AI offers many opportunities for information access among the other benefits. However, if the basics of web accessibility are not followed, the promise of that access will be difficult or more to some parts of the population.
Both Bing and Bard, from Google and Microsoft respectively, currently need to improve at one of the most basic tests here in my trials. My instruction to both AI services:
Show me a list of U.S. states in a table based on population.
In both cases I received a table with proper column headings but row headings were not present. I tried a range of commands to get them to appear with no success. I tried more with Bard leading to Bard eventually acknowledging that it didn’t know how to add row headers yet. I suspect trying similar additional instructions would yield some equivalent result with Bing.
I added an instruction to both services as my first attempt to improve the output where I added the instruction to ensure the table had proper row and column headers for accessibility. This had no impact on the result.
It is vital that the information from AI technology be accurate. It is equally as critical that proper accessibility be used for that output.
Asking both services for details on how to create an accessible table yields good results talking about both row and column headers among other points that would be common from an accessibility perspective. So both services should be following their own advice here.
As these AI experiences become, if they are not already, more mainstream in society, developers need to ensure proper standards are used for information display. My intent is not to single out Bard and Bing exclusively. These are two services I have immediate access to for experimentation but I suspect other AI experiences would yield equivalent results. If you know of a service that passes this test today, please share it in the comments.
In order to get the response one wants from these LLM models, as pointed out in many articles, it is important to provide an explicit prompt and clear directions.
I input the following text into BARD and got proper row and column titles:
“Produce a table where each row is a state and having a column for the population of each state in 2010 and another column for the population of that state in 2020.”
This navigated well with JAWS and I could easily ascertain which row and which column I was in.
I just tried that exact text. I get a table yes but the first column is not marked as row headers. Unless I change the setting in JAWS to not use just marked headers, the row header is not read as I move down a different column.
The raw HTML also does not show row headers marked.