Why is AI text hard to read
AI text is hard to read because the newest models pack too much meaning into too few words. A new paper by Clark, Martin, and Mahowald (NYU and UT Austin) measured it. The paper looked at 14 models. The nine newest and strongest overuse sentence shapes that are known to slow human readers down.
This is a different problem from the old one. The old complaint was filler and padding. The new complaint is text that is short, dense, and reads like notes the model wrote for itself.
Asking for shorter text likely makes it worse. What works is to tell the model who is reading, and to run one plain rewrite pass before you send anything. The workflows are below.
Six habits that slow readers down
The paper counted six habits in AI text from chat, coding, and essay tasks. Reading research links each one to extra effort for the reader.
| Habit | Dense version | Plain version |
|---|---|---|
| Rare word pairs | The plan has a hiring-shaped hole. | The plan does not say who we will hire. |
| Long gaps between words that belong together | The report that the team in Toronto that handles renewals sent us is wrong. | The Toronto renewals team sent us a report. It is wrong. |
| Dropped "that" | The issue the client raised is fixed. | The issue that the client raised is fixed. |
| Noun stacks (three or more in a row) | customer churn risk score update | an update to the score that shows which customers may leave |
| Nouns used as verbs | Let's whiteboard it, then inbox the client. | Let's sketch it on a whiteboard, then email the client. |
| Things that act like people | The deadline does not care about your holiday. | The deadline will not move for your holiday. |
Compared with human writing, all nine top models overused rare word pairs and things that act like people. Most also used longer gaps, more noun stacks, and more dropped "that" in chat.
The researchers also sent each AI sentence through a translator into Japanese and back. The meaning stayed, and most of the habits dropped off. So a plainer way to say the same thing was there, and the model did not pick it.
Stack a few of these in one sentence and you get this:
Dense: The pricing page redesign signup lift the team reported hides a churn problem the dashboard never caught.
Plain: The team reported that more people signed up after we redesigned the pricing page. But more customers are also cancelling, and the dashboard does not show that.
Why models picked up these habits
The likely cause is training that rewards right answers and short output, with no human reader in the loop.
Human language is shaped by listeners. Sentence shapes that people cannot follow fall out of use. People also add small helper words like "that" when the next part is hard to predict.
The newest models get a second round of training on math, code, and reasoning tasks. That training checks whether the answer is right. It does not check whether a person can read the output with ease. Context windows are also limited, so short dense text wins.
The authors tested the idea on a small open model, Qwen 2.5 (1.5B). They trained it to turn structured data into English sentences, with a penalty for length. Nobody told it to be terse.
In the runs with no guard against drift from the base model, it picked the shorter wording over 98% of the time on the two "that" choices it was trained on. It also dropped "that" in sentence types it never saw in training.
Three other results point the same way:
- In one family of small open models (Olmo), the habits were weakest in the base model, stronger after instruction tuning, and strongest in the reasoning model.
- The habits were weaker in single essays and stronger in long chat and coding sessions.
- It can be fixed. Claude Opus 5.5 scored lower than Opus 5 on 19 of 21 comparisons.
What this changes for a regular user
Three things follow. These are inferences from the evidence. The paper did not test prompts.
- "Be concise" is the wrong instruction. Pressure on length is what produced the habits in the experiment. A prompt that asks for short text pushes the same way. The old advice to cut filler does not help either, because dense text has no filler left to cut.
- Long chats produce worse prose. By turn 30 the model is using labels it made up at turn 5. It assumes you remember them.
- The model writes for a reader like itself. It can keep a whole long context in view at once. You hold a few words at a time. Text that the model can read back with ease can still be hard for you.
Five workflows to make AI text read like a person wrote it
Use the first one every time. Add the others as the stakes go up.
1. Name the reader before you ask
Add these lines to any writing request. Save them as custom instructions so every chat starts with them.
Write this for [who]. They were not in this conversation and do not know our terms.
Use full sentences and common words. Keep helper words like "that" and "which".
If the shorter wording is harder to read, use the longer one.2. Move the writing to a fresh chat
Do this after any long working session.
- In the long chat, ask: "List the facts, decisions, and numbers from this session as plain bullets. Explain any term we made up."
- Open a new chat and paste the bullets.
- Ask for the email, post, or report there, with the reader named.
3. Run a plain rewrite pass
Do this for anything you will send or publish. Paste the draft under this prompt.
Rewrite the text below for a reader who is new to the topic. Keep every fact. Do not make it shorter.
1. Split any sentence over 25 words.
2. Keep each subject close to its verb.
3. Put back "that", "which", and "who" where they were dropped.
4. Break up any run of three or more nouns.
5. Swap any noun used as a verb for a plain verb.
6. Where an object or idea does something only a person does, say plainly what happens.
7. Swap any term we made up for a common phrase.4. Give it your voice
Do this for anything that goes out under your name.
- Paste two or three things you wrote yourself.
- Ask the model to match your sentence length and word choice, not your content.
- Fix the first paragraph by hand. Then ask it to carry your edits through the rest.
5. Test it on a cold reader
Do this for high stakes text.
- Read it out loud and mark each place you stumble. This is the real test.
- Paste it into a new chat and ask: "Say what this means in two sentences. List any term a newcomer would not know."
- Fix only the marked sentences.
A model cannot do step 1 for you, because dense text is easy for it. Use it only to catch terms that were never explained.
Two quick extras
- If one model keeps writing dense text, try another one for the writing step. The paper found real gaps between models.
- You can copy the paper's method and translate the text into another language and back. That strips most of the six habits. It also strips your voice, so save it for plain notices.
What the paper does not prove
The paper counts habits that are known to slow readers. It did not time real people reading this AI text.
- It covers English only, with 15 conversation starters and 7 essay prompts.
- The training test used one small model on one narrow task. The link to how the top models are trained is still a hypothesis.
- The human baselines are movie subtitles, StackExchange posts, Wikipedia, and high scoring essays by US students in grades 6 to 12. They are rough matches for chat, coding, and essay writing.
- For rare word pairs and things that act like people, the direct comparison with human writing comes from the essay task only.
Source: Why is AI-generated language so hard to understand? by Thomas Hikaru Clark, Matias B. Martin, and Kyle Mahowald (2026).
