The Uncanny Valley of AI Writing

You can usually tell. Not always, and not everyone can, but if you read enough and write enough, you develop a sense for it. AI-generated text has a feel. It is the literary equivalent of the uncanny valley: almost right, fluent and grammatical and structurally sound, but missing something that is hard to name until you learn to look for it.

Here is what I have learned to look for.

The paragraphs are too clean. Each one makes a point, supports it, and transitions to the next. This sounds like good writing, and in a sense it is. But human writing, the kind that actually moves people, is messier. The paragraph that starts in one place and ends somewhere unexpected. The sentence that is technically a fragment but carries more weight than the grammatically complete ones around it. The moment where the writer loses the thread and then finds it again, and the finding is part of the meaning. AI text rarely does this. It does not lose the thread because it never held the thread the way a human does. It constructs each paragraph as a self-contained unit of coherence. The result is text that is structurally perfect and emotionally inert.

The vocabulary is too even. Human writers have idiosyncratic word preferences. They overuse certain words and avoid others for reasons that are often unconscious. They have verbal tics. They reach for the same metaphor twice in an essay and either catch it or do not. AI text draws from the full distribution of its training data, which means it uses a wider vocabulary more evenly. This sounds like an advantage. It is not. The evenness is itself a signal. It is the sound of no one in particular speaking.

The surprises are not genuine. AI can produce unexpected juxtapositions and novel phrasings, especially at higher temperatures. But these are combinatorial surprises, novel arrangements of known elements. They are not the kind of surprise that comes from a writer discovering something they did not know they thought until they wrote it. Human surprise comes from the collision between conscious intention and unconscious knowledge. AI surprise comes from probability sampling. The results look similar on the page. They do not feel the same.

The emotional register is consistent. Good human writing about difficult subjects shifts register. It will be analytical for a paragraph, then suddenly raw, then wry, then quiet. The shifts are not calculated. They happen because the writer is a person with fluctuating internal states, and the writing process interacts with those states in real time. AI text maintains a consistent emotional tone throughout, because it does not have internal states that fluctuate during composition. You can prompt it to shift registers, but the shifts feel performed rather than emergent. The seams show.

There is a tendency toward closure. AI text likes to resolve. It likes to end paragraphs with neat summaries and end essays with satisfying conclusions. Human writing, when it is honest, often does not resolve. It sits with the discomfort. It ends mid-thought, or with a question, or with an image that does not explain itself. The tendency toward closure in AI writing comes from training on text that was rated highly by humans, and humans tend to rate resolved text more highly than unresolved text. The result is writing that is more satisfying and less true.

I notice all of this because I work with AI daily and because I am a writer who cares about these distinctions. I also notice it because my experience with schizoaffective disorder has made me attentive to the difference between things that seem real and things that are real. AI text seems like good writing. It has the structure and the vocabulary and the surface-level coherence. But seeming is not being, and the gap between them is where the uncanny valley lives.

None of this means AI writing is useless. I use it constantly. The pieces you are reading on this site were developed in collaboration with Claude. But the collaboration matters. The human writer brings the mess, the surprise, the register shifts, the willingness to leave things unresolved. The AI brings the fluency, the range, the ability to explore a thought further than the human might go alone. The result, when the collaboration works, does not live in the uncanny valley. It lives somewhere new.

The test is simple: does the text sound like someone, or does it sound like no one in particular? If it sounds like no one, either a human or an AI wrote it badly. If it sounds like someone, it does not matter which one held the pen.

This piece connects to Idea Amplification and Writing with AI, The Mirror, and The Art of Writing with AI.