US Government Asks Meta to Share Its AI Models for Review

What happened?

According to an exclusive report by Engadget citing anonymous sources familiar with the matter, the US government has urged Meta to share its artificial intelligence models, including the Llama family, for review by federal authorities. The request, which has not yet been officially confirmed by the White House or Meta, comes amid growing concern over the safety and ethical risks associated with generative AI. Meta, owner of models such as Llama 2 (released in July 2023 with 7 billion, 13 billion, and 70 billion parameters) and the more recent Llama 3 (April 2024, with 8 billion and 70 billion parameters), has been a key player in open-source AI development, raising questions about control, transparency, and the potential proliferation of malicious uses.

Historically, the relationship between the US government and big tech on AI has been one of voluntary cooperation. In July 2023, seven companies (Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI) signed voluntary commitments with the White House to conduct internal and external safety testing, share information, and develop watermarking systems. However, the new approach appears more coercive: a direct request for access to model weights and architectures, representing a qualitative leap in government oversight.

Why is this important?

This government request represents a step toward greater AI regulation in the US, a country that has so far favored a self-regulatory and innovation-promoting approach. Unlike previous voluntary commitments, where companies voluntarily shared information about safety testing, here we see direct pressure on a tech giant to reveal the internal details of its models. If realized, it could set a precedent for other companies like OpenAI, Google, or Microsoft to also be required to open their models for government review, especially those considered high-risk or with dual-use capabilities.

The open nature of Meta's models (such as Llama 2 and Llama 3, distributed under open-source licenses) contrasts with the closed systems of other competitors, like OpenAI's GPT-4 or Google's Gemini. This adds complexity to the debate over intellectual property and national security. On one hand, open models allow for greater transparency and community auditing, but they also facilitate misuse by malicious actors. According to a 2023 RAND Corporation report, open-source large language models can be modified to generate disinformation, malware, or hate speech with relative ease. The government's request may seek to mitigate these risks without hindering open development, but it also raises questions about the confidentiality of Meta's training data, which includes user conversations and social media content.

Moreover, the timing is crucial: we are less than a year away from the 2024 US presidential elections, and generative AI has already been used to create deepfakes and disinformation. The Biden administration has issued executive orders on AI (October 2023) and established the AI Safety Institute (AISI) under NIST, but still lacks legal authority to demand access to proprietary models. This request could be a preliminary step toward stricter regulation, similar to the European Union's AI Act, which classifies models by risk level and requires conformity assessments for high-risk ones.

Potential consequences

The implications are multiple and profound. On one hand, it could improve safety by allowing the government to identify vulnerabilities or biases before models are widely deployed. For example, racial or gender biases in training data could be detected, or dangerous capabilities such as generating biological or chemical weapons. A 2023 study by Anthropic showed that models like Claude could be tricked into providing dangerous instructions, and government review could help close these gaps.

On the other hand, it creates tensions over the confidentiality of algorithms and training data. Meta could face dilemmas between complying with the request and protecting its competitive advantage. Its Llama models have been adopted by startups and companies for their low cost and customizability; revealing the weights could allow competitors to replicate its technology. Additionally, Meta's training data includes information from Facebook and Instagram users, which could violate privacy agreements if shared without consent. The company is already under scrutiny for its data handling in the Cambridge Analytica case and GDPR fines.

This move could accelerate the creation of similar regulatory frameworks in other jurisdictions. The European Union, with its AI Act approved in March 2024, already requires transparency and risk assessment for general-purpose models. If the US follows this path, global AI companies will have to comply with multiple standards, increasing compliance costs. For startups that rely on open models like Llama, there could be uncertainty about the future availability of these models under permissive licenses. Meta might be forced to restrict access or introduce versions with more restrictive licenses, as it already did with Llama 2 (which requires approval for certain commercial uses).

In the market, Meta's shares fell slightly after the news, reflecting investor concern over potential regulations. Competitors like OpenAI, which keep their models closed, could benefit in the short term if Meta is forced to share its intellectual property. However, in the long run, all players will face similar scrutiny.

What readers should know

It is important to understand that this report has not yet been officially confirmed by either party. The information comes from unspecified sources cited by Engadget, so it should be taken with caution. However, it reflects a growing trend: governments are seeking to closely oversee AI development. For users, this could mean safer AI products, but also a possible slowdown in innovation due to increased regulatory requirements. Companies using Meta's models should be alert to potential changes in terms of use or availability. For example, if Meta decides to withdraw Llama 3 from open source, developers building applications on it could be affected.

The historical context is key: in 2022, the US government already asked AI companies like OpenAI to share information about language models to assess national security risks, but informally. Now, the pressure is more explicit. Additionally, in 2023, the FTC launched an investigation into OpenAI for possible consumer protection violations, indicating a general tightening of scrutiny. Meta's response will be crucial: if it refuses, it could face legal or political action; if it complies, it will set a precedent for the industry.

Finally, readers should consider that open-source AI is a double-edged sword: it promotes innovation and democratization, but also facilitates misuse. The balance between safety and openness will be the central debate in the coming years. As an analyst from TheVortiq noted: 'The request to Meta is a symptom that the era of self-regulation in AI is coming to an end.' Stay informed and critically evaluate sources.