From Open Source to “Open Washing”
Open source is back. But “open” AI often isn’t. Llama-like and “open-source-plus” licenses keep code visible while limiting real freedoms. Copyleft (GPL) regains relevance as a clear, non-discriminatory counterweight to open-washing.
How copyleft becomes relevant again in the era of Llama-like licenses
The renaissance of open source, and the backdoor of restrictive licenses
Open source is alive again. Not only developers, but also decision-makers without a programming background now talk about LLM weights and permissive licenses. But that renewed interest has a downside: some of the most visible “open” AI projects come with terms that are anything but open.
Meta set the tone when Mark Zuckerberg presented the Llama models in 2023 as “open-source AI.” Yet the accompanying license excludes commercial redistribution, directly violating the Open Source Definition’s “free redistribution” criterion.
https://opensource.org/definition-annotated
The Open Source Initiative (OSI) publicly corrected Meta; Llama 3 is still “not open source by any definition.”
https://opensource.org/blog/metas-llama-license-is-still-not-open-source
Still, Meta’s campaign inspired other companies to hijack the term. From small startups to established names: they call their model “source-available,” while adding restrictions on competition or cloud distribution.
Eigent: British multi-agent hype with its own “open-source-plus” license
A closer-to-home example (well—across the Channel) is Eigent, a UK-based startup that maintains a “multi-agent workforce” on GitHub. The README showcases an “Eigent Open Source License,” formally based on Apache 2.0 but with additional conditions that restrict commercial SaaS exploitation.
https://github.com/eigent-ai/eigent/blob/main/LICENSE
With clauses like these, the source remains visible and verifiable, but the freedom to integrate the project wholesale into closed-source products is missing. That puts Eigent in the same category as Llama: source-readable, but not “free” in the OSI sense. In other words, the source is open, but it is not open source.
That is not inherently a mortal sin; companies are free to choose their own licensing strategy. The problem starts when marketing ignores the nuance and still labels the project “open source.” Then you’re not only violating community norm, you’re also creating legal risk for integrators who miss the fine print.
Copyleft as a counterweight
Ironically, this proliferation of “open-washing” is making classic copyleft license —the GPL family— attractive again. Where permissive licenses (MIT, Apache 2.0) mostly maximize freedom, copyleft licenses enforce reciprocity: anyone may use your code, commercially included, as long as improvements flow back.
If you distribute binaries without source, you must provide a written offer for the source and require later resellers to do the same.
https://www.gnu.org/licenses/gpl-faq.en.html
Yes, that is a “commercial restriction” in the sense that you can’t easily hide proprietary forks. But it’s a restriction that is clear upfront, applies to everyone, and does not discriminate by business model. That fundamentally distinguishes copyleft from Llama-like “non-commercial” clauses, which are selective—and therefore not approved by the OSI.
Why more and more companies choose source-available anyway
Legal advisors are signaling a broad shift in licensing: from open source to “source-available,” giving investors more control over IP and data exclusivity.
AI companies in particular fear that competitors can undermine their market plans with a simple fork plus cheap GPU time. The reflex is understandable—but the result is a gray zone where terms like “community edition,” “ethical open source,” and “commercial-friendly license” circulate without any uniform definition.
What does this mean for Dutch CIOs and development teams?
- Read the license, not the press release. Never trust the “open source” label without checking the clauses—especially for AI models and agent frameworks.
- Check for OSI approval. An OSI identifier guarantees the ten OSD criteria are met, including free redistribution and non-discrimination.
- Re-evaluate copyleft. The obligation to contribute changes back may feel inconvenient, but it provides contractual clarity—and prevents later dependency on a vendor who changes the rules.
- Watch downstream effects. If you integrate code with a Llama-like clause into your SaaS, it may complicate future negotiations with partners, investors, or acquirers.
Conclusion: precise language restores trust
The open-source ecosystem has always drawn its strength from transparency and unambiguous definitions. Now that AI dominates the hype curve, marketing departments seem to forget those hard rules. By consistently distinguishing copyleft and OSI-approved permissive licenses from “source-available” and proprietary solutions, we keep the conversation honest.
Computable readers don’t need to become lawyers—but they do need their terminology in order. Feel free to call Eigent an interesting British agent experiment, but don’t label it “open source” without a disclaimer. And the next time you hear a keynote calling Llama an “open model,” calmly ask: “Under which OSI license exactly?”
Because real open source begins —and ends— with the license text.
Original article in Dutch
Externe auteurs
