Meta Quietly Rewrote Its Terms of Service — and the AI Training Question Isn’t Going Away
AI news this week has a theme running underneath the headlines: who controls the tools, the data, and the decisions. Meta is quietly editing what it says it can do with your information. OpenAI is shipping steady updates to the software developers use to build AI products. And the companies building on top of AI are starting to pick sides. These stories are connected, even if they don’t look like it at first glance.
Meta Edited Its Terms of Service. Here’s What Changed — and What Didn’t.
Terms of service — the legal documents almost nobody reads but everyone agrees to — just became a little more interesting. According to TechCrunch, Meta recently removed language from its terms that had raised alarms about whether the company could scrape websites to collect data for AI training. Meta says the change is a clarification, not a policy shift. The company insists it hasn’t changed how it actually trains its AI systems.
The concern that prompted the change is worth understanding. Web scraping means using automated software to copy text, images, or other content from websites at scale — the same way a spider crawls a web. AI companies have used scraped data to teach their models how language works, what the world looks like, and how to answer questions. When people spotted language in Meta’s terms that seemed to allow this broadly, it sparked worry among publishers, journalists, and anyone who puts content on the internet.
For everyday people, this matters because the content in question often isn’t abstract. It’s news articles, forum posts, product reviews, and personal blogs. Whether or not Meta actually changed its practices, the fact that users and publishers pushed back hard enough to trigger a public rewrite signals that scrutiny of AI training data is intensifying. People are paying attention now.
Why this matters: The legal language companies use around AI training is becoming a battleground. What’s written in a terms of service document can define what’s legally permitted — and what isn’t — for years.
“Meta removes language that raised scraping and AI training concerns.”
OpenAI’s Python Library Just Got Two Updates in One Week
A software library — think of it as a pre-built toolkit that developers plug into their own applications — just received two updates from OpenAI in quick succession. Version 3.35.0 of OpenAI’s Python library added support for something called typed Live sideband events. Version 3.35.1, released shortly after, fixed a bug. Both are available on OpenAI’s GitHub repository.
The new sideband events feature sounds technical, but the idea is approachable. When a developer builds an app where users have a live conversation with an AI, sometimes the app needs to handle other signals at the same time — things like transcription status, connection health, or metadata about what the model is doing. Sideband events let developers manage those signals in a structured, predictable way, rather than writing custom workarounds. The bug fix that followed addressed a case where the connection handling those signals was quietly dropping chunks of data.
For people who don’t write code, these updates are invisible. But they matter because every AI app on your phone or laptop is built on libraries like this one. When the underlying toolkit handles data more reliably, the products built on top of it become more stable and less likely to behave strangely mid-conversation.
Why this matters: Two releases in one week suggests OpenAI is iterating quickly on live, real-time AI features. That category of AI — where the model responds instantly to voice or text in real time — is where a lot of consumer products are heading.
“Support for typed Live sideband events in v3.35.0.”
Also Happening in AI
A few other developments worth tracking today. Ollama, a tool that lets people run AI models on their own computers rather than through the cloud, released version 0.40.6 to fix a Windows bug that was leaving hidden background processes running after use — a small fix that matters for anyone who cares about what’s quietly running on their machine. LiteLLM, an open-source library that lets developers switch between different AI providers without rewriting their code, shipped version 1.106.0 with cryptography-related updates, according to GitHub. On the business side, Replit — a popular platform for writing and running code in a browser — has switched its default AI assistant to OpenAI’s models, moving away from Anthropic, according to TechCrunch. Separately, TechCrunch reports that Anthropic’s own workforce has shifted toward more senior employees who are spending more time in meetings, a pattern that often signals a company moving from scrappy startup mode toward something more structured.
What to Watch
The Meta terms-of-service story is a preview of a larger legal fight that hasn’t fully arrived yet. As governments in the US and Europe push toward clearer AI training data rules, companies will face real pressure to make their practices explicit — not just buried in legalese. Watch whether other major AI labs quietly update their own terms in the coming weeks. Replit’s switch from Anthropic to OpenAI is also worth following: when developer platforms pick a default model, they shape what millions of users interact with daily, and those choices tend to consolidate power quickly.