What do Deaf People Really Think of AI-Generated Captions?

by Bit Blair

We’ve talked about automated captions and their impact on transcribers’ morale and job prospects

But what really matters the most to us is whether Deaf and hard-of-hearing people are getting their needs met. How do Deaf and hard-of-hearing people feel about AI-generated captions? Do they think that automated captions are a suitable replacement for human service providers?

To find out, the Association of Transcribers and Speech-to-Text Providers, or ATSP, sent out a survey to people who use captioning services. Here is what they learned.

Who responded to this survey?

Most survey respondents were experiencing some degree of hearing loss and were either college-aged or over 50. Respondents typically used captions in college-level classes or for work.

Colorful pie chart showing the proportion of survey respondents who identified as Deaf, deaf, hard-of-hearing, hearing impaired, and other.

80% – 90% of people who responded to ATSP’s survey identified as D/deaf, hard of hearing, or hearing impaired. The remaining respondents marked “other” or “prefer not to say.”

Who were the “others?” We don’t have data on this, but survey respondents most commonly reported ADHD as a co-occuring disability.

Colorful pie chart showing the proportion of survey respondents who were 13-17, 18-29, 30-49, 50,60+, and other.

Most respondents were young, with 42% saying that they were between the ages of 18 and 29. The second largest cohort was comprised of people 50-60+, who represented 31% of respondents.

Colorful pie chart showing the proportion of survey respondents who use captioning services in K-12, Higher Ed, Work, and Other.

59% of respondents used captioning services in higher education. Another 23% used captioning services at work.

What kind of services did respondents prefer?

Most respondents have used closed captioning, TypeWell/CART, and AI-generated captions. They reported that live, on-site, human-generated transcription services were the best option in classrooms or meetings where there is a lot of interaction between participants.

Colorful bar chart showing the number of survey respondents who have used closed captions, CART, AI captions, TypeWell/C-Print, ASL interpretation, and Assistive Listening Devices.

Almost all survey respondents (80%) have used closed captions.

Only 40% of respondents have used an ASL interpreter as an accommodation. This challenges a common misconception that all people who experience hearing loss know and use sign language. 

This population seems to be experienced with AI-generated captions. 59% of respondents have used this accommodation.

Onsite vs. Remote

Colorful bar chart showing the number of survey respondents who prefer on-site services in lectures that require participation, no preference, all the time, STEM classes, Lectures with one speaker, and never.

Respondents had a slight preference for in-person services. They reported that on-site captioning is more appropriate to capture content in classes with lots of participation and interaction (39%) and dense, technical material (21%). Some preferred in-person services all the time (25%). However, almost one-third of them had no preference (29%). Only 12% of respondents want remote captioning all the time.

AI vs. Human-generated

Respondents did not feel that AI-generated captions were satisfactory or equitable.

This is not because of a lack of exposure to or awareness of AI-generated captions. Almost 60% of respondents have used these services, whereas only 34% found them satisfactory.

Colorful pie chart showing the proportion of survey respondents who feel that human captions are more equitable than AI-generated captions: Strongly Agree, Agree, Neutral, Disagree.

87% of respondents “agree” or “strongly agree” that human-generated captions are more equitable than AI-generated captions.

Colorful pie chart showing the proportion of survey respondents who are satisfied with the quality of AI-generated captions: Strongly Agree, Agree, Neutral, Disagree, Strongly Disagree.

Only 34% agree or strongly agree that AI-generated captions are satisfactory. 47% disagree or strongly disagree.

When asked to name specific elements of AI-generated captioning that they didn’t like, respondents said:

  1. No punctuation (64%)
  2. No environmental context (e.g., background noises) (74%)
  3. No non-verbal communication (54%)
  4. Tone and sarcasm conveyed only “some of the time” (64%)

A consistent number of Deaf and hard-of-hearing respondents, around 20-30%, felt neutral about the abilities of AI captioning software. Maybe they feel as though AI-generated captions are better than nothing when human-generated captions aren’t available. Maybe they don’t have enough experience with AI-generated captions in a classroom setting to have a clear preference. Or maybe they feel like the technology isn’t “there” yet, but they are open to it and believe that it may improve. It will be interesting to see if these data points change as the technology evolves and becomes more prevalent.

Conclusion

As providers of live captioning services, our primary mission is to provide accurate, context-aware, equitable access to Deaf and hard-of-hearing people. No matter how much we love our jobs and believe in what we do, the needs of our clients must always come first.

If Deaf and hard-of-hearing people felt that AI-generated captions were superior to human-generated ones – cheaper, more convenient, more accessible, or easier to understand – we would celebrate AI-generated captions as a pivotal advance in disability access.

But this is not the case. 

On-site, human-generated captions are still a clear favorite among Deaf and hard-of-hearing people. The people who need captioning services are not relying on sign language interpretation or assistive listening devices as their primary means of access in classrooms and meetings. Many survey respondents may have become deaf or hard-of-hearing later in life. They need captions to help them keep up with heavy, fast-paced lectures or meetings, and the group discussions that take place in those settings. They need communication access that provides them with accurate spelling and punctuation, environmental context, and non-verbal cues.

The push for AI-generated captions is likely being driven by a desire to cut costs, not by an earnest intention to improve disability access.

As of today, AI-generated captions are not an adequate substitute for high-quality, trained human captionists.

About the Author

Bit Blair has been transcribing since 2013, in-person and remotely. She has a degree in Library Science and lives in North Carolina with her son, where she is busy turning her suburban yard into a paradise of permaculture.