You can hear which model wrote it

This is a shift from where we were two years ago. Early LLM output was detectable mainly by its failures - the bland sentence construction, the over-hedged claims, the telltale phrases like 'it's important to note'. Those failures are mostly gone. What's replaced them is subtler: each major model now has a recognisable default cadence, a distribution of sentence lengths, a set of preferred transitions, and a way of structuring arguments. Read enough of it and you start hearing the model through the content. Once you hear it, you can't un-hear it.

Why this is a problem

The effect on reading is specific. When you recognise a model's voice in a piece of writing, your attention shifts. A small background process starts running that evaluates which model produced this, how much of it is the writer and how much is the tool, whether the writer edited the output or passed it through unchanged. That process runs continuously alongside your reading comprehension, and it consumes some of the cognitive budget that would have gone to the actual content. By the time you've finished the piece, you remember the meta-observation - 'that was probably GPT' - more than whatever the piece was trying to say.

This is the same mechanism as noticing typos. A typo pulls your attention sideways for half a second, and the sentence's actual meaning is lost in the shuffle. A detectable model voice does this continuously, at every transition word and every cadence you recognise. The writing feels competent. The experience of reading it is tiring.

The 'unsigned prose' problem

A useful framing I've borrowed from conversations with editors: detectable AI-assisted prose is unsigned. It has no point of view. The writer's decisions are mostly absent because the model's defaults fill the gaps. You can tell the piece was produced, not composed. Competent but unsigned is the specific failure mode, and it's worse than bad writing because bad writing at least has a person making distinguishable choices.

Corporate communications have quietly been producing a lot of unsigned output over the last eighteen months. The difference is visible on brand terms. Readers don't know exactly what changed, but they read the communications and feel nothing. The companies producing them assume the unchanged metrics mean nothing is wrong. What's happening is that attention is migrating away from the channel, slowly, without the kind of signal that shows up in open rates or click-through. The brand quietly becomes furniture.

The workflow fix

The technical solution to this problem is easy to describe and hard to apply. You have to edit more than one pass. Measured in user research, it takes about three passes of heavy rewriting to remove a model's stylistic signature from a piece of prose. Most writers do one. The difference is the distance between 'sounds like GPT' and 'sounds like you wrote it.' The effort budget that would have gone to writing it from scratch has been displaced to rewriting it afterwards - often with similar total cost, which is why the productivity advantage of AI writing is smaller than it looks when measured against high-quality output.

Writers who produce undetectable AI-assisted output tend to do something specific. They use the model as a first-draft engine and rewrite the draft aggressively. They don't accept sentences whole. They reorder paragraphs. They inject specific examples and personal phrasing the model wouldn't have produced. The final piece is the product of the writer's editorial judgement, not the model's default. That workflow exists. It's about three times more effort than 'paste the output in', which is why most people don't do it.

Competent but unsigned is the specific failure mode. It's worse than bad writing.

A side-by-side exercise

If you want to train your ear for the detection, a useful exercise: give the same prompt to three different LLMs and read the outputs side by side. After a few weeks of this, you start recognising each model's default rhythms and patterns. You'll find that you can identify which model produced a piece of text cold, on a first read, about 70% of the time for long-form writing. This is not a skill that requires special training. It's a skill that develops once you've been exposed to enough examples.

Once you have this skill, your own writing with AI assistance changes. You notice when a sentence sounds like the model rather than you, and you rewrite it. The noticing is the whole discipline. You can't want to remove the signature if you can't detect it in the first place.

The adversarial future

There's an adversarial direction to this that's just starting. Once users are reliably detecting model voice, the next generation of models will be trained to reduce stylometric signatures. We'll get models that produce prose optimised to be undetectable, specifically because the labs will have noticed the distaste for detectable output. This will work, for a while, until users calibrate to the new default and the detection skill catches up. This arms race is more subtle than the current deepfake-detection debate, but it's happening on the same timeline.

The durable answer, for any writer who cares about their voice, is to keep writing. The texture and specificity of your own prose is the signature that a model trained on general text will never fully reproduce. Your LinkedIn post about a specific client engagement, with the actual numbers, with the phrases your colleagues use, with the jokes nobody else would make - that's hard for any model to fake because it requires knowledge the model doesn't have. If your AI-assisted output reads like the model, it's because you didn't give it enough of you. Fix that, and the detection problem takes care of itself.