What AI mistakes can teach human interpreters

Like so many others, I’ve been reading about the emergence of generative AI, trying to understand what this means for our world. What are the implications of relying on generative AI models as a source of information? News coverage of the new AI chatbots released by Microsoft and Google have been quick to highlight their fallibility.  Microsoft’s ChatGPT powered Bing reported several factual errors in financial data, and Google’s first demo of their AI chatbot Bard provided false information about the James Webb telescope.  In one analysis piece about the implications of such visible mistakes, a quote caught my eye: Search powered by AI and large language models “can confidently make factual mistakes and spread misinformation.” 

Confidently making factual mistakes, huh? That seems familiar.

I work as an interpreter. The core focus of my work is in the space between people and their understanding of one another. Language is my business and interpreting runs on a foundation of trust. You could say that my grasp of the languages I use (English and American Sign Language) is one piece of the “language model” I hold in my brain, not unlike the language models that AI chatbots are trained with. In addition to language, my brain is also filled with random facts I’ve learned about the world over the course of my life. I’m using all of these bits of information to form a cohesive re-wording of other people’s inputs into my own brain. My output is, in a way, very much like ChatGPT answers – only as good as the information on which it is trained. 

Now, what if I have an incomplete language model? What if I have incomplete information about the context of the subject at hand? Just like a generative AI model, I use the information that is available to me and fill in the gaps to form my response. I can (and do) confidently make factual mistakes in the course of doing my job, and that has a bearing on the foundation of trust I hope to build with people who use my services.  I stand to cause harm if I am not actively paying attention to my interpreting output.  Luckily, I have something that AI doesn’t have – a human brain.

A critical skill for interpreters to develop is to notice when we’ve made a mistake, recognize the impact, then make a repair. This first requires developing the fundamental skill of monitoring your own work – to listen to your own output in real time. It’s next level stuff (something that AI can’t currently do) but it’s critical if we are going to build trust with the people whose words we hold in our hands. So how do we build this muscle?

Here are a few exercises to build linguistic self-monitoring skills –

  1. Repeat what you hear/see. Listen to talk radio or a podcast in spoken language and repeat exactly what you hear as you’re listening.  You can do the same for sign language – watch a video and copy sign exactly what you see. Can you listen to yourself as you’re also listening to the speaker? It’s a little like rubbing your belly and tapping your head at the same time. Notice your prosody (rhythm, stress, intonation, affect). How’s the fidelity of your message output? 
  2. Develop a practice of reading books aloud. Not only is this a lovely opportunity to engage in literature and bond with others, but it also has a host of benefits for your brain. I find that reading aloud helps me to practice a smooth, natural delivery while focusing my attention on listening to my output, and also keeping an eye ahead to what’s coming up in the next sentence.
  3.  Record and watch/listen to your work. I know, it’s painful. Most people can’t stand the sound of their own voice or to watch themselves on a video. But the hard truth is – if you expect others to listen to or watch you work, you should be willing to subject yourself to it as well. Develop a practice of observing your interpreting work regularly enough that you move beyond self judgement to valuable self-observation and learning.

There is much to be discovered about the impacts Generative AI will have on our lives.  At this early stage, one such impact is already clear – an AI chatbot that is unable to monitor itself can deliver false information with confidence, spread misinformation and have potentially harmful impacts on users. The same could be said of a human interpreter. 

We have a responsibility to those who use our interpreting services to make sure that we are monitoring our output, recognizing the impact and making repairs. To err is human, and it is through the process of making repairs that we can work toward building trust.