Can AI translate your marketing videos without putting viewers off? What a Journal of International Marketing study found
Tools that re-voice a finished video and move the speaker’s lips to fit a new language promise one shoot for every market. Two experiments, run in Indonesia and with British and American viewers, found the AI speaker sounds foreign and slightly artificial, yet people’s stated plans to like it, share it or comment held steady.
Wahid et al. find AI-translated marketing videos sound less native and natural than human-voiced ones, yet viewers were as willing to like, share or comment.
Marketing in 3 · 2025 in research · No. 48 · Paper published 23 November 2025
The words and the lip movements can now be changed after the shoot. Whether the speaker still sounds local is another matter. Photo: Voyage Pro on Unsplash.
The short answer
Mostly, yes, with a catch you can hear. Across 301 viewers, a short video translated by an AI tool into Indonesian or into English drew as much stated interest in liking, sharing or commenting as a version spoken by a native speaker. But people rated the AI speaker as sounding much less like a native and a little less natural, and whether they understood it better or worse depended on which language it went into.
Evidence level
★★★One study
★★★Several studies
★★★Meta-analysis
Key takeaways
Out of 7, the AI-translated speaker scored 1.9 points below a native speaker for sounding native in Indonesian, and 2.7 points below in English.
Plans to engage with the post (liking, sharing, commenting) moved by less than half a point between the two versions (0.08 in Indonesia, 0.35 in the English test), and neither gap was statistically significant.
Indonesian viewers understood the AI version slightly less well (5.93 against 6.31), while British and American viewers understood it slightly better (5.62 against 5.25).
The AI version came across as less natural in both directions, by about 0.4 points each time (3.83 against 4.26, and 3.80 against 4.22).
What the researchers did
Risqo Wahid of the University of Jyväskylä in Finland and Universitas Indo Global Mandiri in Indonesia,
Jiseon Han of the University of East Anglia in the UK,
Nizar Fauzan of Khon Kaen University in Thailand and Universitas Islam Bandung in Indonesia,
and Heikki Karjaluoto, also of Jyväskylä, tackled a problem that global content teams now face. AI tools can take
a finished video, swap the spoken language, keep a voice close to the speaker's own and reshape the lips to fit the new
words. Do viewers judge that version differently from one spoken by a native?
They built two short online experiments around a social media clip for an invented mobile game company. In it, the
founder announces a new game world and asks players to post ideas in the comments for a chance to win a prize. The first
test had a British firm talking to its audience in Indonesia: one group saw an Indonesian man deliver the script in
Indonesian, the other saw an English man's video turned into Indonesian by the AI tool HeyGen. The second experiment
reversed the direction. An Indonesian firm addressed Americans and Britons, with a British man voicing the human
version while an Indonesian man's clip was translated into English by AI.
After one viewing, people rated how well they understood the words, how natural the clip looked and sounded, how far
the speaker sounded like a native, and whether they would engage with it by liking, sharing or commenting. The analysis held constant
age, gender, language skill and attitudes to mobile games and to calls to action that offer a reward.
Participants
301 in all: 140 in Indonesia (average age 23), reached through the authors' social networks, and 161 Americans and Britons (average age 44) on Prolific
Videos
One founder clip of under a minute per version, for a made-up mobile game firm
Translation
AI versions made with HeyGen, which changes the language, the voice and the lip movements
Measured
Understanding, naturalness, how native the speaker sounded, and planned liking, sharing and commenting
Checks
Separate tests with 160 Indonesians and 91 Americans found the speakers equally attractive and convincing
The accent gives the AI away
By far the largest gap was in accent. In both experiments the AI speaker came across as plainly less native, and
the way the video was translated accounted for 31% of the differences in accent ratings in Indonesia and 41% in the
English test. Nothing else in the paper came close.
Part of the reason lies in the tool. HeyGen did not let the team pick which English accent the translated voice would
carry, which is why the second experiment recruited Americans as well as Britons. The two nationalities answered
alike on all four ratings.
Naturalness slipped as well, though only slightly. People saw the AI clip as a little more machine-made, and the
translation method accounted for roughly 4% of the variation in both studies.
How native the speaker sounded (1 to 7 scale)
How native the speaker sounded (1 to 7 scale)
Group
Value ()
Indonesian, native speaker
5.78
Indonesian, AI-translated
3.9
English, native speaker
6.25
English, AI-translated
3.52
Study 1 (140 viewers in Indonesia) and Study 2 (161 Americans and Britons); adjusted means from the published figures.
Understanding depends on the direction
Comprehension is where the two experiments parted ways. Turned into Indonesian, the AI version was a little harder to
follow than the native one. Turned into English, it was a little easier. Both gaps were small, with the method
explaining 3% and 5% of the variation.
The authors suggest a reason they did not test: AI systems learn from far more English than Indonesian, so their
English output may simply be more accurate. The web appendix gives a glimpse of what viewers actually heard. A
transcript of the AI-made Indonesian clip marks two short stretches as unintelligible and a hesitant "umm" halfway
through a sentence. The English transcript shows no such gaps, although the game's prize currency came out as "cash coin".
Engagement did not move
For a marketer, the finding that matters most is the one that stayed flat. In neither experiment did the AI version
make people less inclined to interact with the post. In the second it even scored slightly higher, by a margin that
chance could easily explain.
Look at the level as well as the gap. Every group's average fell under 4, the scale's halfway mark, so this was not a
clip people were eager to pass on, whoever voiced it. The authors take the result to mean that weaker sound and look
need not drag engagement down. A narrower reading, and ours rather than theirs, is that for an ordinary promotional
clip the translation route was not what held engagement back.
Intention to like, share or comment on the video (1 to 7 scale)
Intention to like, share or comment on the video (1 to 7 scale)
Group
Value ()
Indonesian, native speaker
3.55
Indonesian, AI-translated
3.63
English, native speaker
2.78
English, AI-translated
3.13
Study 1 and Study 2, adjusted means from the published figures. Neither gap between the AI and human versions was statistically significant.
Advice for teams working across markets
Their advice is measured. AI translation is not simply worse or better than a human voice: the weak spots are the
sensory cues, accent and naturalness, so that is where to look hard. In some directions, such as into English, the
AI may already hold its own. They describe it as an emerging option for adapting content across markets, provided
quality is kept under review.
This is a short Insight Paper, and it closes with 12 open questions. Among them: whether telling viewers a clip was
AI-translated changes trust, how regional dialects fare, how bilingual viewers react, and whether any of this shows
up in clicks and sales rather than stated intentions.
Marketing in 3 steps
Ready. Set. Market.
1
ReadyDecide and prepare
Start with one proven video. Choose a short video that already does well in its home market. Pick two new markets to try it in and write down the language each one speaks.
Plan extra review for smaller languages. Expect cleaner output when the target is a widely used language such as English. For a language with less content online, set aside twice the review time.
2
SetBuild it
Have a native speaker watch the AI cut. Ask one native speaker per market to watch the translated video twice. They should mark any mumbled word, any wrong name or price, and any line that sounds foreign. Fix or re-render before you post.
Put the key facts on screen. Show the brand name, the offer and any price or prize as on-screen text in the local language. If a word comes out garbled in the voice, viewers can still read it.
3
MarketLaunch and measure
Run it against a local voice. Where budget allows, post the AI version and a version voiced by a local presenter to two matched audiences. After a week, compare likes, shares and comments for every 1,000 views.
Ask viewers how it sounded. Add a two-question poll to the post: was it easy to follow, and did the speaker sound local? Repeat the check every three months, because these tools change fast.
Found this useful?
Marketing in 3 is free and ad-free, with new issues every Monday and Thursday. A coffee keeps it that way.
Fewer reshoots. The appeal is time and money: re-voicing or reshooting even a 30-second social clip for each language soon adds up, as the authors note.
An accent you cannot fully hide. Plan for local audiences to hear that the speaker is not one of them, even when the lips match the words.
Some fixing by hand. Mumbled words and mangled names can slip through, so budget for a review and the odd re-render.
Results that date quickly. The authors stress that these tools are improving fast, so a verdict on today's output may not hold for long.
Where it does not work
One tool, one language pair. Only HeyGen, and only English and Indonesian, were tested.
One clip per version, different men. Each version rested on a single video, and the human and AI versions featured different speakers, so one man's delivery may colour the ratings.
Two different audiences. Young Indonesians reached through personal networks and older panel members in Britain and America also differ in more than the direction of translation.
Intentions only. People only said what they planned to do; nobody's real behaviour was tracked.
A lukewarm video. Engagement was low in every group, so the picture may change for content people are keen to share.
Questions marketers ask
Do consumers notice when a marketing video has been translated by AI?
Mostly through the voice. In two experiments, viewers in Indonesia, the US and the UK judged an AI-translated speaker as far less native-sounding and somewhat less natural than a human speaker (Wahid et al., 2026).
Does AI video translation hurt engagement?
Not in these tests. Intended engagement was statistically the same for both versions in each experiment. These were answers given after a single viewing, not recorded behaviour.
What is AI video translation?
Software that takes a recorded video, translates the speech into another language, recreates it in something like the original speaker's voice and adjusts the lip movements to match. The study used HeyGen; the authors also name Adobe Firefly and Invideo AI as tools that do this.
Should brands tell viewers that a video was translated by AI?
Disclosure was not part of these experiments. The authors leave it for future work, alongside dialects, live translation and the effect on real purchases.
Podcast hosts have only a voice and a story to win people over. Across two studies, listeners who saw a host as authentic were more loyal to the channel, keener to buy the brands the host endorsed and readier to defend the host, with emotional attachment and trust carrying the effect. The pull on trust and buying grew when the host confirmed how listeners saw themselves. Sincerity travels by sound alone, which makes host fit worth checking before booking a podcast read.
Three questionnaire studies compared live-stream product demos shot from the presenter's own point of view with demos shot from a third-person angle. The third-person view did more to keep people watching and make them want to buy, because the extra angles let viewers inspect the product, almost feel it and get absorbed in the stream. Product type shifted the result: in first-person streams functional products fared better than hedonic ones, and in third-person streams the reverse. Camera position, it seems, is a sales choice as much as a creative one.
Source
Wahid, R., Han, J., Fauzan, N., & Karjaluoto, H. (2026). Generative AI for Video Translation: Consumer Evaluation in International Markets. Journal of International Marketing, 34(1), 27-33. https://doi.org/10.1177/1069031X251404843
This brief is written for practitioners from the published article. Open access (CC BY 4.0): free to read at the publisher. Numbers are as reported by the authors; read the article for full methods.
Schivinski's machine learning study of 1,356 brand followers found five engagement patterns driven by commitment, and they still matter in the age of AI.
Bruno Schivinski is Associate Professor of Marketing at Gdańsk University of Technology. He studies how consumers engage with brands in digital and social media.