Is Speed Reading Real? What the Science Says

Is Speed Reading Real, and Does It Work?

No. Speed reading as it's sold doesn't work: no controlled study has found a reader sustaining 1,000 or 2,000 words per minute with normal comprehension, and researchers have been testing the claim since the 1960s. The real number is smaller and better established. A 2026 eye-tracking study by Laura Schwalm, Ralph Radach, and Victor Kuperman in Scientific Studies of Reading paced 46 adults through texts at speeds from 180 to 405 wpm and found comprehension stayed stable until the fastest condition, where it fell. Their own summary is that readers can go to roughly 150% of the average rate without sacrificing comprehension, which puts the honest headroom at about 1.5x your comfortable pace, not 10x.

That second half is what most debunkings skip. You probably can read somewhat faster than you do now without losing much. What you can't do is quadruple it, and no eye trick, app, or course moves where that limit sits.

Speed reading is the umbrella name for methods that promise to multiply your reading rate without costing you comprehension. They work, or claim to, by pulling one or more of five levers: widen how much text you take in per glance, silence the inner voice, stop your eyes from looking back, flash words one at a time so your eyes don't move at all, or reformat the text itself, the way Bionic Reading bolds the first half of every word. Each lever runs into a hard fact about how reading works.

Why does speed reading still feel like it works for some people? Usually because they're doing something else and calling it reading. They're skimming, previewing, or skipping. Those are real and useful skills. They're just not the same activity as reading for understanding, and pretending they are is where people get burned.

---

What the 2026 Research Changed

For years the standard debunking rested on one landmark paper. Keith Rayner and four colleagues published "So Much to Read, So Little Time" in Psychological Science in the Public Interest in 2016, a review of the entire science of reading aimed squarely at speed-reading claims. Their conclusion was blunt: there's a trade-off between speed and accuracy, and it's unlikely readers can double or triple their speed, going from around 250 to 500 or 750 wpm, while understanding the text as well as they would at a normal pace.

That finding still stands. What changed in 2026 is our picture of the gentler slope below it.

Schwalm, Radach, and Kuperman asked a question the older literature had mostly assumed the answer to: is your natural reading rate the same as your comprehension-optimal rate? Forty-six native English-speaking university students, mean age 19.82, read texts at a range of imposed speeds using a line-by-line text enhancement method to control the tempo. The sample's own average natural rate was 273 wpm.

The results are worth reading carefully:

• Comprehension stayed relatively stable across the tested range, hovering around 57% correct, rather than degrading the moment readers sped up.
• A significant decline appeared at 405 wpm, where comprehension fell to about 49%. The last speed with no significant loss was 360 wpm.
• Eye movements adapted efficiently to the imposed pace, particularly in later measures, and word length and frequency effects held up, meaning readers were still processing language normally rather than just sweeping their eyes across the page.
• Readers who were naturally faster, and those with higher word reading efficiency, maintained comprehension best at elevated speeds.

The authors frame this as challenging the notion of a strict speed-accuracy trade-off. Skilled adult readers can sustain effective comprehension above their preferred rate, because the preferred rate isn't a hard limit, it's a comfortable cruising speed.

Now the parts that matter for how much weight to put on it, which press coverage of this study tends to skip.

405 wpm was the fastest condition they ran. The paper describes it as "the highest speed tested" and calls for future work to pin down whether the real threshold sits between 360 and 405. So 405 isn't a measured ceiling, it's the edge of the evidence. Where the true limit falls above it is nobody's number yet.

The texts were easy. The stimuli were 12 personal narratives from the Rutgers University Oral History Archives, 427 to 671 words each, with a mean Flesch-Kincaid readability of 6.89. That's roughly a seventh-grade level. The paper's own limitation section says it "remains to be seen" whether the same results hold for longer stretches or "more challenging text materials." If what you actually want to read faster is a dense paper or a technical book, this study didn't test that.

The stability was stable at a middling level. Comprehension hovered at about 57% on four-option multiple choice, where chance is 25%. Steady is not the same as high.

Everyone declined eventually. Better readers held a higher curve at every speed, but the drop between 360 and 405 showed up across the sample. And two eye-movement signals moved at exactly that point: fixation probability dropped 11% from 360 to 405, and surprisal effects diminished significantly, both consistent with readers skipping more words rather than reading them faster.

So hold both findings at once. This is good news for "read a bit faster than you currently do on material you find easy." It's still fatal for "read 1,000 wpm with full comprehension." A tested range topping out at 405 wpm and a marketing claim of 1,500 wpm aren't in the same conversation.

---

Average Reading Speed: How Fast Can Humans Really Read?

The average adult reads English nonfiction silently at about 238 words per minute, and anything from about 175 to 300 counts as a normal reading speed. Against those numbers, the marketing collapses.

Marc Brysbaert published a meta-analysis in 2019 in the Journal of Memory and Language that pooled reading-rate data across 190 studies and 18,573 participants. It remains the best single answer to the question. For adults reading silently in English, the averages land around:

• 238 words per minute for nonfiction
• 260 words per minute for fiction
• A typical range of roughly 175 to 300 wpm for English nonfiction

Those are normal-comprehension rates. Reading aloud is slower, around 183 wpm, because you're physically producing speech. Brysbaert also noted these estimates are lower than the figures commonly repeated in both popular and scientific writing, which is part of why inflated speed claims sound plausible to people.

The bottom row is arithmetic, not opinion. A fixation lasts about a quarter of a second, so 1,000 wpm would demand roughly four words per fixation across a foveal window that resolves only a handful of characters sharply. Your eyes never land on most of the page. You can call that reading if you like, but you aren't taking in every idea on it.

Average reading speed by age

Reading speed climbs through childhood and then holds steady for most of adult life. The classic school-age dataset is old: S. E. Taylor photographed the eye movements of 12,143 readers from first grade through college and counted only those who answered at least 70% of the true-or-false comprehension questions correctly. The data were collected in 1960 and published in 1965, and Brysbaert's review reproduces the rates:

Two cautions before you treat those numbers as targets. When Spichtig and colleagues ran a replication designed to stay close to Taylor's study, with 2,203 students in grades 2 to 12 in 2011, they measured substantially lower rates at every grade they tested except second grade, and almost a third of the recordings missed the 70% comprehension bar. And Taylor's college figure of 280 wpm sits well above the 238 wpm adult average from the meta-analysis. Read the table as the shape of the climb.

At the other end of life, silent reading slows, but Brysbaert's review doesn't expect the decline to start before about 60 to 65 in healthy readers. Older readers make longer fixations, and several studies find them leaning harder on prediction, skipping more words and then looking back more often when a guess turns out wrong, though not every study finds that pattern.

If your own number comes out low, that isn't a verdict on how smart you are. Our piece on slow reading versus being a slow reader unpacks the difference.

Average reading speed by language

Words per minute is an English-shaped yardstick. Languages use different numbers of words to say the same thing, so wpm figures from different languages aren't directly comparable. Brysbaert found at least two studies of both silent reading and reading aloud for twelve languages:

The last column is the one to watch. Brysbaert found that languages needing more words to carry a message tend to post higher wpm, and languages needing fewer, longer words post lower wpm, with a correlation of r = 0.55 for silent reading across these 12 languages and 0.82 for reading aloud. His explanation: "the amount of information transferred seems to be more important than the number of words needed to do so." So if you read in Finnish, an English benchmark of 238 wpm can make you look slower than you are. Reading in a second language does cost real speed, though: Dutch-English bilinguals who read half a novel in Dutch and the other half in English read 17% slower in English (Cop, Drieghe, and Duyck, 2015).

A few notes on the table itself. The English row comes from English-language studies alone (144 of silent reading and 27 of reading aloud), so it reads 236 and 190 rather than the headline 238 (silent) and 183 (aloud), which pool the English studies with those in other languages written in the Latin alphabet. The Chinese figures were often estimated from character counts at about 1.5 characters per word, which can overstate the rate. Portuguese and Japanese didn't have enough studies to make the table; for Portuguese, native speakers reading the standardized International Reading Speed Texts (IReST) aloud averaged 181 wpm, against 228 wpm for English speakers on the same short, simple texts (Trauzettel-Klosinski and Dietz, 2012).

How to test your reading speed

You'll learn more from your own number than from any average. Glasp's free reading speed test takes about two minutes and checks your comprehension with three questions. To test yourself on your own reading material, run this at home:

1. Pick a fresh text of at least 1,200 words, the kind you actually read. Easy fiction will flatter you and a textbook will humble you, so test on your real diet.
2. Read it normally with a timer running. Don't push. You're measuring your cruising speed, not your top speed.
3. Divide words by minutes. 1,500 words in 6 minutes is 250 wpm.
4. Check comprehension before you trust the number. Write down the main points from memory or answer a few questions about the text. Taylor only counted readers who passed a comprehension check, and the same rule applies to you: if you can't recall most of the main points, the speed doesn't count.
5. Repeat on two or three different texts and average them, because a single passage swings the result.

Then use the same check whenever you try a faster pace. If your comprehension collapses, the faster number was skimming.

---

The Eye-Movement Reality: Foveal Vision and Regressions

To understand why speed has a limit at all, look at how the eye takes in text. Your eyes don't glide smoothly across a line. They move in quick jumps called saccades, pausing on words in brief fixations of roughly 250 milliseconds each.

The catch is that you only see fine detail in a tiny part of your vision called the fovea. The high-acuity foveal window is small, on the order of a handful of characters at a time. Outside that narrow center, text gets blurry fast. You can pick up a little information from the next few letters, in a region called the parafovea, but you can't read a whole line, let alone a whole paragraph, in a single glance.

This is the wall that "read a block at a time" and "see the whole page at once" techniques run into. The biology doesn't allow it. If you try to take in large areas at once, you aren't reading those areas, you're guessing at them.

Then there are regressions, the backward eye movements to words you already passed. Speed-reading courses love to frame these as a defect to train away. The science says otherwise. Most regressions repair comprehension: they happen when your brain detects an ambiguity, a parsing problem, or a connection it needs to nail down, and it sends your eyes back to fix it. Some are just oculomotor error correction. Suppress the lot and you don't get faster reading, you get worse understanding, because you've disabled a self-correction system.

The 2026 data adds a wrinkle worth knowing. When readers were pushed to go faster, their eye movements adapted intelligently rather than falling apart, and normal word-level processing continued. So the visual system has more flexibility than "every saccade is sacred" implies. What it doesn't have is the ability to read text it never fixated on. At 405 wpm, fixation probability fell 11% against the 360 wpm condition, which is the mechanism of the comprehension drop showing up in the eye data: past a point, going faster means landing on fewer words.

Either way, the two pillars of most speed-reading methods, "expand your visual span" and "stop looking back," are still working against how reading functions.

---

Is Subvocalization Bad? No.

Subvocalization is the quiet inner voice you "hear" as you read. Speed-reading programs treat it as public enemy number one: silence the voice, they say, and you break free of speaking-speed limits.

It's a tidy story. It's also mostly wrong.

Subvocalization isn't an inefficient verbal tic bolted onto reading. It reflects a phonological process tied to how you understand language, especially anything complex. The inner voice helps you hold words in working memory, parse difficult sentences, and keep track of who did what to whom. For easy material you can lean on it less. For dense or syntactically tricky text, it's doing real work: attempts to eliminate inner speech have been shown to impair comprehension when texts are reasonably difficult and require readers to make inferences.

People can reduce subvocalization to a degree, and for light skimming that's fine. But trying to fully suppress it as a path to 1,000 wpm tends to backfire. Either you fail to suppress it, so the "technique" does nothing, or you succeed and comprehension falls, because you've knocked out part of the machinery that builds meaning.

The takeaway: don't wage war on your inner voice. It isn't slowing you down so much as it's helping you understand.

---

Debunking RSVP, Bionic Reading, Skimming, and 3x Apps

A whole category of apps promises 3x speed through Rapid Serial Visual Presentation, or RSVP. RSVP flashes words one at a time in a fixed spot, so your eyes never move. The pitch is that by killing saccades, you remove "wasted" time.

The flaw is the one from the eye-movement section. Those eye movements aren't waste. Saccades and regressions are how your brain navigates text, re-checks ambiguous passages, and controls its own pace. RSVP takes that control away. Rayner and colleagues made the point directly: these technologies don't let people reread the text to correct misunderstandings, so readers would maintain their misinterpretations as they're pushed forward and comprehend the text less well. There's no review button for a sentence that already flashed by.

Notice how this interacts with the 2026 finding. Readers pushed to a faster pace held comprehension partly because their eye movements adapted. RSVP removes exactly the mechanism that made that adaptation possible. Going faster in a format that still lets you control your eyes is a different proposition from going faster in one that doesn't.

Bionic Reading tries something gentler. Instead of flashing words at you, it bolds the first half of each one, on the theory that your eyes will catch the bold anchor and your brain will fill in the rest. It turns up as a toggle in reading apps and browser extensions, and the pitch is familiar: faster reading, same understanding.

The research has caught up with it. Joshua Snell timed readers on 100 paragraphs shown either in Bionic style or in normal type and found no significant difference in reading times (Acta Psychologica, 2024). Then in September 2026, Adam Parker and three colleagues published a registered report in Royal Society Open Science that tracked the eye movements of 90 skilled adult readers through sentences that were unbolded, bolded at the start of each word, or bolded in the middle. The bolding did change something: it pulled the first fixation closer to the start of long words. It just didn't pay off. The authors found no evidence of differences in word skipping, word reading times, overall sentence reading times, or comprehension, and concluded that selective bolding doesn't make word identification more efficient for skilled adult readers.

The study covered skilled adults reading single sentences, so it doesn't settle whether bolding helps people with dyslexia or other reading difficulties. If you like how bolded text looks, use it. Just don't expect it to raise your speed.

Skimming gets a similar overclaim. Skimming is real and genuinely useful. But skimming isn't reading. When you skim, you sample: headings, first sentences, keywords, the occasional full line. You build a rough map of the gist. You don't absorb the argument, the evidence, or the details. Calling fast skimming "speed reading with full comprehension" is the central sleight of hand, because it quietly swaps the definition of reading.

Here's the claim-versus-reality scorecard:

None of this means the apps are useless for everything. They're fine for breezing through easy email or low-stakes text. They just can't deliver the headline promise.

---

ChatGPT and Speed Reading: What AI Can and Can't Do

Can ChatGPT make you read faster? Not in the sense speed-reading courses mean. ChatGPT doesn't touch your eye movements, your foveal span, or your subvocalization, so it can't move your reading rate by a single word per minute. What it changes is which pages get your reading minutes, and that's a real gain. What it can't do is the reading itself, and two 2025 studies put numbers on what happens when you ask it to.

Shiri Melumad and Jin Ho Yun, writing in PNAS Nexus in October 2025, ran seven experiments with 10,462 participants. People learned about topics like planting a vegetable garden or handling financial scams, either through an LLM synthesis or through ordinary web search links. Those who learned from the LLM developed shallower knowledge. When they later gave advice on the topic, it was sparser, less original, and less likely to be adopted by the people who received it. The mechanism is the interesting part: it isn't that the AI was wrong. It's that receiving a pre-digested synthesis removes the work of discovering and connecting sources yourself, and that work is where depth comes from.

Hudson Etkin and colleagues, in Frontiers in Education in 2025, ran a pre-registered randomized crossover trial with 195 college-aged participants reading ACT-derived passages under four AI conditions: generated summaries, generated outlines, a question-and-answer tutor chatbot, and a Socratic discussion chatbot. The headline result is a split:

• AI tools significantly improved comprehension for lower-performing readers. The Socratic chatbot helped them most (d = 0.86).
• AI tools significantly worsened comprehension for higher-performing readers. The summary tool hurt them most (d = 0.83).

Read that second bullet again if you consider yourself a capable reader. The single most popular AI reading behavior, "just summarize it," was the most harmful intervention for the strongest readers in the study. The likely reason is the obvious one: summaries strip out the detail and nuance a good reader would have extracted anyway, and hand back less than that reader could have gotten alone. One caveat on the study: participants were split into high and low performers by a median split on a short test, a design that tends to exaggerate differences between the groups, so treat the split as directionally real rather than precisely sized.

A behavioral finding fits alongside this. An eight-week study of 15 undergraduates using AI on real course readings described them as "reading through AI rather than with it," using generated summaries as the primary material and letting them decide which sections deserved attention. It's a small, not-yet-peer-reviewed preprint, and the AI use was assigned rather than spontaneous, so treat it as a description rather than evidence. The pattern will still look familiar to anyone who has done it.

Three prompts that use ChatGPT the way the evidence supports. The pattern is to work on either side of the reading, never in the middle.

Before, to decide whether to read at all: "Here is an article. In five bullets, tell me what claims it makes and what evidence it offers for each. Don't summarize the prose. If it makes a claim without evidence, say so."

Before a hard text, to build the scaffold: "I'm about to read this paper. List the six terms or concepts I need to already understand, and define each in one sentence. Don't explain the paper."

After, to test yourself instead of rereading: "Ask me eight questions about this text, one at a time, hardest first. Don't give me the answers until I've tried."

The one to avoid is "summarize this so I don't have to read it." That's the exact intervention that cost higher-performing readers d = 0.83 in the Etkin trial. If you want to see how the current tools compare for the triage job specifically, we tested them in best AI summarizer tools.

---

Does Watching Videos at 1.5x or 2x Speed Hurt Learning?

Speed reading's modern cousin is speed watching: lectures, podcasts, and YouTube explainers at 1.5x or 2x. Here the news is decent up to about 1.5x, and the reason fits everything above. Brysbaert's review notes that normal reading rates are in line with maximum listening speed, which suggests that ordinary silent reading already runs at about the same language-processing limit that caps listening. Speech starts well below that limit. Audiobooks are narrated at roughly 140 to 180 wpm, under the 238 wpm of silent reading, so you can speed up a speaker quite a bit before you hit the bottleneck.

Where's the ceiling for listening? A 1969 review of accelerated-speech research by Foulke and Sticht concluded that listening comprehension "declines at a slow rate as word rate is increased, until a rate of approximately 275 wpm is reached, and at a faster rate thereafter." Do the arithmetic for a speaker at 150 wpm: 1.5x is 225 wpm, comfortably under that bend, while 2x is 300, just past it, and 2.5x is 375.

Modern lecture-video studies roughly line up with that math. A 2025 meta-analysis in Educational Psychology Review by Theepan Tharumalingam and colleagues pooled 110 effect sizes from 24 studies. The cost of speeding up a lecture was small and not statistically significant at 1.25x and 1.5x, small to medium and significant at 2x, and large at 2.5x. The best-known single study is more generous. Dillon Murphy and colleagues at UCLA, publishing in Applied Cognitive Psychology in 2022, found minimal costs at 1.5x and 2x on tests given right away and a week later, with performance declining only beyond 2x, even though 2x pushed their lecturers past 300 wpm. They also found that watching twice at 2x, back to back, did no better than watching once at 1x, but students who rewatched at 2x immediately before the test did better than students who had watched once at normal speed a week earlier. Murphy's team adds its own caveat: results may differ for videos with different speech rates, harder material, or a different balance of audio and visuals.

Here's what the common playback speeds do to your watch time:

The formula is length divided by speed, so a 100-minute lecture at 1.5x takes 66 minutes and 40 seconds, and a 20-minute video at 1.5x takes 13 minutes and 20 seconds.

How to use this without fooling yourself:

• Default to 1.5x, and treat 2x as a deliberate trade. That's where the pooled evidence puts the line: in the authors' own illustration, a class averaging 75% on a test would lose about 2 points at 1.5x, about 7 at 2x, and about 17 at 2.5x. If you go to 2x, plan to review afterward.
• Spend the saved time on review, not on another video. The UCLA authors suggest using it for more studying or a second viewing shortly before you need the material.
• Triage before you press play. Skim a summary first, from YouTube Summary or any of the YouTube summarizer tools we compared, then use the transcript to jump to the part you need instead of scrubbing. Highlight the lines worth keeping so you have something to come back to.

---

What Actually Works Instead of Speed Reading

If you can't buy your way past the limit, what do you do when there genuinely is too much to read? You change the goal. Instead of reading everything faster, read less but better, and get strategic about what earns your slow attention.

1. Match your reading rate to your purpose. Not everything deserves 238 wpm of careful attention. A news blurb, a contract, and a textbook chapter call for different gears. Decide what you're reading for before you start. Looking for one fact? Scan. Need to truly learn it? Slow down and accept that slow is correct here. The skill isn't a faster top speed, it's choosing the right gear.

2. Claim your actual headroom, on the right material. The 2026 study is a genuine invitation: your comfortable pace probably isn't your comprehension limit. Measure your baseline with the test above, then try pushing familiar, moderately easy material toward 300 to 350 wpm and check yourself afterward by summarizing what you read from memory. If your summary holds up, that pace is yours. If it doesn't, back off. Do this with material you could already handle, since that's the only kind the study tested, and don't treat 405 as a target.

3. Use skimming as triage, not as reading. Skim to decide whether something deserves a real read. Run your eyes over the structure, the headings, the first and last lines of sections. That gives you the map. Then either read it properly or move on. Skimming is excellent at this job and terrible as a substitute for the job of reading.

4. Preview before you read. A few minutes on the table of contents, the abstract, the subheadings, and the conclusion primes your brain for what's coming. You read the full text faster and understand it better afterward, not because your eyes move quicker, but because you already know the shape of the argument.

5. Read actively, because retention is the real throughput. This is the big one. Speed obsession measures the wrong thing. What matters isn't words per minute but ideas retained per minute. If you "read" 1,000 wpm and keep 10%, you've effectively read 100 wpm of usable material. If you read 238 wpm and keep most of it, you're far ahead. Active reading, meaning highlighting the load-bearing ideas and then recalling and reviewing them, is what raises your effective throughput.

This is where Glasp's web highlighter earns its place. Highlighting as you read forces you to decide what actually matters, which is itself a comprehension boost, and it leaves a durable record instead of a vague memory. For why that beats passive reading, see the science of highlighting.

For the triage step, AI is a legitimately fast tool as long as you use it to decide what to read rather than as a replacement for reading what matters. Drop a long video into YouTube Summary to find out in seconds whether it's worth your 40 minutes. Use Glasp's AI chat to interrogate a piece before committing to it. That's the honest version of fast: a quick scan to triage, then real reading on what survives the filter.

And because retention is the goal, close the loop. Export your highlights and revisit them, ideally by testing yourself rather than rereading. Our guides on how to remember what you read and active recall go deep on the second half of that loop.

---

Frequently Asked Questions

Is speed reading real, or is it a myth?

The version sold by courses and apps is a myth. You can't read 1,000 or 2,000 wpm while keeping full comprehension, and no trained eye technique changes that. But "you can never read faster than you do now" is also wrong. A 2026 eye-tracking study by Schwalm, Radach, and Kuperman found comprehension stayed stable up to 360 wpm and fell at 405 wpm, the fastest speed they tested. Their summary is that readers can reach about 150% of the average pace without sacrificing comprehension. So there's real headroom of roughly 1.5x, on easy material. There isn't 10x.

Does speed reading work?

Not as advertised. Rayner and colleagues concluded in their 2016 review that readers are unlikely to double or triple their speed, from around 250 to 500 or 750 wpm, without losing comprehension. What people call "speed reading" is usually skimming or skipping, which are useful skills but aren't the same as reading for understanding. Pushing modestly above your comfortable pace on familiar material is the part that actually works.

Can you actually read 1,000 words per minute?

Not with comprehension. 1,000 wpm is more than four times the 238 wpm adult average and roughly 2.5 times the 405 wpm point where measured comprehension dropped. At that rate you'd need about four words per fixation, while the high-acuity foveal window resolves only a handful of characters sharply, so most of the page never gets fixated at all. People who hit 1,000 wpm on a stopwatch are skimming, and skimming scores far lower on comprehension tests than normal reading.

How fast can humans really read?

For adults reading silently in English with normal comprehension, the averages are around 238 wpm for nonfiction and 260 wpm for fiction, with a typical nonfiction range of roughly 175 to 300 wpm. Those figures come from Brysbaert's 2019 meta-analysis of 190 studies and 18,573 participants. Under imposed pacing on easy text, comprehension has been shown to hold up to 360 wpm and decline at 405. The fastest competent readers aren't far above average, and their edge comes mostly from word reading efficiency and vocabulary rather than a special technique. Figures in other languages differ, partly because some languages need more words than others to say the same thing.

What is the average reading speed by age?

In Taylor's classic data on more than 12,000 American students, silent reading with comprehension rose from about 80 wpm in first grade to 158 in fourth, 204 in eighth, and 250 in twelfth, with college students at 280. A 2011 replication measured lower rates at every grade it tested except second grade, so treat Taylor's figures as a rough curve. Healthy adults from 18 to 60 average about 238 wpm on nonfiction and 260 on fiction, and the slowdown that comes with age isn't expected before about 60 to 65.

Can ChatGPT help you speed read?

It can't change your reading rate, because it doesn't touch your eyes or your inner voice. It's genuinely useful on either side of the reading: before, to decide whether a piece deserves your time, and after, to quiz you on what you read. What doesn't work is asking it to summarize so you can skip the text. In a pre-registered randomized trial of 195 participants, AI-generated summaries significantly worsened comprehension for higher-performing readers, the most harmful of four tools tested (Etkin et al., 2025). A separate seven-experiment study of 10,462 people found that learning from AI syntheses produced shallower knowledge than ordinary web search (Melumad and Yun, 2025).

Is subvocalization bad?

No. The quiet inner voice you hear while reading is part of how your brain processes language, and it helps with working memory and parsing difficult sentences. You can lean on it less for easy text, but trying to suppress it entirely usually hurts comprehension rather than buying you speed. Don't treat it as a habit to eliminate.

Is looking back while reading a bad habit?

No. Those backward eye movements, called regressions, mostly support comprehension. They happen when your brain spots an ambiguity or a connection it needs to resolve, and it sends your eyes back to fix it. Training yourself to never look back doesn't make you a faster reader, it removes a built-in self-correction system and degrades understanding.

Does Bionic Reading work?

Not for speed or comprehension, as far as the evidence goes. A 2024 study found no significant difference in reading times between Bionic and normal text across 100 paragraphs (Snell, 2024). A registered report published in September 2026 tracked the eyes of 90 skilled adult readers and found the bolding shifted where they first looked in long words, but found no evidence that it changed word skipping, reading time, or comprehension (Parker et al., 2026). It only tested skilled adults, so whether bolding helps readers with dyslexia is still an open question.

Is watching lectures at 2x speed bad for learning?

Somewhat. A 2025 meta-analysis of 24 lecture-video studies found that 1.25x and 1.5x cost very little, while 2x carried a small-to-medium cost on tests of the material and 2.5x a large one (Tharumalingam et al., 2025). The best-known single study, from UCLA, reported minimal costs at 2x and found that rewatching at 2x right before a test beat a single 1x viewing a week earlier (Murphy et al., 2022). If you speed up, 1.5x is the safe default, and the time you save is best spent on review.

What should I do instead of speed reading?

Read less but better, and triage hard. Match your reading rate to your purpose, test whether you have headroom on easy material, skim only to decide what deserves a real read, preview before you read, and read actively by highlighting key ideas and testing yourself on them later. Effective throughput is ideas retained per minute, not words per minute.

---

Conclusion

Speed reading sells a fantasy: that comprehension is free if you just learn the trick. The science has said otherwise for decades, and the 2026 evidence doesn't rescue the fantasy. What it does is soften the edges. You have perhaps 1.5x of headroom above your comfortable pace on material you find easy, the readers who use it got there through practice and vocabulary rather than a technique they bought, and past about 400 wpm the eye-tracking data shows people skipping words rather than reading them faster.

The useful reframe is to stop optimizing words per minute and start protecting ideas retained per minute. Take the headroom you have. Video gives you about the same room: lectures held up to about 1.5x across 24 studies, with a real cost at 2x, so put the minutes you save into review. Triage ruthlessly so your slow attention lands on what deserves it. Be skeptical of handing the reading itself to an AI, especially if you read well already, because the measured cost falls hardest on exactly those readers.

That's the loop Glasp is built for: triage with YouTube Summary, highlight what survives the filter, and export it for review. Real fast isn't reading everything quickly. It's keeping what counts. Start reading better with Glasp.