AI writes faster than anyone can read

Human writing is uneven, and the unevenness is load-bearing. A writer gets excited and the sentences shorten. They hedge and the sentences lengthen. They lose interest in a sub-point and it gets one line. Those variations are navigational: your eye uses them to work out what matters before you have read any of it.

Generated prose is flat by construction. Every paragraph is roughly the same length, every sentence roughly the same complexity, every section equally confident. There is no prosody on the page, so there is nothing for your eye to grip. You are not reading a landscape, you are reading a lawn.

That flatness is why skimming fails here in a way it does not fail on a news article. Skimming works by sampling the high points; if the text has no high points, sampling returns a representative average of nothing in particular. So you either read all of it properly or you absorb very little of it, and most people, most of the time, do neither.

The W3C's accessibility guidance has a success criterion on reading level that makes the underlying point: when text is demanding, the remedy is supplementary means of understanding it, not a demand that the reader try harder. An audio version is explicitly one of those means.

The instinct is to treat listening as a way of not reading: put it on, do something else, absorb it by osmosis. That works for undemanding material and it is not the interesting use.

The interesting one is having a voice read while you are still looking at the text. Reading silently is entirely self-paced, which sounds like an advantage and is where attention leaks away. Nothing notices when you stop. You can sit on one sentence for forty seconds and the page will wait, because a page has no opinion about whether you are present.

A voice does not wait. It moves at a rate you set, and that steady forward pressure does two things at once. It discourages the micro-pauses that become drift, and it makes drift visible, because the gap between where the voice is and where your eyes are becomes obvious inside a sentence rather than a page and a half later.

On flat generated text this matters more than on anything else, precisely because the page gives you no other signal that you have stopped taking it in.

A voice on its own solves half of it. It sets a pace and it exposes the drift. What it cannot do is help you get back in, because sound has no position. Look away, look back, and audio offers nothing to look at.

So the reader marks where it is, at three levels at once while it speaks: a soft tint on the paragraph, a stronger one on the sentence being spoken, and the single word the voice is on. Look away for five seconds, look back, and your eye goes straight to the lit word. The hunt for your place, which is the expensive part of losing concentration, costs nothing.

Being able to click any sentence and have the voice resume from exactly there matters more than it sounds like it should. Re-reading a dense passage stops being a decision about whether it is worth the interruption and becomes one click. Over a week that changes how often you actually do it.

This works the same way on a PDF, an EPUB, a pasted block of text, or the page you are already on, including the chat interfaces where most generated text is actually read.

There are two honest places for a model inside a reading tool, and a lot of dishonest ones.

The first is the voice. Modern neural voices are models, and they are better than the ones built into your computer. Lumotext offers a small set of them, and they are optional: the free voices that ship with macOS and Windows run on your own machine, cost nothing, work offline, and are much better than the one your browser picks by default, which almost nobody changes. Start there before you consider paying anyone, including us.

The second is timing. Hosted audio arrives as a sound file with no word boundaries in it, so the highlight would have to be guessed from the length of the clip. Lumotext instead sends the audio through a transcription model to get real word timings back, which is why the highlight lands on the word being spoken rather than near it. That is the whole of what the model is for.

What is deliberately missing: the reader does not summarise, does not answer questions about your document, and does not add a chat panel. Those features exist in a dozen products already, and every one of them asks you to trust a summary of something you did not read. A tool whose purpose is to get you through the text should not be optimised around not reading the text.

Four habits, one session to build.

1. Install a decent voice first. Two minutes, free, and it changes your opinion of the entire category. Here is where they are hidden on macOS and Windows.
2. Set your speed on hard material, not easy material. Most people land between 1.1× and 1.4× for familiar subjects and slower for new ones, and almost nobody arrives there by accident.
3. First pass at pace, second pass by clicking. Let it run the whole way through for the shape of the argument, then click back into the three or four passages that carry the weight.
4. Stop when comprehension goes, not when the text ends. The point is not to finish. The point is to have read it.

Worth saying plainly, because this category tends toward inflated claims.

You will not get through twice as much. Anyone promising that is selling speed reading with a new interface, and speed reading trades comprehension for words per minute at roughly one for one.

What tends to happen instead is that the second half of a long piece gets the same quality of attention as the first half, which is usually not true otherwise. Long items stop being the ones you skim because you already spent your attention on the short ones.

And nothing here replaces deciding what deserves a close read. The most effective thing available to anyone drowning in generated text is still being ruthless about which of it needs reading at all. No tool fixes reading the wrong things carefully.