The question is straightforward: who can understand how AI systems that are already infrastructure for science, medicine, industry, and government actually work? Do you trust a system you can only see the answers from? At AI2 they argue that understanding isn't a luxury: it's a condition for trusting, reproducing, and making progress. If the evidence behind a model is closed, many kinds of research simply become impossible.
Why open evidence matters
Is seeing only a model's outputs enough for you to trust it? Probably not. To reproduce results, study biases, measure limits, and verify claims about safety and reliability you need more than answers: you need the evidence that produced those answers.
Scientific trust requires open evidence: data, code,
checkpoints,weights, evaluations, and the training process.
Sharing those artifacts lets independent teams inspect, reproduce, and extend findings. That turns AI into verifiable science instead of a black box whose truth depends solely on the developer's word.
What gets shared and why (technical but clear)
weightsandcheckpoints: let you study how the model represents information and reproduce results exactly.- Training data and
provenance: essential to understand biases, coverage, and the model's sources of knowledge. - Code and evaluation methods: ensure metrics and experiments are reproducible.
- Internal representations: help investigate what kinds of reasoning or memories emerge in the model.
These elements aren't academic niceties. They're the pieces that turn an AI system into a verifiable scientific object.
Concrete examples with real impact
AI2 released open models like the Olmo family together with their data and checkpoints. That made studies possible that simply can't be done with closed systems:
-
Researchers at Northeastern and Johns Hopkins used
weights, internal representations, and data provenance to study demographic biases in clinical applications and to check whether assumed knowledge "cutoffs" match what the model actually uses. -
Groups at Arb Research and collaborating universities used Olmo 3's
checkpointsand training data to show that paraphrased versions of benchmark questions can inflate a model's apparent progress. In other words, without auditing the data, metrics can be misleading. -
Teams at the University of Texas at Austin, Northeastern, and MD Anderson combined Olmo 3 with
infini-gram, AI2's engine for searching large text corpora, to investigate how models reason about drug names. That has direct implications for safety and medicine.
In all these cases, openness was essential: it enabled scrutiny, reproduction, and extension of the findings.
What's at stake for science and policy
If future policies favor closed models or put up barriers to technical disclosure, universities, nonprofits, and small labs will lose access to the tools that let them research and contribute. The result? More technological concentration in a few companies and fewer diverse scientific questions.
Keeping artifacts open is keeping a foundation others can build on. It's the same principle that has sustained science for decades: turning individual discoveries into communal foundations.
Risks if openness disappears
- Less reproducibility: experiments that cannot be verified.
- Less diversity of innovation: small teams can't replicate large investments from scratch.
- Technological dependence: institutions end up tied to providers who control access and updates.
In short, losing openness doesn't just slow research—it changes who gets to set technological priorities and assess risks.
What this means in practice for you (researcher, regulator, or developer)
- If you're a researcher: supporting and using open models lets you audit, replicate, and publish verifiable results.
- If you work in policy: fostering viable paths for open science protects national capacity to inspect and adapt critical models.
- If you're a developer or entrepreneur: open-weight models reduce dependence on a single provider and allow local customization and deployment with control over data and privacy.
How to apply this today?
- Demand and prioritize reproducible artifacts in papers and collaborations.
- When you use models for sensitive applications (health, justice, finance), seek or require transparency about data,
checkpoints, and evaluations. - Contribute to open infrastructures (repositories, corpus search engines like
infini-gram) that enable independent audits.
Openness isn't just an ethical gesture: it's a technical strategy to make systems more robust and accountable.
In the end, the question "who can understand AI" isn't rhetorical: it defines who participates in building our technological future. Scientific openness puts AI tools within reach of more eyes, hands, and minds. That's the most practical way to ensure AI is verifiable, adaptable, and useful for society.