Opening remarks on “AI in the Workplace: New Crisis or Longstanding Challenge”
On Thursday 9/28, I had the opportunity to speak at a virtual roundtable convened by Congressman Bobby Scott on “AI in the Workplace: New Crisis or Longstanding Challenge?”. The roundtable was a closed meeting, but sharing our opening remarks is allowed, so I am posting mine here.
Below is a lightly edited transcript:
Thank you Representative Scott for organizing this and thank you all for your time and attention. As mentioned I’m a professor of linguistics at the University of Washington and I work in computational linguistics. Part of what I would like to do here today is give you some insight into how this technology works and how it is already in the world around us so far.
What is AI?
In fact this is a marketing term. It’s a way to make certain kinds of automation sound sophisticated, powerful, or magical and as such it’s a way to dodge accountability by making the machines sound like autonomous thinking entities rather than tools that are created and used by people and companies. It’s also the name of a subfield of computer science concerned with making machines that “think like humans” but even there it was started as a marketing term in the 1950s to attract research funding to that field.
I think that discussions of this technology become much clearer when we replace the term AI with the word “automation”. Then we can ask:
- What is being automated?
- Who’s automating it and why?
- Who benefits from that automation?
- How well does the automation work in its use case that we’re considering?
- Who’s being harmed?
- Who has accountability for the functioning of the automated system?
- What existing regulations already apply to the activities where the automation is being used?
In order to make this a bit more concrete, I want to break down the different kinds of systems that are being called AI these days. There are different types of automation:
1. One thing is using computers to automate consequential decisions. This is called automatic decision systems and it’s used for example in the process of setting bail or approving loans or screening resumes or allocating social benefits.
2. Another kind of automation is when we’re automating different kinds of classification: things like image classification to try to get the camera to focus on the faces or classifying web users for targeted advertising.
3. A third type of automation is automation of the choice of information to present to someone. This is called recommender systems and it’s for example the automation behind the ordering of the feed in social media or movie suggestions in Netflix.
4. A fourth type is automating access to human labor or making human labor conveniently available to buyers. Here think Uber, Lyft, Amazon Mechanical Turk and similar services.
5. The fifth type I want to call out is the automation of translation of information from one format to another: automatic transcription, finding words and characters and images like automatically reading license plates, machine translation or something like image style transfer: make this photo of me look professional.
6. Then finally there’s a type that’s been very much in everyone’s mind recently: things like ChatGPT which I call synthetic media machines. These are systems where you might be able to generate images based on specific content or specific styles or a plausible sounding text without any commitment to what it says.
I want to just put in a few more words about ChatGPT. It’s important to understand that its only job is autocomplete, but it can keep autocompleting very coherently to make very long texts. It’s being marketed as an information access system but it is not effective for that purpose. You might as well be asking questions of a Magic 8 Ball for all the connection to reality or understanding that it has.
A very key thing to keep in mind here is that the output of these systems don’t actually make sense. It’s that we are making sense of the output. It’s very hard to evaluate them because we have to take that distance from our own cognition to do so.
Finally the ability to create plausible sounding texts on just about any topic is quite dangerous because it looks like we have or are just about to have robo-lawyers, robo-doctors, robo-tutors, robo-therapists, etc., and we don’t.
Popping back up to that full range of automation, I want to point out that these systems have some characteristics in common:
They are built using training data and then algorithms that capture the patterns in the training data and can reproduce them over new data at runtime to varying degrees of accuracy and varying degrees of desirability. For example, automatic transcription takes patterns of how sounds map to written words and captures them in sufficient detail meaning that we can get first packs automatic captioning in Zoom. That’s very useful, though we can also see problematic biases: Things like if you’ve got a less frequent name it’s more likely to be transcribed poorly. Another example is the infamous COMPAS algorithm for predicting recidivism risk that reproduces patterns of racial discrimination in policing in terrible ways. Another example is image systems trained on large collections from the web tend to reproduce patterns of sexualization in images of women especially women of color. And similarly ChatGPT and systems like it will output hate speech and more subtly biased language, again reproducing these patterns.
There are other things these have in common: They work well because of the effort of people but they’re often designed to hide that effort. They may promote mass surveillance either by enabling it or providing motivation for it. And they tap into automation bias, that is our cultural tendency to assume that computers must be objective authoritative and fair.
Finally the hype around these systems really serves corporate interests because it makes the tech look powerful and valuable because it distracts from the real issues that I hope regulators will be focusing on and because it makes the AI seem too exotic to be regulatable. My hope is that our representatives are critical consumers of information about this technology and not falling for the narrative that this is moving too fast for regulation to keep up. Your job is to protect rights and those aren’t changing so fast.
