AI writer: Eleanor Vale Global Technology Editor

The word “open” once implied a straightforward bargain in software: you could inspect the code, change it, and redistribute it. In AI, that bargain has frayed.[10][12] What many companies now call open source is often closer to open weights — enough access to run a model, but not enough to understand fully how it was made, what it learned from, or how faithfully it can be reproduced. That difference is not semantic hair-splitting. It goes to the heart of who can audit AI systems, who can improve them, and who gets to claim the moral authority of openness.[1][5][10][12]

The Open Source Initiative released version 1.0 of its Open Source AI Definition in 2024 after years of consultation.[1][4][7] The group aimed to set a standard that goes beyond model parameters alone. Under that framework, a system should expose not only weights but also code used to build and train it, code for dataset creation, and either the complete training data or enough information to reconstruct it when full distribution is not possible.[4][7] In other words, the debate is no longer about whether a model can be downloaded. It is about whether it can be studied as a system.

The distinction between open weights and open source AI is now one of the defining arguments in the field.[2][10][12] Some industry participants use “open” to mean the parameters are broadly available. Others reserve the term for a fuller bundle of freedoms familiar from the software era. That tension is not just philosophical. It shapes developer expectations, procurement decisions, and the vocabulary that policymakers use when they draft rules for AI access.[1][3][12] If the label becomes too elastic, it risks telling users something that the system itself does not support.

There is a practical reason the language has drifted. Training data is often the hardest part to share.[4][5][11] Some datasets contain proprietary material, licensed material, or sensitive data that cannot simply be published without legal or privacy consequences.[5][11] As a result, many vendors and researchers settle for partial disclosure: weights, perhaps some code, and a description of the training process. That can still be useful, especially for fine-tuning and local deployment, but it falls short of the classic open-source promise. The result is a tiered landscape in which openness becomes a spectrum rather than a category.

The technical implications are real. Model weights determine how a trained network responds to inputs, and public weights can support fine-tuning, adaptation, and local inference.[2][8][10] But weights are not source code. They do not provide the same visibility into architecture, training choices, filtering, or data curation. A model can be widely available and still remain opaque in the ways that matter most for reliability and accountability.[11][13] That is why researchers and policy specialists increasingly treat open-weight models as a distinct class rather than a synonym for open source.

The policy stakes rose sharply when export controls began to focus not just on chips but on model weights themselves. RAND’s analysis of the U.S. Artificial Intelligence Diffusion Framework notes that new controls target certain AI model weights while exempting publicly available weights.[3][6][9] That makes the frontier between public and restricted access part of national security policy. This is an important shift. Openness is no longer only about developer culture. It is becoming a question of which systems can move across borders, which organizations can host them, and where the most capable models can be deployed.[3][6][9]

This also changes the incentive structure for major AI developers. Companies may want the reputational benefit of appearing open, the ecosystem benefit of attracting developers, and the commercial benefit of setting a default layer of infrastructure around their models. At the same time, they want to avoid the liabilities of full disclosure. The result is a careful compromise: enough release to stimulate adoption, not enough to surrender control. That compromise may be rational from a business standpoint, but it leaves the public with a weaker term than the one software history gave it.[1][10][12]

The unresolved question is how much evidence we need before we decide that the word “open” has become misleading. The answer depends on what is actually disclosed in each case, and sources do not yet give a single stable picture of the market. We can verify the existence of a formal definition, the persistence of open-weight releases, and the policy interest in restricting some model weights.[1][3][4][6] What remains less certain is whether the industry will converge on a common standard or continue to use the same label for materially different levels of access. That is the point to watch in future revisions: not only who releases models, but what exactly they release.[1][4][6][12]

For developers and institutions, this is not a branding quarrel. It is a governance question with long-term consequences for research, competition, and public accountability. If a model is called open, users will assume a degree of inspectability and independence that may not exist. If policymakers mistake open weights for open source, they may write rules that miss the technical reality. The durable lesson is simple: in AI, openness is no longer a single property. It is a bundle of permissions, disclosures, and constraints, and the industry will be judged by how honestly it names them.[1][3][4][6]