My Eyes Started Hurting From Reading Long AI Answers — E-Paper Cuts the Strain

September 18, 2026

Since I started running local LLMs on my own machine, my eyes have started getting tired.

I’m reading a lot more text, and using my eyes a lot more. Ask a question, and the AI comes back with a fairly long answer. Following it with your eyes and checking whether it’s correct — that keeps happening, over and over.

This article works out from actual measurements how much more reading time an AI’s output adds, then lays out the reading-device options. It’s not just about buying a new screen — I also cover what you can do with the hardware you already have.

What This Article Covers

Three things: how much more reading a local LLM adds, what screen options exist for reading, and what you can do without buying anything.

The reading-volume part is calculated from generation speeds I measured on my own RTX 3090. The screen-characteristics part is written from published specs and my own experience using an e-paper tablet. I have not run a measurement comparing eye fatigue across different screens reading the same text.

How Much More Reading Does This Add?

Working from my own measurements: running Qwen3.8-27B on an RTX 3090 (24GB), generation speed came to 32.9–38.0 tokens per second depending on the quantisation level.

Measured September 2, 2026, with llama-bench (Vulkan), 3 runs each, generating 128 tokens.

A token is the unit an AI uses to count text. For Japanese, roughly one character works out to 1–2 tokens; in English, a token is roughly three-quarters of a word.

Running continuously at 35 tokens per second comes to 2,100 tokens per minute — in English, that’s roughly 1,500 words a minute, about six to seven double-spaced pages. Leave it running for 10 minutes and you’ve accumulated 15,000–20,000 words.

Since you end up reading all of that, your eyes are locked onto the screen continuously — and that’s exactly where the fatigue comes from.

Types of Screens on Electronic Devices

Worth a quick refresher here on the common types of screens found on electronic devices.

Both LCD and OLED mean you’re looking directly at a light source. Dimming it reduces the strain, but the fact that it’s emitting light doesn’t change.

E-paper works differently. Particles inside the screen rearrange themselves to form a pattern, and it only becomes visible once ambient light hits it — the same principle as reading ink printed on paper.

What Changes When You Read on E-Paper

I’ve had an e-paper tablet for a long time and used it mainly as an e-reader, but lately I’ve been using it more and more to read the long text generative AI outputs.

My eyes tire less even with extended viewing. Reading the same amount of text for the same amount of time, my eyes are noticeably more comfortable than doing the same thing on an LCD.

It’s also, unsurprisingly, larger than a phone screen and usable anywhere, which makes it genuinely convenient.

Where E-Paper Falls Short

To be upfront about it: it’s useless in the following situations.

• Refresh is slow. Watching a response stream in as it generates makes the display flicker. Not suited to this use
• No color. Syntax highlighting in code, or colors in a chart, become hard to tell apart
• Can’t be read as-is in the dark. You need a model with a built-in front light, or a separate light shining on it

It works better if you treat it as a device purely for reading back what’s already been generated — that’s where it shines most. And of course, pair it with a mobile keyboard and it can handle light work just fine too.

Making the Device You Already Have Easier on Your Eyes

Eyes get tired most when bright light hits them directly. Straining to read small text also naturally wears them down over time.

Even on a light-emitting screen like LCD or OLED, adjusting the display settings can ease the strain somewhat.

• Lower the brightness. The closer your screen brightness is to your room’s ambient light, the less strain on your eyes
• Lower the color temperature. Cut the blue tones at night
• Increase the font size. The smaller the text, the more your eyes have to move to follow it
• Narrow the line width. Don’t stretch text across the full screen — constrain the reading width
• Have the AI answer shorter to begin with. Just adding “give the conclusion first, under 300 words" to your prompt cuts down how much you have to read

That last one might be the most effective. As a root fix, cutting down how much you have to read is probably the quickest lever to pull.

What I Haven’t Confirmed This Time

• I haven’t run a measurement comparing eye fatigue across different screens reading the same text. This is written as a subjective impression
• I don’t own a light meter, so brightness isn’t given as a number
• I don’t own an e-paper phone. What I have on hand is a tablet and an e-reader
• The generation speed is specific to Qwen3.8-27B. A different model would change the speed

Closing

Running a local LLM adds to how much you read. At 35 tokens per second, that’s roughly 1,500 words a minute in English. The machine takes care of the writing side, but the reading side stays a human job.

Reading devices split into three types. LCD and OLED mean looking directly at a light source; e-paper reflects light instead. For the specific task of reading back generated text, e-paper is clearly easier on the eyes. But its slow refresh and lack of color make it a poor fit for a screen you’re writing on at the same time. Add a keyboard, though, and it can handle light work.

And there are things you can do before buying anything: fix your brightness and text size, and have the AI answer more briefly.

What I realized after running AI locally for a while is that I’d been focused entirely on making it run faster. The faster it runs, the more text piles up, and the longer you spend reading it. Which model to pick probably needs to be balanced not just against write speed, but against how much you can actually get through reading.

I’ve written separately about how to choose a model in the first place.

How Do You Actually Choose a Local LLM Model? Five Axes I Found by Running a Doz

The first thing that stumped me when I started with local AI was which model to pick. Open the m ...

 https://kimono-oyaji.com/local-llm-model-choice-5axes-en/