used for model training.

Google is taking a more aggressive approach. The new Search Services History will store images, files, and audio/video uploaded by users through services like Lens, Search Live, and Translate, which can then be used for training generative AI models. Eligible accounts are opted in by default, and users must manually opt out if they wish. Some data, even after being dissociated from an account, can be retained for up to 4 years once incorporated into the training pipeline.

Advertising has also caught up. Meta uses interactions between users and Meta AI to adjust the content and ads on Facebook and Instagram; OpenAI and Google are also exploring advertising models for chatbots. The better AI understands you, the more detailed the profile in the hands of the advertising systems becomes.

Previously, the concern was about chat logs being used for training, but now even how you work on your computer and what you throw into Google are becoming data in the hands of AI companies.