AI Expert Slams Journalist's Misunderstanding: Why Chain-of-Thought Is in English

Overview

Recently, a video hosted by David Sirota featuring interviews with Ed Zitron and Matt Stoller sparked heated discussion on social media. However, AI expert Alexander McCoy issued a sharp critique of the interview, pointing out that it was riddled with fundamental misunderstandings about how artificial intelligence works.

McCoy argued that the interviewees' technical understanding suffered from serious deficiencies, particularly when explaining certain key events and mechanisms, often reaching conclusions contrary to the facts. This criticism has ignited a broader online debate about whether technology journalists possess sufficient technical literacy.

Severe Deviations from Technical Facts

McCoy voiced his opinions on X (formerly Twitter), directly stating that the interview content was "extremely disappointing." He made it clear that the interviewees' understanding of how AI mechanically operates contained fundamental errors.

Among these, the interview's description of the Hugging Face incident was characterized by McCoy as "objectively wrong." As a platform with significant influence in the open-source AI community, the details surrounding the incident should have been conveyed accurately. However, the interviewees' statements clearly failed to meet this standard, instead disseminating misleading information.

Why Are "Chain of Thought" Mechanisms in English?

Discussions surrounding the Chain of Thought (CoT) mechanism in AI models further exposed gaps in the interviewees' knowledge. McCoy pointed out that Zitron's response to the question of "why chain-of-thought outputs appear in English" demonstrated a "deeply embarrassing ignorance."

In fact, the adoption of English for chain-of-thought reasoning—rather than a so-called "neuralese"—is primarily driven by safety considerations. Research indicates that using natural languages like English as an intermediate reasoning process helps human supervisors audit the model's logical paths, thereby better identifying potential risks and biases. Conversely, if chain-of-thought existed directly in unintelligible internal representations, the model's interpretability would increase significantly, thereby raising the risk of misuse.

The Cognitive Gap Between Media and Technology

Although McCoy holds a negative view of the aforementioned interview, he acknowledged Sirota's subsequent decision to invite New York state legislator Alex Bores as a guest. Bores is championing a $30 million initiative aimed at uniting Democrats to advance AI safety measures.

This shift reflects a core contradiction in the current tech discourse: while the public and media are highly focused on AI governance, some commentators lack the necessary technical background, causing discussions to drift away from factual foundations. Internet user Nat Purser also lamented the frustration of hearing people endorse such technologically ungrounded viewpoints.

This controversy once again highlights the importance of enhancing the technical literacy of media practitioners against the backdrop of rapid AI development. Only by establishing accurate technical understanding can the public be guided toward rational and effective policy discussions.