Blog · Per-site setups

Reading ChatGPT with dyslexia

11 min read · June 2, 2026

ChatGPT, Claude, Gemini and the rest of the AI chatbots have quietly become one of the biggest reading surfaces in a modern adult's day. You ask a small question, you get back six paragraphs. You ask for a summary, you get back a wall of bullet points longer than the source. The chat interface is structurally a long-form reading tool wearing a messenger's clothes. For dyslexic readers that mismatch matters - the messenger affordances (no reader mode, no print stylesheet, no native font picker, mid-stream tokens appearing as the model types) work against the long-form content the model actually produces. This is a practical setup that makes ChatGPT's output easier to parse, the prompts that shape it into a friendlier shape, and the TTS pairing that finishes the job when your eyes are tired.

The short answer

Open ChatGPT in a pinned browser tab (not the desktop app), run a font-override extension to swap the default to OpenDyslexic, Lexend or Atkinson Hyperlegible, turn off the streaming animation in Settings -> Personalization so the text stops crawling letter-by-letter, and add a one-line system instruction in Custom Instructions that asks for short paragraphs, plain words and a top-line summary on every reply. Pair the page with Chrome's built-in Read aloud for the longest answers. The rest of this article is the why and the workflow.

Why ChatGPT is hard to read

The chat interface looks calm - a single column, generous line height, a readable system sans-serif. By the standards of dyslexia-friendly web design the baseline is not terrible. The trouble is what the model produces inside that frame.

Replies are long and structurally heavy. A typical answer has a one-line opener, three or four sub-headings, a nested bullet list under each, a code block, a table, and a hedging paragraph at the end. That is more structure than most articles you read on the web. Each structural shift forces the eye to re-anchor, and dyslexic readers pay a higher cost per re-anchor.
The text streams in. By default the response appears one token at a time, a few characters per beat. For most readers this is a friendly "I'm working" signal. For a dyslexic reader trying to parse a sentence, watching a sentence rebuild itself five times before it settles is exhausting. The page is doing the same work your eyes are.
The model leans on jargon when uncertain. Models hedge with long Latinate words ("subsequently", "furthermore", "additionally") and complex sentence structures. Those are the words most likely to trip a dyslexic reader, and the model uses them where a confident human writer would use a comma.
There is no reader mode. Chrome's built-in Reader Mode needs an article element on the page to activate. The ChatGPT canvas is rendered as a chat thread, not an article, so Reader Mode does not light up. The same is true on Claude, Gemini and most of the rest.
Code blocks and inline code reformat the line. Inline code spans switch to a monospace face at a slightly different size and weight. The model uses inline code liberally even in non-technical answers (file names, button labels, the names of menu items). A single paragraph can shift typeface three times mid-line.

The fix is the same family of moves we recommended for Slack and Gmail: calm the surface, swap the font, then shape the content with a small upstream tweak. With ChatGPT you have a lever the others do not - you can ask the model to write differently.

Step 1 - the settings most people skip

ChatGPT has a Settings panel that almost no user opens. Click your avatar in the bottom-left and pick Settings.

Personalization - turn off the streaming animation

Under Settings -> Personalization, find Show typing animation (sometimes labelled Streaming responses in newer builds) and turn it off. The reply now appears as a single block once it has finished generating. You wait a couple of extra seconds and you get a settled page to read instead of a re-flowing one. For dyslexic readers this is the single biggest perceived-comfort change in the whole product.

Appearance - System theme, not a forced one

Under Settings -> Appearance, set theme to System. This lets you control light/dark from your operating system - useful if you have an automatic schedule, or if you already have an opinion from our dark mode and dyslexia piece. ChatGPT's dark mode is a near-black background that some readers find easier in dim rooms and harder in bright ones; try both for a week before committing.

Speech - voice mode is a separate setup

Under Settings -> Speech you can pick a voice for ChatGPT's voice mode and set the playback speed. Voice mode is a separate way of using the product (you talk, it talks back), but the speech settings also affect the read-aloud button that appears under every text reply. The default speed is slightly fast for comfortable listening; 0.9x is a kinder starting point.

Data Controls - turn off chat history if you want a clean slate

This is not a reading setting but it interacts with one. If you leave history on, every conversation persists in your sidebar and the sidebar gets long. A long sidebar is one more visual object competing for attention. If you do not need history across sessions, turning it off under Settings -> Data Controls empties the sidebar after each conversation and the page gets noticeably calmer.

Step 2 - the font layer

ChatGPT's web client uses a system sans-serif stack: the OS default sans on Windows and Linux, the SF system font on macOS. The body text sits at about 16 px with a 1.7 line height. That is a perfectly reasonable starting point, but it is not a dyslexia-friendly one - the letters that dyslexic readers most often confuse (b/d, p/q, n/u, lowercase l next to capital I) are the ones that those system fonts handle least distinctly.

A browser-side font-override extension swaps that for whichever face you find easiest. OpenDyslexic, Lexend and Atkinson Hyperlegible all apply cleanly to chatgpt.com. The chat bubbles change face, the sidebar changes face, the model name in the top bar changes face. The only thing that does not change is code blocks and inline code, which stay monospace - and that is the right behaviour, since you want to read code in a face designed for code (see reading code with dyslexia for a separate fix there).

LexiFont applies OpenDyslexic free in one click; LexiFont Pro unlocks Lexend, Atkinson Hyperlegible and Comic Neue for $14.99 one-time if OpenDyslexic's weighted-bottom letterforms feel too busy.

ChatGPT reply - default system font

Here is a short summary of the differences between the two libraries. The first is older, more battle-tested, and has wider community adoption. The second is newer but has a more ergonomic API and better TypeScript support.

Same reply - Atkinson Hyperlegible

One caveat about ChatGPT's desktop app - it is a packaged version of the web app and, like Slack's desktop client, it does not honour browser extensions. If you want the font swap, you need the browser tab. Pin chatgpt.com in Chrome (right-click the tab, choose Pin) and treat that as your primary surface. The desktop app can stay installed for its global keyboard shortcut, but the actual reading happens in the browser.

Step 3 - the upstream fix: ask for a friendlier shape

This is the lever ChatGPT has that no other reading surface offers. You can ask the model to write the way you find easiest to read, and it will - consistently, across every conversation, if you put the instruction in the right place.

Custom Instructions (under Settings -> Personalization -> Custom Instructions) is a one-time configuration that prepends a system prompt to every conversation. It has two boxes: "What would you like ChatGPT to know about you?" and "How would you like ChatGPT to respond?" The second box is where the reading shape lives. A simple instruction that works:

When you reply, please: - start with a one-sentence summary so I can decide whether to read the rest - prefer short paragraphs (2-3 sentences) over long ones - prefer plain words over Latinate words (use "after" not "subsequently", "also" not "furthermore", "but" not "however") - avoid nested bullet lists deeper than one level - when you must use a list, keep each bullet to a single line - if the answer is long, end with a "TL;DR" of three short bullets I have dyslexia and these choices make your replies easier for me to read. Do not announce the instructions back at me.

That is one paste, and it shapes every reply from then on. The model is consistent enough that you can rely on the top-line summary appearing every time, which lets you decide in three seconds whether the full answer is worth reading. The "do not announce the instructions back at me" line matters - without it, some model versions open every reply by acknowledging the constraints, which is the opposite of helpful.

You can vary the instruction for specific tasks by writing it into the conversation itself: "Write this in short paragraphs with a top-line summary" at the start of a chat overrides Custom Instructions for that thread. Most readers find one global setting and the occasional override is the right balance.

A note on the "I have dyslexia" framing - you do not have to disclose it. The shape instructions work the same without that sentence; we include it in the example because some readers find the model holds the instructions more reliably when there is a stated reason. Either works.

Step 4 - shape long answers further with one follow-up

Even with Custom Instructions, the model sometimes produces a long answer when you wanted a short one. The single most useful follow-up prompt for dyslexic readers is also the shortest:

Rewrite that as three short paragraphs, plain words.

The rewrite is usually half the length, with the structural padding removed. If you want it shorter still:

Now in one paragraph.

Treat the first reply as a draft. The follow-up is where the readable version lives. This is the cadence most dyslexic readers settle into after a couple of weeks - ask, skim the summary, ask for a rewrite if the summary suggests there is something worth reading, then read the rewrite. It costs one extra round trip and saves five minutes of effort.

Step 5 - pair it with text-to-speech

Long ChatGPT replies are a great use case for TTS, partly because the page already has a built-in read-aloud button (the little speaker icon under every reply) and partly because the model writes in a register that synthesised voices handle well - no slang, predictable sentence structure, few proper nouns.

Two options:

Built-in read aloud. The speaker icon under any ChatGPT reply plays it through OpenAI's own voice synthesis at the speed and voice you set in Settings -> Speech. The quality is high - genuinely indistinguishable from a human reader for short replies - and you can listen with your eyes closed or with the text in front of you to read along. Listen-while-you-read is the strongest setup, since the audio carries you through the sentences your eyes would otherwise stall on.
Chrome's Read aloud. If you want to use the same TTS workflow across every site (Wikipedia, news articles, your inbox, ChatGPT), Chrome's built-in Read aloud works on ChatGPT too. The voices are a little less natural than OpenAI's, but the consistency across surfaces is the point.

For very long answers (anything over about 600 words), TTS at 0.9-1.0x with the font swap on the page is the sweet spot. You are reading the page in a dyslexia-friendly face and hearing it in a calm voice; the two streams reinforce each other, and the comprehension cost drops sharply compared to silent reading alone.

Step 6 - the same recipe applies to Claude, Gemini, and the rest

This article uses ChatGPT as the worked example because it is the most-used chatbot, but every move here transfers:

Claude (claude.ai) - has a similar Settings panel; turn off streaming under Settings -> Appearance. The font swap applies cleanly. Custom Instructions are called System prompt in the API and Project instructions on the web; the prompt above works without modification.
Gemini (gemini.google.com) - lives inside the Google account, so any account-wide reduce-motion or font-size preference you have set applies. The font swap applies cleanly. There is no per-account custom instruction equivalent yet, so you paste the shape prompt at the start of each conversation.
Perplexity, Copilot, Mistral Le Chat, etc. - same shape. Pin the tab, turn streaming off if available, run the font extension, paste the shape prompt if there is no custom-instruction slot.

The "ask the model to write in your shape" lever is the reason chatbots are quietly the most accessible long-form reading surface on the web for dyslexic readers - more so than news sites, more so than Wikipedia, more so than PDFs. You can renegotiate the writing style every time, which is a thing you can do with no other publisher.

Common mistakes to skip

Using the desktop app for serious reading. The desktop ChatGPT app is a packaged web view and does not honour browser extensions. The font swap does not apply. If you find yourself reading more than three paragraphs of output, switch to the browser tab.
Letting the streaming animation run. Watching the page rebuild a sentence while you try to parse it is the single most exhausting part of ChatGPT for dyslexic readers. The setting to turn it off is one toggle deep.
Asking for "a summary". Models default to a wall of bullets when you ask for a summary. Ask for three sentences or two short paragraphs - the specific shape is what shrinks the reply, not the word "summary".
Trusting the model on names and numbers. Not a reading issue but worth saying: dyslexic readers are sometimes more prone to misread proper nouns and numbers, and chatbots are sometimes prone to fabricating them. Verify any name, date or figure you plan to use in writing.
Pasting whole articles to be summarised without first dropping them into Reader Mode. The model summarises better from clean text. Reader Mode on the source page, then copy, then paste into ChatGPT, gives noticeably tighter summaries than pasting the raw HTML.

Putting it all together

A complete dyslexia-friendly ChatGPT setup, in order:

Open chatgpt.com in a pinned browser tab, not the desktop app.
Under Settings -> Personalization, turn off the streaming animation.
Under Settings -> Appearance, set theme to System.
Install LexiFont and pick OpenDyslexic, Lexend or Atkinson Hyperlegible.
Paste the shape prompt into Custom Instructions once and forget it.
When a reply still comes back long, send "Rewrite that as three short paragraphs, plain words."
For replies over 600 words, hit the speaker icon and listen-while-you-read at 0.9x.
Apply the same recipe to Claude, Gemini and any other chatbot you use - the levers are the same.

Total setup time is around ten minutes. The payoff is a reading surface that, once configured, is more comfortably readable for dyslexic adults than most newspapers - because you can ask the writer to write differently, and the writer always says yes.

Get LexiFont Pro - OpenDyslexic, Lexend, Atkinson Hyperlegible and Comic Neue for $14.99 one-time