LLMs and Echo Chambers

The June edition of Best’s Review ran this article — “Large Language Models Examine Rising U.S. Homeowners Rates“ — which contained this introduction:

Best’s Review asked three artificial intelligence-based, large language model programs [this question]: Homeowners rates are rising in many U.S. states. What are the forces driving that increase and what might slow or speed the rise? The following are excerpts from the responses.

The three programs were ChatGPT-4o, Grok 3, and Perplexity AI. The substance of article notwithstanding, two things struck as important in reading it and in considering its implications:

Large language models (LLMs) aggregate (curate) their responses from massive datasets of text from diverse sources like books, websites, social media, and more. This data fine-tunes the model, based on its programming, to predict next words or tokens in given sequences. Each model is then trained by its programming to improve performance for targeted applications as determined by its developers.

Existing LLMs aggregate their responses from existing textual datasets. But at some point, they will learn to aggregate their responses from responses previously generated (aggregated) by AI-powered models. At that point, the entire process becomes some combination of a feedback loop and an echo chamber. That may not be bad necessarily. It depends on our level of awareness and ability to determine and consider the sources of the data being aggregated.

Life Imitates Art

These deliberations put us in mind of a short science-fiction story, “The Wall of Darkness”, written by Arthur C. Clarke and published in 1949. We won’t spoil the story here. But we will share this excerpt, taken from the moment at which Shervale, the story’s protagonist, is lifted atop the Wall and is about to walk across its surface. The excerpt seems chillingly prescient of what we now face with LLMs, regardless of when the story was published:

Behind the Wall, so Grayle had once said, lay Madness … “I shall be gone only a few minutes,” [Shervale] said with elaborate casualness. “Whatever I find, I’ll return immediately.” He could hardly have guessed how small a choice was his.

Should we be afraid? No. Should we be curious enough to explore the sources LLMs use and the ways in which they’re programmed to curate their responses? Yes.

If we remain curious enough to explore the sources of the models and the biases of their programmers, we might not face a wall of darkness. But our experiences may very well be as revealing and surprising as Shervale’s.