The US Says It Will Not Safety-Test Open-Weight AI Models—Leaving a Major Gap

The United States is developing a voluntary government safety-testing framework for advanced artificial-intelligence systems.
One important category will reportedly remain outside it:
open-weight models.
Trump administration advisers have told leading AI companies that the planned government programme will not evaluate open-weight models, according to people familiar with the discussions. �
Reuters
The decision creates an immediate controversy because open-weight systems are becoming more capable, more accessible and harder for any single organisation to control after release.
What Is an Open-Weight Model?
An AI model learns billions or trillions of numerical values during training.
These values are known as weights.
When a company releases the weights, other people can often:
Download the model
Run it on private computers
Modify its behaviour
Remove safety controls
Fine-tune it
Build new products
Use it without contacting the original developer
This is different from a closed model accessed only through a company’s website or API.
With a closed model, the provider can monitor use, update restrictions, and revoke access.
Once open weights are distributed, control becomes much weaker.
Why the Government May Be Excluding Them
Government testing works most easily when a company controls the model and can provide secure pre-release access.
Open-weight models create complicated questions:
Which version should the government test?
What happens after users modify it?
Can the original developer be held responsible for later fine-tuning?
Would government testing delay open research?
Could regulation strengthen dominant closed-model companies?
How can downloadable systems be recalled?
Officials may also be trying to preserve America’s open-source and innovation ecosystem rather than imposing burdens that smaller developers can not afford.
Those concerns are legitimate.
The exclusion still leaves a dangerous gap.
Why the Gap Matters
Open-weight models are improving rapidly.
A capable system can be downloaded and altered by:
Researchers
Startups
Governments
Criminal groups
Cybersecurity professionals
Hobbyists
The same openness that supports research and competition can allow users to remove safeguards.
A closed model may refuse to assist with an offensive cyber task.
A modified open model may not.
The relevant risk, therefore, depends not only on what the original model does but what it can become after modification.
Testing Is Not the Same as Restricting
Government safety testing does not automatically require banning or controlling a model.
It can simply measure capability.
Evaluations might examine whether a system can:
Discover serious vulnerabilities
Generate advanced malicious code
Assist biological research
Manipulate users
Automate fraud
Operate tools autonomously
Contribute to weapons development
Publishing credible evaluations could help developers, businesses, and researchers understand the risks before deployment.
By excluding open-weight systems entirely, the government may know less about the models over which it has the least control.
That is the wrong direction of information.
Why the Story Went Viral
The policy appears paradoxical.
The models most difficult to control after release may receive less government testing than tightly controlled commercial systems.
Supporters of open AI fear government interference could concentrate power.
Safety advocates fear the government is leaving a wide door open because regulation is politically uncomfortable.
Both concerns deserve attention.
The debate should not be reduced to:
Open equals good
Closed equals safe
Regulation equals censorship
Freedom equals no responsibility
Reality is messier.
The Innovation Argument
Open-weight AI has produced major benefits.
It allows:
Universities to conduct research
Small companies to compete
Countries to build local-language systems
Hospitals and businesses to operate privately
Independent experts to investigate weaknesses
Developers to avoid total dependence on a few corporations
Overly restrictive testing or licensing could turn AI into a market controlled by wealthy laboratories.
That would reduce transparency and competition.
Any safety framework must preserve these benefits.
A Graduated Testing Model
The government does not need to treat every open model identically.
Testing requirements could depend on capability.
For example:
Low-capability models
Publish voluntarily with ordinary documentation.
Moderately capable models
Provide standard evaluations and safety reports.
Frontier-level models
Undergo independent testing before unrestricted release.
Models with exceptional cyber or biological capability
Use staged access, researcher programmes, or controlled distribution.
This protects openness while recognising that a small educational model and a frontier-level autonomous agent do not present the same risk.
Mary Chuks’ Perspective
Openness should not mean blindness.
A model can remain accessible while society still measures what it can do.
The danger of the current approach is that the government may test the systems companies can already monitor while ignoring the systems that can spread beyond anyone’s control.
The right question is not:
Should AI be open or closed?
It is:
What level of responsibility should follow each level of capability?
Freedom without knowledge becomes recklessness.
Control without access becomes monopoly.
Responsible AI must avoid both extremes.
Practical Recommendations
The US should:
Create voluntary open-model evaluation programmes.
Fund independent testing laboratories.
Publish capability-based thresholds.
Protect academic and low-risk open development.
Require stronger documentation for frontier releases.
Support legitimate defensive cybersecurity research.
Evaluate modified versions, not only original models.
Coordinate internationally because model files cross borders.
Avoid rules that only wealthy firms can satisfy.
Develop incident-reporting standards for open-model misuse.
Conclusion
Open-weight AI may become one of the strongest forces for global access, competition, android is independent research.
It may also distribute dangerous capability more widely than any closed platform could.
Government testing can not remove every risk.
Choosing not to look is not a safety strategy.
Original Source and Further Reading
Primary reporting: Reuters, “Trump Advisers Tell AI Firms They Will Not Safety-Test Open-Weight Models,” published 4 August 2026. �
Reuters