AI and subtitling: How machine translation affects the viewing experience
Subtitling has long been a central part of the Swedish television experience, but with the rapid development of AI and machine translation, the subtitling industry has undergone a significant change. But how do viewers perceive these machine-translated subtitles, and what impact do they have on our comprehension of moving content? Hanna Hagström, PhD student in Translation Studies, is conducting a research study that suggests there are clear differences between AI-generated subtitles and those created entirely by humans.
Viewers’ reactions: Do they notice the difference?
An ongoing study on machine-translated and post-edited subtitles shows that viewers can actually perceive differences between these “new” subtitles and the more traditional, human-translated subtitles.
- Most viewer comments so far focus on translation errors and note that the Swedish text often feels unidiomatic and less nuanced than the English dialogue, says Hanna Hagström.
Some examples to illustrate this:
Translation error:
Original: We just got to that structure, and something's not right.
Subtitle: När vi kom till byggnaden var det nåt som inte stämde.
Comment: In this context, “structure” refers to two branches tied together into a cross, not a building.
Unidiomatic language:
Original: What Jeff’s not seeing is this Black Wolf maybe his death sentence.
Subtitle: Jeff ser inte att vargen kan vara hans dödsdom.
Comment: Literal translation. In this context, “see” rather means “understand”.
Lacking in nuance:
Original: Let’s go get this son of a bitch!
Subtitle: Nu tar vi honom!
Comment: The Swedish text is toned down compared to the strong and angry tone of the original.
Previous research by Hanna Hagström and Jan Pedersen shows that machine-translated subtitles processed by post-editors also have almost seven times as many errors as human-translated subtitles.
Genres and programs studied
The current study aims to investigate whether the difference between AI-produced subtitles and human-translated subtitles has an impact on viewers. It builds on a previous study by Hanna Hagström and Jan Pedersen that focused on subtitles in reality shows and documentaries, two popular TV genres where language is often more spontaneous and context-dependent than in drama programs, for example. Among the programs included in the analysis are well-known titles such as Deadliest Catch, Chasing Classic Cars, How the Universe Works and Bizarre Foods with Andrew Zimmern. These programs offer a wide variety of linguistic challenges, from technical terms and jargon to everyday speech and idiomatic expressions, making them particularly interesting to study from a translation perspective.
AI and the trust in subtitles
An interesting aspect of AI-generated subtitles as well as human-generated subtitles is that viewers often express a high level of trust that subtitles accurately reflect what is being said. This becomes particularly evident when dialogue is difficult to hear due to background noise or unclear speech.
- In some cases, machine translations have contained incorrect information that nevertheless fits the context, and viewers have accepted it as correct, says Hanna Hagström.
This phenomenon is currently being further investigated in an ongoing study that analyses how these errors can affect viewers’ comprehension and experience when watching subtitled programs.
The future of subtitling: Will AI take over?
AI and machine translation are increasingly present in the subtitling industry:
- The professional role of subtitlers is already changing, with more people now working on post-editing machine-translated subtitles rather than creating subtitles from scratch, says Hanna Hagström.
But while the technology has the potential to save time and resources, an important question remains: How much loss of quality are we willing to accept? And can we ever trust AI to understand the linguistic nuances and contextual subtleties that human translators master?
- Subtitling is a complex form of translation. Translating moving images requires the ability to analyze multiple elements at once—image, text, sound, dialogue, background music, and more—and understand how they interact to create meaning. It also demands the skill to condense and adapt what is being expressed to fit the often restrictive space available for subtitles. On top of that, analyzing human communication is a challenge in itself, as it’s often multi-layered, context-dependent and subtle. In this regard, humans have a clear advantage over AI. I don’t think AI will be able to replace humans, at least not in the foreseeable future. At this point, it’s more about how humans can use AI to create high-quality subtitles, concludes Hanna Hagström.
Those who want to dive deeper into the topic can refer to a previous study in the field, access the article here or a popular science blog post.
Last updated: April 23, 2025
Source: Områdeskansliet för Humanvetenskap