“The open source debate in AI is not really about open source. It is about who controls the most transformative technology in human history and under what terms. Every other argument is downstream of that one.”
Neal Lloyd · Inside The Machine, Day 14In March 2023, Meta released the weights for its LLaMA language model. The release was, technically, a research preview with a licence that restricted commercial use. Within days, the weights had been leaked to 4chan and were proliferating across the internet at a pace that made the licence agreement functionally irrelevant. The genie, as several commentators observed at the time, was out of the bottle. Meta subsequently leaned into this reality — releasing Llama 2, Llama 3, and Llama 4 with progressively more permissive licences — and in doing so ignited the most consequential ideological battle in the history of artificial intelligence. The question it poses is deceptively simple: should the most powerful AI models be free for anyone to use, modify, and deploy? The answer, depending on who you ask, is either obviously yes or obviously no. And both answers are, in their own way, completely correct.
The Definition Is More Complicated Than the Debate Assumes
Before the argument can be had honestly, the terminology needs to be held carefully. “Open source AI” is used to describe several different things that are not the same thing. At minimum, a model being “open” might mean releasing the weights — the numerical parameters that encode what the model has learned — so that anyone can download and run the model locally. It might additionally mean releasing the training code — the software used to train the model. It might additionally mean releasing the training data — the dataset the model was trained on. It might mean releasing the full technical specification of the training process including hyperparameters, hardware configuration, and compute budget. And it might mean making all of this available under a licence that permits unrestricted commercial use, modification, and redistribution.
Almost no “open” AI model does all of these things. Meta’s Llama 4 releases the weights under a relatively permissive licence but does not release the training data. Mistral releases weights but not full training details. DeepSeek released weights for its R1 model, which produced a global shockwave in January 2026 when the AI community discovered it had achieved reasoning capabilities comparable to OpenAI’s o1 at a fraction of the reported training cost — but DeepSeek is a Chinese company whose model’s full provenance is not independently auditable. The open source label covers a spectrum from “you can download and run this” to “you can reproduce, verify, modify, and redistribute everything about how this was made.” The debate often conflates the two ends of that spectrum, which makes it considerably more heated and considerably less illuminating than it needs to be.
Fully open: weights, training code, training data, full methodology, permissive licence. Examples: almost none at frontier scale. Partially open (weights-only): download and run, no training data, often restricted licence. Examples: Meta Llama 4, Mistral, DeepSeek R1. Closed source: API access only, no weights, no training details. Examples: GPT-5.5, Claude Opus 4, Gemini Ultra. Open weights with safety restrictions: weights released but with use-case prohibitions. Examples: several Llama variants. The policy debate almost exclusively concerns frontier-scale models — systems capable of complex reasoning, code generation, and autonomous action. Below that frontier, the open source ecosystem is vast, robust, and largely uncontroversial.
Why the Open Source Argument Is Genuinely Compelling
Auditability and accountability. Closed models are black boxes in a more profound sense than the technical literature’s use of that term. Not only are the internal mechanisms opaque — that is true of open models too — but the training data, the fine-tuning choices, the safety interventions, and the commercial incentives that shaped the model are entirely invisible to the public. When a closed model produces a biased output, discriminates against a demographic, or gives dangerous advice, there is no independent path to understanding why. Open weights at least allow the research community to probe the model’s behaviour systematically, to identify failure modes, and to hold developers accountable for what their systems actually do.
Concentration of power. If the only competitive AI models are produced by three or four large companies and accessible only through their APIs, those companies have effective control over a capability that increasingly determines how knowledge is produced, how decisions are made, and how work is done across every sector of the economy. This is not a theoretical concern. It is a description of the current situation. The argument that this concentration is acceptable because the companies involved have good intentions is precisely the argument that the history of platform monopolies has taught us to be most sceptical of.
Innovation and access. The open source software ecosystem produced the internet’s infrastructure: Linux, Apache, PostgreSQL, Python. The thesis is that the same dynamic — many independent actors building on a shared foundation — produces better outcomes than proprietary development. DeepSeek’s R1 release in January 2026, which demonstrated frontier-level reasoning at a fraction of the training cost previously assumed necessary, was a direct product of open research building on prior open releases. The efficiency gains it revealed have been adopted by closed developers too. The entire field moves faster when research is shared.
The argument that powerful AI should remain in the hands of a small number of well-intentioned American companies is, at its core, an argument that concentrated power is acceptable when the people holding it have good values. History has a consistent view on this argument. It is not encouraging.Neal Lloyd · Inside The Machine, Day 14
Why the Safety Argument Is Also Genuinely Compelling
Irreversibility. Once model weights are released publicly, they cannot be recalled. If a vulnerability is discovered, if a failure mode is identified, if the model is found to enable a class of harms that were not anticipated at release — the weights are already on hard drives across the world. Pharmaceutical companies cannot un-release a drug that turns out to be dangerous; they can recall it from shelves. Software companies cannot un-release source code that turns out to contain a catastrophic vulnerability; they can patch it. AI companies cannot un-release model weights that turn out to enable bioweapon synthesis, sophisticated cyberattacks, or mass-scale influence operations. The asymmetry between release and recall is a genuine and serious argument for caution.
Capability concentration at the frontier. The debate over open-sourcing AI models looks different at different capability levels. Releasing the weights for a model that can write essays and answer questions is a different decision from releasing the weights for a model that Anthropic’s president describes as “very good at cyber warfare.” Claude Mythos Preview is not being released publicly, and Anthropic has been explicit about why: the cybersecurity capabilities are powerful enough to cause serious damage without appropriate safeguards. This is not a hypothetical concern. As models become more capable at precisely the tasks — autonomous action, complex reasoning, technical problem-solving — that make them most useful, they also become more useful for the full range of human purposes, which includes purposes that most people would prefer remain difficult.
The responsibility gap. Closed models are developed, deployed, and monitored by organisations that have legal accountability for their outputs. When something goes wrong, there is an identifiable responsible party. Open weights, once released, can be fine-tuned by anyone, deployed by anyone, and used for any purpose. The organisation that released the original weights may bear moral responsibility for foreseeable misuse, but they bear limited practical ability to prevent it. The argument that this diffuses accountability is correct. The question is whether diffused accountability is better or worse than concentrated accountability in the hands of a small number of commercial entities with their own interests.
Meta has a compelling business reason to advocate for open AI: it cannot compete with OpenAI and Anthropic at closed frontier models, but it can compete — and has competitive advantage — in an open ecosystem where its distribution and infrastructure matter more than model exclusivity. OpenAI and Anthropic have a compelling business reason to advocate for closed models: their entire value proposition depends on the weights being proprietary. The safety arguments on both sides are genuine. They are also conveniently aligned with the financial interests of the companies making them. Holding both of these things simultaneously is essential to thinking clearly about this debate.
The Binary Is False. The Question Is Where to Draw Lines That Actually Hold.
The framing of “open versus closed” is useful for generating heat and not particularly useful for generating light. The real question is not whether AI should be open or closed but where, on the spectrum of capability and risk, the default should shift from one to the other — and who gets to make that call, accountable to whom, under what process.
A reasonable position might look something like this: models below a certain capability threshold — where the marginal uplift to a bad actor is low and the benefits to the research community and independent developers are high — should default toward openness. Models above that threshold, where the capabilities include meaningful uplift toward weapons of mass destruction, sophisticated autonomous cyberattacks, or large-scale manipulation, should be subject to structured access regimes rather than either full public release or opaque proprietary deployment.
Structured access — the ability for vetted researchers, auditors, and regulators to inspect model capabilities and behaviour without full public weight release — is an approach that several researchers have proposed and that the EU AI Act gestures toward without fully operationalising. It is not perfect. It requires trusted intermediaries, robust audit processes, and enforcement mechanisms that do not yet exist at scale. It is considerably better than the current situation, in which the most powerful models are either fully closed and commercially controlled or fully open and uncontrollable once released.
The capability threshold for structured access is not fixed. What required the resources of a national laboratory five years ago can now be approximated by a well-resourced bad actor with access to open weights and cloud compute. The line needs to move with the technology — which means the institutions responsible for drawing it need to be faster, more technically sophisticated, and more internationally coordinated than anything that currently exists. That gap between what the situation requires and what our institutions can provide is, in a sentence, the central governance challenge of the AI era.
The open source debate will not be resolved by argument. It will be resolved by an incident — a specific, attributable harm enabled by openly released model weights that is significant enough to force a policy response. The question is what kind of incident, and whether the response will be proportionate or panic.Neal Lloyd · Inside The Machine, Day 14
Inside The Machine, Day 14 · June 2026
Neal Lloyd writes about technology, human adaptation, and the uncomfortable questions nobody wants to answer at dinner. Inside The Machine is his ongoing daily series on AI.
- Day 01What Is This Thing?
- Day 02Survive the Machine
- Day 03The Great Debate
- Day 04Who Gets Hurt?
- Day 05Who’s In Charge?
- Day 06The Industries That Win
- Day 07The Human Edge
- Day 08The Creativity Question
- Day 09Does AI Feel Anything?
- Day 10The Data Problem
- Day 11The Trust Question
- Day 12The Accountability Gap
- Day 13The Rewired Brain
- Day 14Open vs ClosedYou are here
