Subtitles: "Captions Benefit Everyone" Is Not What the Evidence Says

Something has changed about how people watch television. A YouGov survey of 1,000 US adults, fielded in July 2023, found that 63% of under-30s prefer to watch with subtitles on, against 29% of 45-64s and 30% of over-65s. Overall, 13% of Americans say they always have subtitles on and another 17% say they do most of the time.

The usual defence of this is a paper title: Video Captions Benefit Everyone. It is cited constantly, and it says what it says.

It is also a narrative review with no inclusion criteria, no search strategy and no pooled effect sizes, and by its own abstract, the benefit is "particularly" for "persons watching videos in their non-native language, for children and adults learning to read, and for persons who are D/deaf or hard of hearing."

That is three groups. Notice which one is missing: a fluent native speaker watching a drama in their own language, which is what most of those under-30s are doing.

This article is about that gap, and about the one piece of evidence that speaks to it, which does not say what you would expect.

The evidence that is genuinely strong is about someone else

The captions literature is large and the effects are big. They are also, almost without exception, second-language studies.

Meta-analysis	Population	k	Effect
Montero Perez, Van Den Noortgate & Desmet (2013)	L2 learners	18	Listening comprehension g = 0.99, 95% CI [0.598, 1.377]; vocabulary g = 0.87, 95% CI [0.578, 1.154]
Alotaibi, Mahdi & Alwathnani (2023)	L2 classrooms	26 studies, 34 effect sizes	d = 0.69, 95% CI [0.371, 1.021]
Reynolds, Cui, Kao & Thomas (2022)	L2 vocabulary from video	20 in the meta-analysis	d ≈ 0.87; intralingual captions d = 1.73-1.81

Those are among the larger effects in educational research. If you are learning a language, subtitles in that language are one of the best-evidenced things you can do.

Montero Perez and colleagues explicitly excluded studies using L1 subtitles. Alotaibi and colleagues included no native-speaker studies. These findings do not transfer, and the papers do not claim they do.

The literacy evidence is similarly strong and similarly not about you. Kothari and Bandyopadhyay tracked same-language subtitling of Bollywood film songs on Indian television across five years, with a random sample of 7,409, and found that regular viewers showed "substantially greater mean improvement on all the indicators of literacy skill" than non-viewers. That is a striking result about people learning to read.

The one meta-analysis that does cover native speakers

There is a body of work that speaks to the actual question, and it is filed under a different name: verbal redundancy, what happens when you hear speech and simultaneously read the same words.

Adesope and Nesbit (2012) meta-analysed 57 independent studies with 3,452 participants, most of them postsecondary students working in their own language.

Comparison	Effect
Overall	g+ = 0.15, 95% CI [0.08, 0.22]
Speech + text versus speech alone	g+ = 0.29, 95% CI [0.20, 0.39]
Speech + text versus text alone	g+ = -0.04, 95% CI [-0.14, 0.06]

So adding on-screen words to spoken words produces a real benefit, and it is small: g = 0.29. Against reading alone, it produces nothing at all.

That is the honest headline number for a native speaker. Not the 0.99 from the language-learning literature. About 0.29, and only against audio on its own.

And then it gets more complicated

Yue, Bjork and Bjork ran two experiments with mostly-native English speakers on a narrated lesson, comparing on-screen text that was identical to the narration against text that was slightly reworded.

The identical text, which is exactly what a subtitle is, performed worse.

Experiment	Recall with abridged text	Recall with identical text	Statistic
Experiment 1 (N = 105)	.39 (SD .15)	.25 (SD .14)	p = .002, d = 0.95
Experiment 2 (N = 137)	.45 (SD .19)	.31 (SD .19)	p = .008, d = 0.70

And the metacognitive twist: roughly half of participants preferred the identical text; only about a quarter preferred the version that produced better recall. Their paper's subtitle is An Undesired Desirable Difficulty.

One caution before you generalise: the material was a narrated animated lesson about the life cycle of a star, not a drama with actors. The authors did not claim it transfers to Netflix, and neither do we.

Why it matters that reading subtitles is not optional

Bisson, van Heuven, Conklin and Tunney eye-tracked 54 native English monolinguals watching a 25-minute cartoon under four conditions.

The condition to look at is English audio with Dutch subtitles, a language the viewers did not speak, attached to a soundtrack they understood perfectly. They had no reason to read them at all. They still spent 786 ms per subtitle and fixated them 3.30 times. In the standard condition, fixations ran at almost one per word.

The authors discuss the finding in terms of "automatic reading behavior." Subtitles are not an optional overlay you consult when you need them. If they are on screen, you are reading them.

That is what makes the verbal-redundancy result matter. You are not choosing between "watch normally" and "watch with a safety net." You are choosing between watching and reading-while-watching.

Why everyone turned them on

The behaviour has a supply-side explanation that has nothing to do with attention research.

Broadcast engineers writing for the European Broadcasting Union in 2022 attribute declining dialogue intelligibility to production and playback changes: budget pressure meaning "the use of a second boom operator is now almost a thing of the past"; multi-camera shooting that "causes the background noise level for the scenes to increase... the boom is pushed back, replaced by wireless clip-on microphones"; downmix loudness deviations "of up to 3 LU"; and, on the receiving end, that "most receiver devices simply ignore the set downmix preference." Their example table shows Fast & Furious dialogue sitting at -4 LU against a programme loudness of 0.

That document also asserts that poor speech intelligibility ranks first among viewer complaints, but supplies no survey numbers for it, so we are not repeating that as data.

The plausible story is that people turned subtitles on because dialogue got harder to hear, and then kept them on because reading them became automatic. Both halves of that are documented. The causal link between them is not.

The gap, stated plainly

We searched six different ways for the study this article needs: native speakers, native-language audio, same-language subtitles on versus off, measuring comprehension or memory of the content.

We did not find one.

Everything available turned out to be about second-language learners, deaf and hard-of-hearing viewers, sound-on versus sound-off comparisons, or narrated instructional animations. The most relevant recent study, 161 participants watching English video with English subtitles, compared sound on against sound off, with subtitles present in both. It found comprehension higher with sound (81.30 vs 77.70, F(1,158) = 7.050, p = .009) and effort dramatically lower, which tells you subtitles do not fully replace audio. It cannot tell you whether adding subtitles to working audio helps.

A separate study of 125 non-native English speakers watching educational video found no effect of subtitles whatsoever, a difference of 0.04 grade points, Cohen's d = 0.02, with a credible interval spanning zero, and a Bayes factor favouring the model without subtitles by 10.30 times. Complexity and language proficiency mattered; subtitles did not.

So should you turn them off?

Probably not, and here is the honest reasoning rather than a recommendation dressed up as one.

If you are learning the language, keep them on. That case is settled and the effects are large.

If you cannot hear the dialogue clearly, keep them on. Subtitles are unambiguously better than missing half the lines, and the audio problem is real and documented.

If you are a fluent native speaker with clear audio, the honest answer is that nobody has measured it. The nearest evidence points two ways: a small positive effect for speech-plus-text over speech alone, and a substantial negative effect specifically for text that duplicates the speech word for word. Which of those applies to a television drama is not known.

What is worth knowing is the pattern underneath. In the one place where preference and performance were measured together, people preferred the arrangement that produced worse recall. That is the most reliable finding in this whole literature about people like you: not that subtitles help or hurt, but that your sense of whether they are helping is not evidence that they are.

This is the video half of a question the site covers in two other places: reading on a screen versus on paper asks whether the medium changes what you take in, and watching lectures at 2x speed asks what changing the rate costs.