Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
On June 6, 2023, Senators Richard Blumenthal and Josh Hawley sent Meta CEO Mark Zuckerberg a bipartisan letter asking how the company assessed risks before giving researchers access to its LLaMA AI model—and what it did after the model weights spread online. The senators requested answers by June 15. Their letter raised concerns about possible misuse; it was an oversight request, not proof that LLaMA caused a specific crime or that Meta broke the law.
What happened
Meta announced LLaMA—short for Large Language Model Meta AI—in February 2023 as a family of text-generating models intended to support AI research. Meta’s initial arrangement was for approved researchers, not an unrestricted public download. Within days, the full model appeared on BitTorrent, according to the senators’ account in their June 6 letter.
The distinction matters. The controversy was not simply that Meta deliberately published the model for everyone on its first day. It was whether the restricted distribution approach was secure enough, whether wider dissemination was foreseeable, and whether Meta had adequately assessed the consequences. Once model weights are copied and circulate independently, the original developer has far less ability to monitor use or revoke access than it would with a hosted service.
Free tools Windows power users keep installed
One-click scans. No signup required.
The senators and the political context
The letter was sent by Richard Blumenthal, a Democrat from Connecticut, and Josh Hawley, a Republican from Missouri. At the time, Blumenthal chaired and Hawley was the ranking member of the Senate Judiciary Subcommittee on Privacy, Technology, and the Law. Their joint action was bipartisan oversight amid a broader debate about AI safety and accountability.
#1 Best Overall
The letter followed a May 2023 subcommittee hearing featuring OpenAI CEO Sam Altman, IBM executive Christina Montgomery, and New York University professor Gary Marcus. It was a request for information—not a subpoena, lawsuit, or final congressional finding. The senators asked Zuckerberg to respond by June 15, 2023.
What the senators criticized
Blumenthal and Hawley argued that Meta’s vetting and safeguards for approved researchers appeared limited and questioned whether the company had carried out a meaningful pre-release risk assessment. They warned that once the weights were publicly available, the model could be used without Meta’s monitoring or oversight. They also pointed to examples of LLaMA producing responses to prompts involving fraud, self-harm, antisemitism, and criminal activity, and questioned the adequacy of the company’s abuse-prevention documentation.
Those statements are the senators’ allegations and concerns, not independently adjudicated findings about Meta’s intent, negligence, or the model’s real-world effects. The letter’s comparison of LLaMA with centralized systems such as GPT-4 and Bard was not a controlled safety benchmark.
What the letter asked Meta to disclose
The five-page letter set out nine numbered questions with multiple subparts. Taken together, they sought answers in several areas:
Rank #3
- Pre-release review: What risk assessments Meta conducted, how many researchers it approved, what selection and vetting criteria it used, what safeguards it put in place, and who inside or outside the company was consulted.
- Release choices: What alternative protections Meta considered, how it balanced research benefits against release risks, and whether it considered an API, a Meta-controlled sandbox, or narrower researcher access.
- Response to dissemination: What the company did to limit or mitigate harm, track distribution and downstream use, identify possible fraud or spam targeting Meta users, and whether it issued DMCA notices, cease-and-desist letters, or similar demands.
- Abuse prevention: What training-time or later measures were taken to reduce the model’s ability to assist with fraud, self-harm, cybercrime, or other dangerous tasks.
- Future release policy: Whether Meta had changed its model-sharing practices, what lessons it drew, and whether it used thresholds—such as capabilities, task performance, parameter count, or other risk indicators—to decide when a model should not be publicly released.
- Transparency: Whether Meta could provide risk documentation comparable to OpenAI’s system-card approach for GPT-4.
- Training and personal data: Whether LLaMA was trained on Meta customer data, Facebook, Instagram, or WhatsApp posts, or other user-created content—and when Meta uses personal data for AI research, including work involving outside researchers or public models.
These were questions, not proof that Meta used those platforms’ user data to train LLaMA. A request to clarify data practices should not be mistaken for a finding about what training data the model contained.
Why model weights differ from a hosted chatbot
A hosted chatbot runs on a provider’s infrastructure. The provider can impose rate limits, log activity, restrict accounts, change refusal behavior, and respond to abuse reports. Downloadable weights can instead be copied, run locally, modified, and fine-tuned without the original provider’s ongoing visibility. A copy cannot reliably be recalled once it has spread.
That makes distribution choices consequential, but it does not make local models inherently harmful. Running a model locally can support independent research, privacy, lower latency, and development without dependence on a single provider’s API. Broader access can also let researchers examine weaknesses and reduce concentration of capability. The policy question is how to weigh those benefits against reduced control and harder-to-investigate misuse.
Recommended Free Tools
Terminology also matters. The senators discussed “open source” in the wider AI debate, but Meta’s original release was controlled access, followed by unauthorized dissemination. “Open source” does not necessarily mean the same thing as publicly downloadable weights, and a public download can still have license terms. For this episode, “researcher access” and “weights later circulated online” are more precise descriptions than saying Meta openly released LLaMA to everyone at launch.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What was known about misuse—and what was not
The letter pointed to troubling model responses and downstream projects, including Stanford’s Alpaca, as reasons for concern. It also referred to a LLaMA-based chatbot reportedly taken down after inaccurate or troubling responses. Such examples can show why misuse seemed plausible to the senators, but they do not establish a confirmed wave of fraud, malware, or other crimes caused by LLaMA.
A model’s ability to produce harmful instructions is not, by itself, evidence that a person used it to commit a particular offense. The cited Senate materials do not document a specific confirmed LLaMA-enabled crime. Claims of measurable real-world harm require separate evidence linking an incident to use of the model.
What is established about Meta’s response?
The available record summarized here establishes the letter and the senators’ statements, but does not establish that Meta publicly answered every question by the requested June 15, 2023 deadline. It also does not establish a definitive congressional conclusion that Meta violated a law. The letter’s lasting significance is the oversight question it raised: should a model developer account for foreseeable unauthorized spread when assessing the risks of a restricted release?
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Quick Recap
Sources
- Full June 6, 2023 letter to Mark Zuckerberg
- Senator Blumenthal’s announcement of the letter
- Meta’s original LLaMA announcement
- LLaMA research paper
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
