Open vs Closed: The Defining Fault Line of Modern AI

The question is simple should the

The question is simple: should the most powerful AI models be freely available for anyone to download, study, modify, and use? Or should they be controlled by the companies that build them, accessible only through paid APIs? The answer divides the AI industry down the middle. On one side are OpenAI, Anthropic, and Google — the closed-source frontier labs, who argue that controlling access to their models is necessary for safety and commercial viability. On the other side are Meta, Mistral, DeepSeek, and the open-source community, who argue that openness drives innovation, prevents monopoly, and is the only way to ensure that AI benefits everyone, not just a few wealthy companies. The debate is not academic. Its outcome will determine who controls the most important technology of the twenty-first century.

Open vs. Closed AI

The debate over whether frontier AI model weights should be released openly (open weights, sometimes called “open source”) or kept proprietary. Open weights (LLaMA, Mistral, DeepSeek) accelerate research and decentralisation; closed weights (GPT-4, Claude, Gemini) allow safety controls and competitive moats.

The Case for Closed

The closed-source argument is straightforward. Frontier AI models — models at the leading edge of capability, like GPT-4, Claude, and Gemini — are powerful, potentially dangerous tools. If they are released openly, anyone can use them for any purpose, including malicious purposes. A closed-source model, accessible only through an API (Application Programming Interface) — a structured way for one program to talk to another, where the developer sends text and the company sends back the model’s response without the model ever leaving the company’s servers — allows the developer to monitor usage, filter harmful outputs, and revoke access if the model is being misused. An open-source model, downloaded to a user’s own hardware, cannot be monitored, filtered, or revoked.

The specific risks that the closed-source argument focuses on have evolved over time. In the early days of large language models, the concern was primarily about disinformation — the fear that language models could be used to generate fake news, propaganda, or deceptive content at scale. As the models became more capable, the concerns expanded to include the generation of harmful content — instructions for building weapons, conducting cyberattacks, or manipulating people. More recently, the concerns have expanded further to include the use of frontier models in scientific domains — the fear that the models could be used to design biological weapons, discover novel pathogens, or accelerate the development of dangerous technologies.

The closed-source companies — OpenAI, Anthropic, Google — also have commercial reasons to keep their models closed. Their models are their competitive advantage. If they release the model weights — the actual numeric parameters that define a trained neural network, which is what “release” means in the strongest sense — competitors can fine-tune the models, build on them, and compete with the original developers without bearing the cost of training. The closed-source approach protects the investment that went into training the model and preserves the developer’s market position.

The closed-source argument has been influential in policy circles. The EU AI Act (Regulation (EU) 2024/1689, the EU’s binding horizontal AI regulation, entered into force August 1, 2024) establishes special rules for “general-purpose AI models with systemic risk” — the EU’s formal term for what is colloquially called “frontier” models — while granting a limited exception to open-source GPAI models. The US executive order on AI and other regulatory initiatives have included similar provisions. The AI safety community — researchers and organisations concerned about the risks of advanced AI — has generally supported the closed-source approach, arguing that the risks of open release outweigh the benefits.

The Case for Open

The open-source argument is also straightforward. AI is too important to be controlled by a small number of companies. If the most powerful AI models are controlled by OpenAI, Google, and Anthropic, then those companies have enormous power — power to decide who can use AI, what AI can be used for, and how much it costs. This concentration of power is dangerous, regardless of whether the companies are well-intentioned.

The open-source community argues that openness drives innovation. When models are released openly, researchers and developers around the world can study them, modify them, build on them, and compete with them. This produces innovations that the closed-source companies do not produce or do not share. The open-source AI ecosystem — centred on platforms like Hugging Face (the central platform of the open-source AI ecosystem, hosting model weights, datasets, and demos; often called “the GitHub of AI”) — has produced significant innovations in fine-tuning, inference optimisation (techniques for making AI inference faster, cheaper, or more memory-efficient, such as quantisation and distillation), and safety research that have benefited the broader AI community.

The open-source community also argues that openness provides democratic oversight. When models are closed, only the developer knows how they work, what their limitations are, and what biases they contain. When models are open, researchers, journalists, and civil society organisations can study them, identify biases and harms, and hold the developers accountable. The closed-source approach, the argument goes, makes AI less safe, not more — because the models are less scrutinised.

The economic argument for openness is also significant. Open-source models allow startups, small businesses, and developers in lower-income countries to build AI-powered products without paying the high fees charged by the frontier companies’ APIs. This spreads the economic benefits of AI more widely, rather than concentrating them in the companies that can afford to train frontier models.

The LLaMA Watershed

The open-source AI movement was transformed by a single event: the leak of Meta’s LLaMA model weights in March 2023. LLaMA, Meta’s large language model, had been announced on February 24, 2023 and distributed to selected researchers under a non-commercial research license. Someone — the identity is not publicly known — uploaded the weights to 4chan in early March 2023, and from there they spread across the internet. Within hours, the weights were available on Hugging Face, GitHub, and cloud storage services around the world.

Why the LLaMA leak mattered

Before the leak, the open-source AI community was building on smaller, older, or less capable models. LLaMA was a state-of-the-art model that the community could now study, modify, and run. The leak demonstrated that the open-source community could take a frontier-class language model and build on it — fine-tuning it for instruction-following, code generation, and multilingual tasks; optimising it for consumer hardware; running it on laptops — with a speed and creativity that the closed-source companies could not match. Within weeks, researchers had produced instruction-tuned LLaMA variants. Within months, tools for running LLaMA locally were widely available. The leak proved that open-source AI was a viable competitor to closed-source — and it forced Meta to decide whether to fight the openness or embrace it.

Meta’s response was unusual. Rather than attempting to enforce its license or suppress the leaked weights, the company embraced the open-source approach. On July 18, 2023, Meta released LLaMA 2 under the Llama 2 Community License, which permits most commercial use (with a carve-out: entities with more than 700 million monthly active users require a separate license from Meta). The license is not OSI-approved “open source” in the strict sense — but it is permissive enough that most developers and businesses can use LLaMA 2 freely. The decision was strategic: Meta’s core business — advertising on social media — does not depend on selling AI models, so the company can afford to give its models away. (As Zuckerberg himself put it: “selling access to AI models isn’t our business model.”) And by giving them away, Meta positioned itself as the leader of the open-source AI movement, influencing the direction of the broader ecosystem without having to compete directly with OpenAI and Google in the frontier-model market.

The LLaMA releases were followed by other open-source models: Mistral, from the French company Mistral AI; DeepSeek, from the Chinese company of the same name; and many others. Mistral AI — the French AI lab founded in April 2023 by Arthur Mensch (formerly of Google DeepMind), Guillaume Lample (formerly of Meta), and Timothée Lacroix (formerly of Meta) — is Europe’s frontier open-weight company. The open-source AI ecosystem grew rapidly, and the capabilities of open-source models improved — narrowing the gap with the closed frontier models.

The DeepSeek-R1 release on January 20, 2025 was the most dramatic demonstration of the power of the open-source approach. R1, released under the MIT license — a permissive open-source license allowing essentially any use (commercial, modification, distribution) with minimal restrictions — matched or exceeded the performance of OpenAI’s o1 on many reasoning benchmarks, including AIME, MATH-500, and SWE-bench Verified. Its release triggered a market panic: on January 27, 2025, NVIDIA shares fell approximately 17%, erasing roughly $589 billion in market capitalisation — the largest single-day loss in US stock-market history. The rout intensified the geopolitical competition over AI and demonstrated that open-source AI was not just a niche movement — it was a serious competitor to the closed-source frontier.

The Safety Debate

The open-vs-closed debate is, at its core, a debate about safety. The closed-source companies argue that open release is dangerous. The open-source community argues that closed release is dangerous — because it concentrates power, reduces scrutiny, and prevents the broad community from identifying and addressing safety issues.

The safety argument against open release is that once model weights are available, anyone can use the model for any purpose, and the developer has no ability to intervene. A malicious actor could fine-tune the model to remove safety features, elicit harmful capabilities, or produce content that the developer had tried to suppress. The model could be used to generate disinformation, write malicious code, or facilitate cyberattacks — and the developer could do nothing to stop it.

The safety argument for open release is that closed models are less scrutinised, and therefore less safe. When a model is closed, only the developer knows how it works, what its limitations are, and what biases it contains. Independent researchers cannot study the model, cannot identify its failure modes, and cannot hold the developer accountable. When a model is open, the entire research community can study it, and the collective scrutiny produces a better understanding of the model’s capabilities and risks.

The “Brussels effect” and open-source AI

The Brussels effect — a term coined by Columbia Law School professor Anu Bradford in 2012 to describe how EU regulations effectively become global standards by virtue of the EU’s market size — is relevant to the open-vs-closed debate. The EU AI Act treats “general-purpose AI models with systemic risk” differently from open-source GPAI models, granting open-source models a limited exception from certain transparency obligations (copyright obligations apply to all GPAI models without exception). Because multinational AI companies tend to comply with EU rules worldwide rather than forgo access to the EU single market, the EU’s regulatory distinction between frontier and open-source models may effectively become the global standard — shaping the open-vs-closed landscape even outside Europe.

The truth is that both arguments have merit. Open release does enable misuse — this is not a hypothetical concern; it has happened. Closed release does reduce scrutiny — this is also not hypothetical; the closed-source companies have been criticised for a lack of transparency about their models’ capabilities and limitations. The question is not which approach is safer in absolute terms, but which approach produces the best balance of innovation, safety, and democratic accountability.

What the Debate Means

The open-vs-closed debate is not going to be resolved. The two strategies — open and closed — are both viable, both have their proponents, and both have their advantages and disadvantages. The AI industry will likely remain split between open and closed providers for the foreseeable future.

The interaction between the two strategies is complex. The open-source ecosystem benefits from the closed frontier companies’ research — the open-source models are often derived from or inspired by the closed models, and the open-source community learns from the closed companies’ published research. The closed frontier companies benefit from the open-source ecosystem — the ecosystem produces innovations that the closed companies can adopt, and it provides a pool of talent that the closed companies can recruit from.

The competition between the two strategies is, in some respects, beneficial for the broader development of AI. The competition drives innovation — both in the open and closed models. It provides alternatives — users who are concerned about the concentration of power in the closed companies can choose open alternatives. And it provides a check on the excesses of either strategy — the open-source ecosystem prevents the closed companies from becoming too dominant, and the closed companies’ safety practices provide a benchmark against which the open-source ecosystem’s practices can be evaluated.

The question of whether AI should be open or closed is not a question that will be answered once and for all. It is a question that will be asked repeatedly, in different contexts, by different actors, with different answers. The asking of the question — and the contestation that the asking produces — is itself a valuable feature of the AI era.