Suddenly, “open source” is the newest buzzword in AI circles. Meta has pledged to create open-source artificial general intelligence. And Elon Musk is suing OpenAI over its lack of open-source AI models.

Meanwhile, a growing variety of tech leaders and corporations are setting themselves up as open-source champions. 

But there’s a fundamental problem—nobody can agree on what “open-source AI” means. 

On the face of it, open-source AI guarantees a future where anyone can participate within the technology’s development. That might speed up innovation, boost transparency, and provides users greater control over systems that might soon reshape many points of our lives. But what even is it? What makes an AI model open source, and what disqualifies it?

The answers could have significant ramifications for the long run of the technology. Until the tech industry has settled on a definition, powerful firms can easily bend the concept to suit their very own needs, and it could grow to be a tool to entrench the dominance of today’s leading players.

Entering this fray is the Open Source Initiative (OSI), the self-appointed arbiters of what it means to be open source. Founded in 1998, the nonprofit is the custodian of the Open Source Definition, a widely accepted algorithm that determine whether a bit of software might be considered open source. 

Now, the organization has assembled a 70-strong group of researchers, lawyers, policymakers, activists, and representatives from big tech firms like Meta, Google, and Amazon to give you a working definition of open-source AI. 

The open-source community is an enormous tent though, running the gamut from hacktivists to Fortune 500 firms. While there’s broad agreement on the overarching principles, says Stefano Maffulli, OSI’s executive director, it’s becoming increasingly obvious that the devil is in the small print. With so many competing interests to think about, finding an answer that satisfies everyone while ensuring the most important firms play along isn’t any easy task.

Fuzzy criteria

The shortage of a settled definition has done little to stop tech firms from adopting the term.

Last July, Meta made its Llama 2 model, which it known as open source, freely available, and it has a track record of publicly releasing AI technologies. “We support the OSI’s effort to define open-source AI and sit up for continuing to take part in their process for the advantage of the open source community the world over,” Jonathan Torres, Meta’s associate general counsel for AI, open source, and licensing told us. 

That stands in marked contrast to rival OpenAI, which has shared progressively fewer details about its leading models over time, citing safety concerns. “We only open-source powerful AI models once we’ve got fastidiously weighed the advantages and risks, including misuse and acceleration,” a spokesperson said. 

Other leading AI firms, like Stability AI and Aleph Alpha, have also released models described as open source, and Hugging Face hosts a big library of freely available AI models.

While Google has taken a more locked-down approach with its strongest models, like Gemini and PaLM 2, the Gemma models released last month are freely accessible and designed to go toe-to-toe with Llama 2, though the corporate described them as “open” moderately than “open source.”  

But there’s considerable disagreement about whether any of those models can really be described as open source. For a start, each Llama 2 and Gemma include licenses that restrict what users can do with the models. That’s anathema to open-source principles: one in every of the important thing clauses of the Open Source Definition outlaws the imposition of any restrictions based on use cases.

The standards are fuzzy even for models that don’t include these sorts of conditions. The concept of open source was devised to make sure developers could use, study, modify, and share software without restrictions. But AI works in fundamentally other ways, and key concepts don’t translate from software to AI neatly, says Maffulli.

One in every of the most important hurdles is the sheer variety of ingredients that go into today’s AI models. All it’s good to tinker with a bit of software is the underlying source code, says Maffulli. But depending in your goal, dabbling with an AI model could require access to the trained model, its training data, the code used to preprocess this data, the code governing the training process, the underlying architecture of the model, or a bunch of other, more subtle details.

Which ingredients it’s good to meaningfully study and modify models stays open to interpretation. “We have now identified what basic freedoms or basic rights we wish to give you the option to exercise,” says Maffulli. “The mechanics of easy methods to exercise those rights should not clear.”

Settling this debate shall be essential if the AI community desires to reap the identical advantages software developers gained from open source, says Maffulli, which was built on broad consensus about what the term meant. “Having [a definition] that’s respected and adopted by a big chunk of the industry provides clarity,” he says. “And with clarity comes lower costs for compliance, less friction, shared understanding.”

By far the most important sticking point is data. All the foremost AI firms have simply released pretrained models, without the datasets on which they were trained. For people pushing for a stricter definition of open-source AI, Maffulli says, this seriously constrains efforts to change and study models, routinely disqualifying them as open source.

Others have argued that a straightforward description of the information is commonly enough to probe a model, says Maffulli, and also you don’t necessarily must retrain from scratch to make modifications. Pretrained models are routinely adapted through a process referred to as fine-tuning, by which they’re partially retrained on a smaller, often application-specific, dataset.

Meta’s Llama 2 is a living proof, says Roman Shaposhnik, CEO of open-source AI company Ainekko and vice chairman of legal affairs for the Apache Software Foundation, who’s involved within the OSI process. While Meta only released a pretrained model, a flourishing community of developers has been downloading and adapting it, and sharing their modifications.

“Persons are using it in all forms of projects. There’s a complete ecosystem around it,” he says. “We subsequently must call it something. Is it half-open? Is it ajar?”

While it might be technically possible to change a model without its original training data, restricting access to a key ingredient shouldn’t be really within the spirit of open source, says Zuzanna Warso, director of research at nonprofit Open Future, who’s participating within the OSI’s discussions. It’s also debatable whether it’s possible to really exercise the liberty to review a model without knowing what information it was trained on.

“It’s a vital component of this whole process,” she says. “If we care about openness, we must always also care in regards to the openness of the information.”

Have your cake and eat it

It’s necessary to grasp why firms setting themselves up as open-source champions are reluctant at hand over training data. Access to high-quality training data is a significant bottleneck for AI research and a competitive advantage for larger firms that they’re desperate to maintain, says Warso.

At the identical time, open source carries a bunch of advantages that these firms would love to see translated to AI. At a superficial level, the term “open source” carries positive connotations for a number of people, so engaging in so-called “open washing” might be a simple PR win, says Warso.

It may even have a big impact on their bottom line. Economists at Harvard Business School recently found that open-source software has saved firms almost $9 trillion in development costs by allowing them to construct their products on top of high-quality free software moderately than writing it themselves.

For larger firms, open-sourcing their software in order that it could actually be reused and modified by other developers will help construct a strong ecosystem around their products, says Warso. The classic example is Google’s open-sourcing of its Android mobile operating system, which cemented its dominant position at the guts of the smartphone revolution. Meta’s Mark Zuckerberg has been explicit about this motivation in earnings calls, saying “open source software often becomes an industry standard, and when firms standardize on constructing with our stack, that then becomes easier to integrate latest innovations into our products.”

Crucially, it also appears that open-source AI may receive favorable regulatory treatment in some places, Warso says, pointing to the EU’s newly passed AI Act, which exempts certain open-source projects from a few of its more stringent requirements.

Taken together, it’s clear why sharing pretrained models but restricting access to the information required to construct them makes good business sense, says Warso. But it surely does smack of firms attempting to have their cake and eat it too, she adds. And if the strategy helps entrench the already dominant positions of enormous tech firms, it’s hard to see how that matches with the underlying ethos of open source.

“We see openness as one in every of the tools to challenge the concentration of power,” says Warso. “If the definition is speculated to assist in difficult these concentrations of power, then the query of knowledge becomes much more necessary.”

Shaposhnik thinks a compromise is feasible. A major amount of knowledge used to coach the most important models already comes from open repositories like Wikipedia or Common Crawl, which scrapes data from the net and shares it freely. Corporations could simply share the open resources used to coach their models, he says, making it possible to recreate an inexpensive approximation that ought to allow people to review and understand models.

The shortage of clarity regarding whether training on art or writing scraped from the web infringes on the creator’s property rights may cause legal complications though, says Aviya Skowron, head of policy and ethics at nonprofit AI research group EleutherAI, also involved within the OSI process. That makes developers wary of being open about their data.

Stefano Zacchiroli, a professor of computer science on the Polytechnic Institute of Paris who can also be contributing to the OSI definition, appreciates the necessity for pragmatism. His personal view is that a full description of a model’s training data is the bare minimum for it to be described as open source, but he recognizes that stricter definitions of open-source AI may not have broad appeal.

Ultimately, the community needs to come to a decision what it’s trying to attain, says Zacchiroli: “Are you only following where the market goes in order that they don’t essentially co-opt the term ‘open-source AI,’ or are you attempting to pull the market toward being more open and providing more freedoms to the users?”

What’s the purpose of open source?

It’s debatable how much any definition of open-source AI will level the playing field anyway, says Sarah Myers West, co-executive director of the AI Now Institute. She co-authored a paper published in August 2023 exposing the dearth of openness in lots of open-source AI projects. But it surely also highlighted that the vast amounts of knowledge and computing power needed to coach cutting-edge AI creates deeper structural barriers for smaller players, regardless of how open models are.

Myers West thinks there’s also an absence of clarity regarding what people hope to attain by making AI open source. “Is it safety, is it the flexibility to conduct academic research, is it attempting to foster greater competition?” she asks. “We have to be far more precise about what the goal is, after which how opening up a system changes the pursuit of that goal.”

The OSI seems keen to avoid those conversations. The draft definition mentions autonomy and transparency as key advantages, but Maffulli demurred when pressed to clarify why the OSI values those concepts. The document also accommodates a bit labeled “out of scope issues” that makes clear the definition won’t wade into questions around “ethical, trustworthy, or responsible” AI.

Maffulli says historically the open-source community has focused on enabling the frictionless sharing of software and avoided getting bogged down in debates about what that software ought to be used for. “It’s not our job,” he says.

But those questions can’t be dismissed, says Warso, regardless of how hard people have tried over the a long time. The concept technology is neutral and that topics like ethics are “out of scope” is a myth, she adds. She suspects it’s a myth that should be upheld to stop the open-source community’s loose coalition from fracturing. “I feel people know it’s not real [the myth], but we want this to maneuver forward,” says Warso.

Beyond the OSI, others have taken a special approach. In 2022, a gaggle of researchers introduced Responsible AI Licenses (RAIL), that are just like open-source licenses but include clauses that may restrict specific use cases. The goal, says Danish Contractor, an AI researcher who co-created the license, is to let developers prevent their work from getting used for things they consider inappropriate or unethical.

“As a researcher, I’d hate for my stuff to be utilized in ways that may be detrimental,” he says. And he’s not alone: a recent evaluation he and colleagues conducted on AI startup Hugging Face’s popular model-hosting platform found that 28% of models use RAIL. 

The license Google attached to its Gemma follows an analogous approach. Its terms of use list various prohibited use cases considered “harmful,” which reflects its “commitment to developing AI responsibly,” the corporate said in a recent blog post.The Allen Institute for AI has also developed its own tackle open licensing. Its ImpACT Licenses restrict redistribution of models and data based on their potential risks.

Given how different AI is from conventional software, some level of experimentation with different degrees of openness is inevitable and possibly good for the sphere, says Luis Villa, co-founder and legal lead at open-source software management company Tidelift. But he worries that a proliferation of “open-ish” licenses which might be mutually incompatible could negate the frictionless collaboration that made open source so successful, slowing down innovation in AI, reducing transparency, and making it harder for smaller players to construct on one another’s work.

Ultimately, Villa thinks the community must coalesce around a single standard, otherwise industry will simply ignore it and choose for itself what “open” means. He doesn’t envy the OSI’s job, though. When it got here up with the open-source software definition it had the luxurious of time and little outside scrutiny. Today, AI is firmly within the crosshairs of each big business and regulators.

But when the open-source community can’t decide on a definition, and quickly, another person will give you one which suits their very own needs. “They’re going to fill that vacuum,” says Villa. “Mark Zuckerberg goes to inform us all what he thinks ‘open’ means, and he has a really big megaphone.”