Does Duolingo Work? Yes, Roughly as Well as a Beginner Class - and the Famous Number Is Not Evidence

The short answer is that it works about as well as a beginner language class, on the skills it actually trains, and that the specific quantified claim you have probably heard is not supported by anything that would count as evidence.

Those two sentences point in opposite directions, which is why this article exists. The case for Duolingo is better than its most-cited study. The case for that study is very poor.

The claim, and where it comes from

You have seen some version of it: 34 hours of Duolingo equals one university semester of Spanish.

That number has one source. Vesselinov and Grego (2012), “Duolingo Effectiveness Study,” December 2012. Its own conclusion is that it would take “between 26 and 49 hours (or 34 hours on average) to cover the material for the first college semester of Spanish.”

Here is what produced it.

The report states plainly: “This study was funded by Duolingo.”

It has no control group. Nobody was compared to anyone learning another way.

Its sample collapsed. 727 people viewed the invitation page. 556 completed the entry survey. 196 completed the initial placement test. 88 were analysed. That is a loss of 84% between entry survey and analysis.

It was never peer-reviewed and never published in a journal. It exists as a PDF, and the PDF is hosted by the company that paid for it.

The arithmetic is an extrapolation. Participants averaged 22.0 hours of study (SD 20.4) and gained an average of 91.4 placement-test points, which the report converts to 8.1 points per hour and then projects onto a semester’s material.

Stephen Krashen’s critique of the report notes that the sample was unrepresentative - average age 35, 69% college graduates - that the variability was extreme, with a quarter of participants doing eight hours or fewer and one doing 133, and that the median gain was 3.9 points per hour against the mean of 8.1. A mean that is more than double the median is describing a very skewed distribution.

Note also the pattern around the report. The same lead author produced identically structured commissioned efficacy studies for Babbel and Pimsleur. This is a vendor-evaluation service, not an independent research programme.

The peer-reviewed follow-up, and who wrote it

The claim acquired academic respectability from Jiang, Rollinson, Plonsky, Gustafson and Pajak (2021) in Foreign Language Annals, which reported that Duolingo learners’ proficiency was “comparable with those of university students at the end of the fourth semester of study.” N = 225.

The affiliations, as deposited with the paper:

Author	Affiliation as printed
Xiangying Jiang	Duolingo, Pittsburgh, Pennsylvania, USA
Joseph Rollinson	Duolingo, Pittsburgh, Pennsylvania, USA
Luke Plonsky	Department of English, Northern Arizona University
Erin Gustafson	Duolingo, Pittsburgh, Pennsylvania, USA
Bozena Pajak	Duolingo, Pittsburgh, Pennsylvania, USA

Four of the five authors were employed by the company whose product was being evaluated.

The study also has no control group. The comparison to fourth-semester university students is made against previously published proficiency data, not against a concurrent group learning another way.

None of that makes the result false. It does mean that the two studies underpinning the most repeated claim about Duolingo were both funded or staffed by Duolingo, and neither had a control group.

The pattern continued: a 2024 study in Language Learning & Technology, with the same Duolingo-employed lead author, states in its own text that it “was supported financially by Duolingo.”

A citation error worth fixing

One paper is widely miscited in this area, and if you go looking you will meet it.

Loewen, Crowther, Isbell, Kim, Maloney, Miller and Rawal (2019) is in ReCALL 31(3), 293-311 - not in Foreign Language Annals, where it is often placed. It studied nine participants, learning Turkish, and it is a case study rather than a controlled trial.

Nine people. It is cited far more heavily than nine people can support.

What the independent evidence says

This is the part that most critical coverage of Duolingo leaves out, and it is more favourable to the app than the company’s own studies are rigorous.

González-Fernández and de la Viña (2025), in Language Learning & Technology, ran an independent 16-week study with 337 participants. Duolingo learners outperformed classroom learners on general second-language proficiency and on receptive vocabulary. Classroom learners gained more on listening. Productive vocabulary was comparable.

That is an independent result, with a real comparison group, favouring the app on two of four measures.

Kim, Payant, Skalicky and Namkung (2026), in Studies in Second Language Acquisition, compared three conditions over 16 weeks: Classroom-Only (58), Duolingo-Only (65), and Classroom plus Duolingo (60), for a total of 183 participants.

The finding is symmetrical, and it is important not to spin it: all three groups improved significantly across nearly all measures, with similar magnitudes. The app did not beat the classroom. The classroom did not beat the app. The one place a difference emerged favoured the combination - the Classroom plus Duolingo group gained more on pragmatic competence, specifically the tu/vous distinction.

This is a quasi-experimental design using intact groups, not a randomised trial.

Rachels and Rockinson-Szapkiw (2018), in Computer Assisted Language Learning, ran a 12-week quasi-experimental study with elementary students and found no significant difference in achievement or self-efficacy between Duolingo and teacher-led instruction.

That is a null. It is worth noting how the authors read it: because Duolingo did not differ from teacher-led instruction, they conclude the app “functions as an effective instructional tool.” Absence of a difference is being treated as evidence of effectiveness. That inference is common in this literature and should be recognised for what it is - though in this case the direction of the conclusion happens to match what the better-powered independent studies later found.

The thing that is missing after fourteen years

As of this search, no independent randomised controlled trial of Duolingo exists.

Kim and colleagues used intact groups. Rachels and Rockinson-Szapkiw used a non-equivalent control group design. González-Fernández and de la Viña is independent and controlled but not randomised. The company’s own studies had no control group at all.

For a product with hundreds of millions of registered users, that is a remarkable gap.

What this leaves you with

Duolingo produces learning. Every study finds this, including the independent ones. The contested question has never been whether anything happens.

The contested questions are how much, on which skills, and compared to what. The independent evidence suggests: roughly comparable to beginner classroom instruction overall, better on receptive vocabulary and some general proficiency measures, worse on listening.

Do not use the 34-hours figure. It comes from an uncontrolled, company-funded, never-peer-reviewed report with 88 analysed participants surviving from 556 who started, and its central statistic is a mean more than double its own median.

The combination beat either alone on one measure. In the best-designed independent comparison, classroom plus app produced more pragmatic competence than either by itself.

A streak is not a measure of learning. That is a separate mechanism with its own evidence, covered in the Duolingo streak.

For what spacing does and does not guarantee, see spaced repetition. For the wider question of which study tools survive scrutiny, see study tools and apps actually worth it.

Limitations of this article

Two of the six studies here are quasi-experimental with intact groups, which cannot rule out selection effects.

We verified the Jiang author affiliations through the deposited metadata, not the printed page, because the publisher returned an access error. The competing-interests statement as printed was not read. See the source-resolution record.

We could not verify the exact sample size in Rachels and Rockinson-Szapkiw (2018) from the article, its repository record or the source dissertation record. We have therefore described its design and result without printing an N. See the withheld-figures record.

We did not read the funding declaration in Kim and colleagues (2026), which was not visible in the version we accessed.

No meta-analysis of Duolingo specifically was located. A systematic review of mobile-assisted language learning covering 35 included studies exists and is listed below for context.