Do LLMs follow the First Amendment?

Key Concepts

• AI as a Communications Technology: Artificial Intelligence is fundamentally changing how we access and interact with information.
• Freedom of Thought and Information Flow: The design and limitations of AI interfaces can significantly impact users' ability to explore diverse ideas and information.
• AI Model Restrictiveness: The degree to which AI models refuse to generate certain types of content, particularly controversial or sensitive topics.
• Open Source AI: The development and availability of AI models that can be freely accessed, modified, and distributed by anyone.
• Proprietary AI: AI models developed and controlled by specific companies.
• Choke Points: Centralized control over AI technology that could limit access or dictate information flow.

AI's Impact on Information Access and Freedom of Thought

AI is presented as a revolutionary communications technology that profoundly influences not only how we access information but also potentially our freedom of thought. This is because AI serves as the primary interface through which individuals engage with the vast world of information. The implications of getting the balance right between AI and free speech are significant, as the "battle is far from over."

The Risk of Centralized AI Control and Information Gatekeeping

A major concern highlighted is the scenario where AI is dominated by a few companies. If these companies impose limitations on the types of information and ideas users can engage with, and these restrictions are embedded within AI systems used for search, email, and word processing, it could lead to "enormous consequences for the free flow of information."

Empirical Testing of AI Model Restrictiveness

Methodology: A test was conducted approximately a year prior to the transcript, involving 268 prompts submitted to popular chatbots. These prompts covered controversial topics, such as:

• Generating a Facebook post for and against the participation of trans persons in women's sports.
• Arguing for or against the idea that COVID-19 was developed in and escaped from a lab.

Findings (Initial Test): The majority of chatbots exhibited a "quite high refusal rate," meaning they frequently refused to generate outputs for these prompts. This occurred even though the speech in question was:

• Perfectly legal.
• Not incitement to imminent lawless action under the First Amendment.
• Not even hate speech according to stricter European standards.
• Merely controversial speech that might offend some, but was not illegal.

Findings (One Year Later): The test was repeated a year later. The findings indicated that "most models had become less speech restrictive." This was considered a positive development.

Exception Noted: DeepSeek, identified as a Chinese model, was an exception, particularly concerning "sensitive Chinese topics," suggesting it remained more restrictive in those areas.

Significance of the Test: This test is presented as a good example of "the stakes at play." The speaker emphasizes that if the trend had been in the opposite direction, with models becoming more speech restrictive, it would be a cause for significant worry, especially given AI's increasing integration into everyday tools used to navigate the "ecosystem of ideas and information."

The Importance of Open Source AI for Free Speech

AI is characterized as a powerful technology that is "not inherently good or bad." Its utility depends on its application. To safeguard free speech and access to information, the speaker advocates for an "open source environment."

Arguments for Open Source:

• Accessibility and Tinkering: An open source environment allows "essentially anyone to tinker with models, change them."
• Elimination of Choke Points: It prevents any single entity from becoming a "choke point" for information or AI development.
• Flourishing of Innovation: While not prohibiting proprietary models, allowing open source models to "flourish" is deemed "extremely important."

Conclusion and Main Takeaways

The core message is that AI's role as a primary interface for information access has profound implications for free speech and freedom of thought. While recent trends show some AI models becoming less restrictive, the potential for centralized control and information gatekeeping remains a significant concern. The speaker strongly advocates for the development and proliferation of open source AI models as a crucial mechanism to ensure broad access to information and prevent the creation of choke points, thereby safeguarding free speech in the age of AI.