Monday, March 20, 2023

Word prediction on steriods and the authorship questions it raises

I've been thinking lately about the future of word prediction (and phrase and sentence prediction). We’re at a point now where, without the user typing a single letter, but instead just selecting predicted words, syntactically and semantically coherent messages can emerge. This is obviously a huge boon for anyone who needs help typing and actually knows the meanings of the predicted words and what they want to say with them.

But what about those who don’t? What about all those individuals with autism who use AAC not because they have problems with motor control, but because they have problems with language? How do we know that someone isn’t simply selecting words at random that they don’t understand? Worse, given how text-prediction software can adapt to the styles and content of particular users, how do we know that the AI hasn’t been trained through earlier sessions that were mediated through on or another form of facilitated communication to output messages that originated with a facilitator—even when that person is no longer in the room?

The answer, as with the more standard, immediate cases of FC, is message-passing testing.

But while the essence of a message-passing test for text-prediction is still the same, there are some interesting differences—and text-prediction tests might actually be easier to conduct.

That is, if everyone agrees that the autistic person's assistance is entirely through text-prediction, there’s no excuse for any potential human facilitators to be in the room or within cueing range. Having removed these humans, a tester might safely ask any questions whose answer are known to the testers, or even questions whose answers aren’t known to the testers. Not just questions about transient circumstances like “How many fingers am I holding up?” and “What color is this shape?”, but also questions that tap into general background knowledge (“What day is it today?” “Who is the president of the United States?”) or, even, open-ended questions like “What would you like the world to know about autism?” 

Even if some of these questions have been anticipated and rehearsed with a human ahead of time, the typist would still have to have acquired enough language to comprehend each specific question when it was asked. Indeed, this would be an interesting way to probe how much the person can comprehend--independently of how much they can express.

No comments: