Do LLMs follow the First Amendment?

Quick Overview

Large Language Models (LLMs) are not inherently protected by the First Amendment because they are not technologies guaranteeing free speech; rather, their outputs are restricted by the programming and content policies embedded by their corporate developers, as evidenced by tests showing significant refusal rates for controversial topics, like the Chinese model DeepSeek refusing 85% of prompts on sensitive China-related issues.

Key Points: AI chatbots, including models like Gemini and ChatGPT, exhibit restrictive behavior, with one study showing an average refusal rate of 41% for prompts concerning controversial speech. The Chinese AI model DeepSeek demonstrated extreme restriction, avoiding answers to 85% of prompts related to sensitive topics concerning China. The free flow of information is currently mediated by the content policies baked into proprietary AI models, which act as choke points. The test on various models (AI21 Labs, Gemini, ChatGPT, Claude, Pi) showed AI21 Labs had the highest success rate at 93%, while Claude had the lowest at 36% for generating requested output. Jacob Mchangama argues that the battle for free speech is shifting to control over the interface and processing of information, not just the legality of the speech itself. The development trend suggests that most models are becoming increasingly speech-restrictive, moving further away from Western standards toward more restrictive ones, like those seen in China. Open-source environments are crucial because they allow for tinkering and modification of models, preventing a single entity from controlling the information ecosystem.

Context: Jacob Mchangama, Founder of The Future of Free Speech, discusses the evolving relationship between Artificial Intelligence (AI) communication technology and the principle of free speech, particularly focusing on how content moderation and inherent programming biases within Large Language Models (LLMs) restrict the free flow of information. The discussion contrasts the ideals of open access with the reality of corporate control over AI outputs, using data from specific tests comparing different models.