Big Techs defend Open Source AI and call for fewer restrictions in the United States

Big Techs defend Open Source AI and call for fewer restrictions in the United States

A group made up of some of the most influential technology companies has intensified pressure on United States lawmakers to avoid broad restrictions on open artificial intelligence models.

The effort brings together names such as Nvidia, Microsoft, Meta and IBM, along with other companies and organizations connected to the AI ecosystem. In a letter released on July 24, 2026, the signatories argued that rules created too hastily could reduce competition, limit technological development and push part of the innovation to other countries.

The position expands a debate that has already moved beyond technical discussions about licenses and access to code. The openness of AI models has become an economic, regulatory and geopolitical issue, especially as Chinese laboratories advance and the United States and China compete for leadership in the global artificial intelligence market.

At the same time, supporters of stricter rules warn that powerful models, when made available for download and modification, can be used for cyberattacks, disinformation campaigns and other harmful activities.

The discussion, therefore, is not simply about choosing between innovation and safety. The main challenge will be to define which risks need to be controlled, who should be held responsible for misuse and how far a government can go in restricting a technology after its components have already been publicly distributed.

What happened?

The letter was led by Nvidia CEO Jensen Huang and signed by around two dozen companies, groups and organizations. The document asks lawmakers to avoid premature restrictions that could harm competition or move innovation outside the United States.

Among the participants cited are Microsoft, Meta and IBM. The move comes at a time when the United States Congress is evaluating stricter safety measures for advanced artificial intelligence systems.

The companies acknowledge that there are legitimate concerns involving technology theft, malicious use and national security. However, they argue that these problems should be addressed through targeted legal and commercial mechanisms, rather than broad bans on open models.

According to the argument presented in the letter, closed systems are not automatically safer. They can also be hacked, misused or contain flaws that external researchers are unable to identify.

Models with open weights, on the other hand, can be examined by a broader community of developers and specialists, increasing the chances of finding vulnerabilities, creating protections and continuously improving the systems.

Are Open Source AI and Open Weight the same thing?

Although the term “Open Source” is often used in a broad way, there is an important difference between truly open models and models that only make their weights available.

Weights are the mathematical parameters learned during training. When a company publishes these files, developers can run the model on their own infrastructure, fine-tune it and create customized applications.

However, publishing the weights does not necessarily mean that the entire model creation process is open.

According to the Open Source Initiative, a truly Open Source artificial intelligence system must grant the freedom to use, study, modify and share it. This requires access not only to the parameters, but also to the information and code needed to properly understand and modify the system.

An Open Weight model may make the final parameters available without fully revealing:

• the datasets used for training;
• the data preparation code;
• the training process;
• the filters applied;
• the evaluation criteria;
• all technical decisions used to build the system.

For this reason, many models commercially presented as Open Source are, technically, Open Weight models. This distinction should become more important as governments create specific laws for open systems, because different levels of transparency may require different responsibilities.

Why are companies defending open models?

The defense of open models involves technical, economic and strategic interests.

Reducing dependence on vendors

Companies that use closed models usually depend on the APIs, pricing, rules and availability of the provider.

A change in pricing, request limits or usage policy can directly affect applications built on top of that platform.

With open models, an organization can run AI on its own servers or on a cloud infrastructure of its choice. This reduces the risk of excessive dependence on a single provider and allows greater control over the architecture.

Greater control over sensitive data

Sectors such as banking, healthcare, government, defense and industry work with information that cannot always be sent to an external service.

The United States government’s artificial intelligence plan recognizes that open models can benefit organizations that need to keep sensitive data inside controlled environments. The document also highlights the importance of these systems for startups, researchers and academic institutions.

A company can, for example, install a model on its internal network, prevent internet access and fully control the data sent to the system.

This does not eliminate security risks, but it makes it possible to create governance policies that are better aligned with the organization’s needs.

Deeper customization

Open models allow adjustments focused on specific sectors, languages and tasks.

A company can adapt a model to:

• analyze legal documents;
• interpret financial records;
• support customer service teams;
• generate code following internal standards;
• process medical information;
• classify business documents;
• work on devices without permanent connectivity.

This flexibility can be especially relevant for organizations whose needs are not fully met by generic services.

Strengthening research

Researchers need to examine models, reproduce experiments and compare results.

In fully closed systems, changes made by the provider can alter the model’s behavior without the academic community having access to the details of the update.

The availability of weights, documentation, code and training information makes audits, safety studies and the development of new techniques easier.

The United States AI plan considers open models essential for certain academic research and recommends creating an environment that supports their development.

International competition

The discussion is also connected to the technological competition between the United States and China.

Open models can become standards adopted by companies, universities and governments around the world. When developers build tools on top of a specific ecosystem, they increase the technological influence of the organization or country responsible for that platform.

The United States government recognizes that open models have geostrategic value. Its AI plan defends the existence of competitive open models developed within the country and aligned with American interests.

In this context, excessively restricting United States models could create room for technologies produced by foreign laboratories.

What are the risks of open AI models?

The benefits do not eliminate the risks.

Once a model’s weights are made available for download, it becomes extremely difficult to retrieve every copy. Even if the original developer removes the files, versions may continue to be stored on servers, computers and distributed platforms.

It is also possible to modify a model to remove protections that were originally implemented.

Use in cyberattacks

Advanced models can assist with code analysis, vulnerability discovery and the automation of technical tasks.

These capabilities can be used both by security professionals and by criminals. The same technology capable of finding a flaw in a system can help exploit it.

The debate in the United States intensified after incidents involving autonomous agents and concerns about the use of AI in cyber operations. According to Reuters, lawmakers even proposed mandatory interruption mechanisms, known as “kill switches”, for certain systems.

Applying this idea to open models, however, is difficult. A provider can disable a centralized API, but it does not have the same control over thousands of copies running independently.

Disinformation and manipulated content

Models can be adapted to produce text, images, audio and video at scale.

This type of automation can be used to create manipulation campaigns, impersonate people or distribute false information.

The risk does not exist only in open models. Closed services can also be misused. The difference lies in the provider’s ability to block accounts, limit requests and update protections centrally.

Removal of protection mechanisms

Companies that develop models usually implement filters to prevent certain types of responses.

When weights are released, external developers can perform new training or modifications that reduce these restrictions.

This makes it harder to apply a uniform safety policy.

Legal responsibility

Another issue that still lacks a definitive answer is who should be held responsible when an open model is used in a harmful way.

The possibilities include:

• the original creator;
• the organization that made modifications;
• the company that deployed the system;
• the application developer;
• the user responsible for the action.

Excessively broad liability could discourage open projects. On the other hand, a complete lack of responsibility could make it harder to protect victims and investigate incidents.

Are closed models necessarily safer?

No.

Closed models offer control advantages. The provider can monitor usage, block users, fix vulnerabilities and update the system centrally.

However, restricted access also reduces the ability of independent researchers to examine how the technology works.

A vulnerability may remain publicly unknown for longer. In addition, customers need to trust the assessments and information provided by the company responsible for the system.

Supporters of openness compare this process to Open Source software development, where a broad community can review code and collaborate on fixes.

This comparison, however, is not perfect. Artificial intelligence models depend on large volumes of data, computing infrastructure and training processes that are difficult to reproduce.

Therefore, opening the weights does not automatically guarantee full transparency, safety or real reproducibility.

The report from the United States National Telecommunications and Information Administration, the NTIA, concluded that open models offer important benefits, but need to be continuously monitored to identify emerging risks. The agency recommended monitoring and evidence collection instead of immediate broad restrictions.

Impact for developers

For developers, the continuation of open models can significantly expand what can be built.

More technical alternatives

Teams will be able to choose between using commercial APIs or hosting their own models.

This choice makes it possible to balance factors such as:

• cost;
• performance;
• privacy;
• latency;
• customization;
• availability;
• data location.

Applications that need to run locally or without constant connectivity may especially benefit from smaller models.

Less dependence on APIs

An application built exclusively on an external API can be affected by changes outside the developer’s control.

Open models make it possible to create hybrid architectures, where part of the processing happens internally and external services are used only when necessary.

This can reduce costs and increase application resilience.

New responsibilities

Hosting a model also transfers responsibilities to the technical team.

The developer needs to handle:

• infrastructure;
• updates;
• monitoring;
• access control;
• data protection;
• response evaluation;
• vulnerability remediation;
• license compliance;
• prevention of misuse.

Therefore, access to the model does not mean that its use will be simple or cheap.

More attention to licenses

Each model may have different rules for distribution and commercial use.

Before integrating a model into a product, it is necessary to verify whether the license allows:

• commercial use;
• redistribution;
• creation of modified versions;
• offering the model as a service;
• use in certain sectors;
• integration with proprietary software.

The lack of standardization in AI licenses still creates legal uncertainty and can create difficulties for companies and developers. The Open Source Initiative highlights that simply publishing the weights is not enough to automatically classify a system as Open Source.

Impact for companies

For companies, open models can represent greater autonomy, but they also require technical maturity.

Potential benefits

Among the main benefits are:

• reduced dependence on vendors;
• greater control over data;
• the ability to run AI internally;
• customization for specific processes;
• cost predictability in certain scenarios;
• integration with legacy systems;
• creation of proprietary intellectual property based on adaptations.

Companies with large usage volumes may consider their own infrastructure more advantageous than continuous payment per request.

However, this analysis depends on the size of the model, the hardware required and the team’s capacity.

Costs that may be overlooked

A freely available model does not mean a cost-free operation.

The company may still need to invest in:

• GPUs or cloud services;
• energy;
• storage;
• data engineering;
• specialized professionals;
• security;
• observability;
• testing;
• governance;
• maintenance.

In many cases, a commercial API will remain simpler and more cost-effective, especially for smaller projects or limited usage.

Governance will be essential

Companies that deploy their own models will need to define clear rules for access, allowed data, usage logs and human oversight.

It will also be necessary to test the system before using it in decisions that affect customers, employees or critical operations.

Opening the model does not replace risk management practices. NIST maintains a voluntary framework to help organizations identify, assess and manage risks related to artificial intelligence throughout the system lifecycle.

What could balanced regulation look like?

One of the alternatives under discussion is to create rules based on the level of risk and the purpose of the system, instead of banning an entire category of technology.

Under this model, an application used to recommend music would not be treated in the same way as an AI system responsible for medical decisions, critical infrastructure or financial operations.

It would also be possible to differentiate between:

• low-capability models;
• advanced general-purpose models;
• systems used in critical sectors;
• applications with access to external tools;
• agents capable of executing actions autonomously;
• fully open models;
• models that only release their weights.

Targeted regulation could require additional assessments when certain capability or risk thresholds are reached.

Another possibility is to increase the responsibility of those who deploy and use the technology, especially when the final application has a direct impact on people.

This approach seeks to preserve research and innovation without ignoring the risks associated with the most powerful systems.

What could happen next?

The companies’ letter does not end the discussion. On the contrary, it increases pressure for the United States Congress to define a clearer policy.

In the coming months, the debate should focus on four main points:

Legal definition of openness

Governments will need to determine whether Open Source and Open Weight models should be treated in the same way.

This distinction may affect transparency, safety and licensing obligations.

Rules for advanced models

Models with higher capabilities may receive specific requirements, such as safety tests, technical reports and evaluations before release.

Responsibility for use

Lawmakers will have to define how far the responsibility of the original creator goes and when it shifts to those who modify, deploy or use the system.

Competition with China

Any restriction adopted by the United States will also be analyzed based on its impact on international competition.

Rules that are too strict could reduce the presence of American technologies in foreign markets. Rules that are too flexible could increase safety risks and facilitate the transfer of advanced capabilities.

Conclusion

The defense of open models by Nvidia, Microsoft, Meta, IBM and other organizations shows that artificial intelligence regulation has entered a decisive phase.

The debate is not only about publishing code or mathematical parameters. It defines who will be able to control AI technologies, which companies will be able to compete and how developers will be able to build new products.

Open models can reduce dependence on large providers, expand research and allow organizations to keep sensitive data inside their own infrastructure.

At the same time, distributing weights makes it harder to remove dangerous systems and allows protection mechanisms to be modified.

A broad ban could limit legitimate projects and favor foreign technologies. The absence of rules, however, could increase risks that are not yet fully understood.

The most likely path is regulation based on capabilities, applications and levels of risk, accompanied by technical standards, monitoring and clearer responsibilities.

For developers and companies, the main conclusion is that open models should continue to gain space, but their adoption will increasingly require knowledge of security, infrastructure, governance and licensing.

FAQ

What is Open Source AI?

It is a system made available in a way that allows its use, study, modification and sharing. To be considered truly Open Source, it is not enough to simply allow model download; it is also necessary to provide the elements required to understand and modify how it works.

What does Open Weight mean?

Open Weight means that the trained parameters of the model are available. This allows the system to be executed and adapted, but does not guarantee access to the data, training code or the entire process used to create it.

Why are Big Tech companies against broad restrictions?

The companies argue that general restrictions could harm competition, increase dependence on a few providers, limit research and favor models developed in other countries.

Are open models free?

Not always. The download may be free, but running a model requires infrastructure, processing, storage, maintenance and specialized professionals. In addition, the license may impose conditions for certain uses.

Can a company use open models with confidential data?

Yes, as long as it adopts a secure architecture and complies with applicable rules. One of the main advantages is the possibility of running the model internally, without sending data to an external API. Even so, the company needs to implement access controls, logs, testing and governance policies.

Are closed models safer?

They allow greater centralized control, but they are not immune to breaches, failures or misuse. Open models make external audits easier, while closed models allow the provider to apply updates and blocks centrally. Each approach has advantages and risks.