AI 的十万个为什么

One of the most painful arguments I keep having with fellow techies is the question of whether you can distinguish between human-written and AI-generated text.

Their skepticism is rooted in reason: at their core, LLMs are state-of-the-art statistical models of how humans talk. If so, the output from the model should be almost by definition indistinguishable from human language under any statistical test.

I don’t think this is always argued in good faith; at least some of the debates are started by folks who wish to maintain deniability for their own underhanded use of the tech. But if you sincerely hold this belief, I present you the following collage:

The image shows about 220 Amazon book covers that appear if you search the site for “100000 whys” (link). Some of these books are category bestsellers in children literature. You can view a zoomable, full-resolution version here.

There’s nothing inhuman about any of these titles or covers. At the same time, I probably don’t need to convince you that you’re staring at the purest form of AI slop that now fills up many nonfiction book categories on Amazon. More specifically, what we’re seeing here is the artifact of the tools being quasi-deterministic: if a hundred “authors” give their favorite AI tool a similar prompt — say, “generate a reference book for children” — the model will produce functionally identical output perhaps 80% of the time.

The similarities in the collage go far beyond the choice of titles: for example, all the covers in the two top rows feature a roaring T-Rex on the left. There are many other clusters in the data, too. Look for a recurring red-and-white cartoon rocket, a golden retriever, a lion, and so forth. The similarities extend even to author names: Ethan Bright, Nolan Bright, Pamela Bright, Daniel Bright, Thomas Bright, Andrew W. Bright, Mayan Bright, Mary Bright, Levi Bright — the Brights must be a big and exceptionally talented family.

This is precisely what makes LLM writing distinctive: it’s not that the models’ individual mannerisms are different from ours. It’s that they resort to the same, complex set of mannerisms in response to almost any normal prompt. This is a fuzzy signal, so you shouldn’t fire your intern when they say “it’s not this — it’s that”. But in more casual settings, it’s OK to trust your gut. In fact, these instincts are becoming increasingly important because traditional models of online interactions fall apart if it takes much less effort to produce content than to engage with it.

PS. If you’re using an LLM to automate blogging: yes, the tech is amazing, but chances are, your publication could be renamed to “100,000 Whys”.