Google Released Gemma 4 as Apache 2.0. Someone Immediately Asked If the Weights Are Actually Open.

Google announced this week that Gemma 4 is now Apache 2.0 licensed, framing it as a milestone in a two-decade commitment to open source. The official Bluesky post was warm and promotional — "giving builders the autonomy to innovate without limits" — and collected the kind of modest engagement that corporate announcements usually attract. Then, with 701 likes, came the response that defined how a significant slice of the community actually received the news. A self-described long-time listener asked three questions in plain language: Are the code and algorithmic weights open source? Did the training process use scraped code without compensating the developers behind it? And is Google retaining user data as part of its $100 million Series B — a reference to the broader pattern of AI companies bundling data rights into funding rounds. The questions weren't hostile. They were the kind a careful person asks before trusting something.

This is where the open source AI conversation lives right now — not in arguments about capability benchmarks or inference costs, but in a persistent credibility gap between what companies announce and what builders actually want to know. The Gemma 4 launch is a good example of a genuine concession: Apache 2.0 is a real license with real permissiveness, and the open source crowd has, in other contexts, stopped arguing about Google when the licensing terms hold up to scrutiny. But the questions with 701 likes aren't about the license — they're about the layers underneath it. Open weights without open training data is a known half-measure. A permissive license that coexists with data retention clauses is a known contradiction. The community has been burned enough times to know the difference between a press release and a commitment.

What's telling is that the productive, technical conversation is happening in parallel. Another post circulating this week noted that open source developers using AI saw a 19% productivity drag in a 2025 study — but when the study was rerun this year, the number had flipped to an 18% gain. That reversal, if it holds, is genuinely significant for the argument that open models are worth the governance complexity. Elsewhere, builders are quietly demonstrating what structured open-source models can do: delivering outputs for fractions of a cent per workpaper, running self-hosted on Hugging Face, closing the gap with frontier models on specific tasks. The case for open source AI has never been stronger on the technical merits. The case on the trust merits is exactly as strong as the least-answered question in that 701-like thread.

Google will almost certainly point to the Apache 2.0 license as evidence of good faith, and that's not wrong. But the person who asked about the weights, the training data, and the data retention wasn't asking in bad faith either — they were asking the questions that determine whether "open source" means anything in practice or just in press releases. Until those answers are as prominent as the licensing announcement, the community's skepticism isn't a communication problem for Google to manage. It's the correct response to incomplete information.