A Use Case for AI – Accessible Documentation
Studio Notes | What We’ve Been Up To
by Sean Casey, Service Designer.
Before we start, trust us – we never expected to be jumping on the “AI” bandwagon. We’ve been experimenting with some tools on a project and thought it might be good to share…
We’ve been working with Dublin City Council designing a new way to involve Disabled Persons Organisations in decision making processes. The work has meant significant time spent producing accessible documentation. One task is the production of “Easy Read” documents based on guidelines from organisations for people with intellectual disabilities or reading difficulties.
We know that this adds an extra task to projects, and so we have been doing some tinkering with (you’ve guessed it!) Chat GPT to see if it might help.
Prototyping
First, we decided to just ask the AI to convert an excerpt from a report to and Easy Read version suitable for people with intellectual disabilities. The result wasn’t great, with the output looking more line a general summary than an Easy Read.
Next, we fed some Easy Read guidelines from our research in and asked it to try again. This produced better results, with the overall structure and language being appropriate but there were still misunderstood terms and some oversimplification that had to be corrected.
With some refinement we tested to see if ChatGPT could handle larger chunks of content, but the output was presented in bullet points, and the language wasn’t simple enough. We asked ChatGPT to modify the output using simpler language and fewer bullet points, which produced a good piece of Easy Read documentation.
We tried other different methods, and the most promising approach we found was to summarise the input first, then do a pass to make sure no important content was excluded, and then ask for an Easy Read version. However, the quality of the output was inconsistent.
Result
For a final additional test, we tried one last method. We had the idea to try asking the AI to convert a piece of the report to Plain English – and then to Easy Read. These two forms of writing are very similar, but with Plain English being a more familiar term we figured it might be more likely to output good results.
This was likely the best result yet – choice of language was very good, and the overall structure was preserved, and the information was accurate. Some slight adjustments were needed here and there but these were minimal. This test was replicated on other excerpts from our report and the outputs were equally reliable.
Want to see the process applied to this article? Find it here!
In upholding the human rights of disabled persons within the UNCRPD, it is necessary for public bodies to communicate in an accessible manner. This benefits everybody, and if just some of the work can be automated like this, then there is potential for huge positive impact on our society.
Whats Next?
There is likely an even better method that we are yet to figure out, but for now this has potential to reduce our workload for producing these accessible documents.
Going forward, we’d love to push even further with AI tools to tackle more complex accessibility challenges – a huge win would be to produce draft text descriptions of street plans and other technical documentation. These draft descriptions could then be reviewed by architects or project managers, saving time, an increasing accessibility and engagement.
Do you work with these kinds of tools? Have ideas that might be able to help? Reach out and let’s talk!