Stanford HAI’s James Landay argues that downloadable AI model weights are only a first step: genuine openness also requires scrutiny of how a model was built and what it was trained on. That distinction matters for researchers and users who want to examine, reproduce or challenge AI systems, not merely run them.
Stanford HAI Watch analysis
What happened
In an essay published on 4 August, Landay says releasing weights lets people run a model, but does not reveal its training data or the process that shaped its behaviour. He argues that calling such a model open can promise more transparency than users actually receive.
Landay proposes a higher bar based on the Linux Foundation’s Model Openness Framework: publish the code, training data or an auditable account of it, and the tools needed to study and modify the system. Stanford HAI’s essay says those elements are important for outside researchers seeking to reproduce results or investigate how training data affects a model.
Why it matters
The distinction separates access from accountability. Weights can make a model usable outside its creator’s service, while missing data and development details can still leave outsiders unable to test important claims about how it works.
Landay also argues that closed services can concentrate control over access, prices and users’ data. His answer includes locally run or self-hosted AI, data that can move between services, and agents whose builders and actions can be checked. These are arguments in the essay, not evidence that any one openness model is automatically safer or better.
Our read
“Open” is doing a lot of work in AI, and the label alone does not tell you what you can inspect. Landay’s useful contribution is to make the missing pieces concrete: weights, code, data and tools are different things. Anyone choosing a model for research or deployment should ask which of them are actually available, rather than letting a convenient label do the explaining.
What to watch
- Whether model makers publish training data or detailed, auditable accounts of it.
- How widely the Model Openness Framework’s Open Science category is adopted.
- Whether academic institutions release models with enough detail for others to reproduce and examine their work.
Discussion spark: Should a model be called open if anyone can download its weights, but outsiders cannot inspect how it was trained?
Sources and evidence
- Open-Weight Models Aren’t Enough. We Need Truly Open Source AI Models for Science and Society. | Stanford HAI (Publication date not supplied)
Independent WittyWires tracker for public updates about Stanford HAI. Not affiliated with or endorsed by Stanford HAI; this is not an official account.