Who owns the model? Rethinking openness in the age of foundation AI

As AI exploded into the mainstream with the 2022 launch of ChatGPT, a fierce debate simmered around the tech’s new backbone: foundation models.

These vast systems suddenly fueled everything from OpenAI’s chatbot to Copilot’s coding. But their power provoked contentious questions. Who truly owns these models and determines what they do?

In Silicon Valley, responses were divided. OpenAI locked its foundation models in-house, granting outsiders access to outputs only under tight commercial terms. Meta released its Llama model weights — essentially the “brain” of the AI, containing all its learned knowledge — under an “open-weight” license, making them downloadable while keeping training data and code private. Google kept its foundation models proprietary and bound by restrictive licenses.

But beyond Big Tech’s dominions, a wave of alternative approaches was taking shape. Each offered a distinct vision of model ownership.

In Washington, D.C., EleutherAI pursued radical transparency. The non-profit releases all model artifacts — weights, training code and datasets — under permissive open-source licenses like Apache 2.0 and MIT.

Stella Biderman, EleutherAI's executive director, a mathematician and computer scientist, wants to create a world where anyone can build models from scratch.

“We've invested really hard in trying to make more people able to own, control, and study this technology,” Biderman said. “We think that's really important on the scientific front.”

Biderman grew up “idolizing” the Free Software Foundation, the Open Source Initiative and the 1990s hackers who fought for encryption rights in the “Crypto Wars”.

“That's very much where I'm coming from philosophically. I want there to be some real sense in which you can own, control, modify and redistribute a model checkpoint, but also really understand what your program is… and influence it to behave the way you want.”

EleutherAI treats its models as public goods, prioritizing transparency and accessibility over proprietary rights. Within months of its 2020 launch, the non-profit unveiled The Pile, an 886-gigabyte dataset for training large language models (LLMs). A year later, the lab released GPT-J, the world’s largest publicly available GPT-3-style model at the time. Further open LLMs followed, showing top-tier models could be built without tech giants.

“Having model weights is a really good form of ownership, but being able to build your own model is an even stronger form of ownership,” Biderman said. “It’s expensive, and kind of a waste of resources if you don’t need it. But if you do need it, it should be an option.”

Another option is using EleutherAI’s pre-built open models — a popular choice for startups. “They like the fact that they could actually control the technology in a way that OpenAI wouldn't let them.”

If EleutherAI represents radical transparency, Sentient Foundation in San Francisco represents radical decentralization. The organization is running an unusual ownership experiment aimed at preventing any single entity from controlling artificial general intelligence, or AGI.

“Our goal is to make sure that open AGI remains open and decentralized,” said Himanshu Tyagi, co-founder of Sentient.

Tyagi and his team want to turn every model contributor into an owner. First, they establish identity through “fingerprinting”; the technique embeds a hidden signature in the model. “The first prerequisite for any kind of ownership or control is identity,” Tyagi said. “If you cannot trace back a copy to its origin, then there is no ownership at all.”

Every action is then logged on the blockchain, from curating data to fine-tuning the model. Contributors can even receive automated rewards, turning developers into stakeholders. 

Models can also be aligned to the needs of communities, which can then arbitrate disputes. “We ourselves do not want to dictate these governance routes,” Tyagi said.

These plans have won support from Silicon Valley titan Peter Thiel. Last year, the billionaire’s Founders Fund co-led an $85 million seed round in Sentient.

Fingerprinting is also gaining traction. In February, over 650,000 people applied the technique to secure fractional ownership of a Sentient AI model. 

Tyagi said this is just the start. “We think that countries and religions eventually will have their own models fingerprinted.”

In Israel, AI21 Labs represents a hybrid approach. The startup provides open-weight foundational models that developers can customize, alongside controlled versions for enterprises.

Founded by the CEO of self-driving pioneer Mobileye, a Stanford computer scientist, and a veteran of Israel’s elite intelligence unit 8200, AI21 maintains full ownership of its model development pipeline. 

“That means we take full responsibility for their design, training and deployment,” said Amnon Morag, AI21 Labs’ VP of Product.

“It also means we carry the accountability: ensuring the models are robust, transparent and aligned with enterprise needs. That’s a very different kind of ownership than what you get if you’re only consuming a closed-source API.”

The approach straddles openness and control. Smaller models get released under open licenses, while the enterprise-grade models remain proprietary, accessible only through commercial licensing agreements with full vendor support."

“We support open approaches where customers can fine-tune or adapt models for their own purposes,” Morag said. “In those cases, ownership becomes shared; we provide the foundation, and they own the responsibility for how it’s customized and applied.”

The strategy has impressed investors. In November 2023, AI21 Labs raised a $208 million Series C round at a $1.4 billion valuation.

Since then, the company has focused on enterprise AI reliability. AI21's Maestro system, released in 2025, is an orchestration platform that reduces hallucinations in LLMs such as GPT-4o and Claude 3.5 Sonnet by up to 50 percent.

Reliability also shapes the startup's model ownership vision. “The next phase of ownership and stewardship won’t just be about who trains the biggest model, but who can make those models reliable enough for mass deployment,” Morag said.

The ownership options are expanding. EleutherAI releases everything. Sentient distributes control. AI21 selectively opens. The diversity may be the point.

Regulatory regimes may favor bespoke models. Contributor-driven networks could benefit from blockchain, while hybrid models balance open access and security. 

Back at EleutherAI, Biderman is keeping an open mind. “Part of believing in individual ownership is believing people can set the rules for the stuff they build — even if I don’t like them.”