You might be surprised to learn it, based on the bulk of my writing, but I am pro-AI. The potential for machine learning to give us life saving, world-changing systems is huge. The reason I write so much stuff about the issues with AI is that I am realistically pro-AI. I love AI doing the things it can do really well in an ethical and sustainable way. What I object to with visceral intensity is AI being hyped as being able to do things it cannot do, in ways that are often incredibly dangerous, both to individuals and society. But I’ve written quite a lot about that, and will no doubt write quite a lot more, so today I want to focus on the things that AI can and should be used to do.
A lot of those things – maybe all of them – require what we call “human in the loop” – ie you get an AI system to do the tedious, tricky bits, and then get a human to check that work, and apply any judgement, reason, or creativity required. All hype aside, AI is not capable of judgement, reason, or creativity. Not even a little bit. That means that you can’t just send the results of AI systems out into the world without human involvement. And there must always, ALWAYS, be a human being in charge who can countermand the system’s results where needed.
So what kinds of things is AI genuinely good for? There are so many, but I’ve chosen a few stories that really hit home for me.
I came across this really beautiful example on threads recently, in this story posted by “writegirl”:
“I’m deaf and rely on lipreading to communicate. I miss a lot of things, usually because I don’t understand someone or have my back to them. I just bought a pair of smart glasses that have ONE function: creating captions of the spoken word that are displayed for the user.
I wore them in public for the first time today! Imagine my joy when I was browsing in the farm supply store and saw the words “Hi how are you? Can I help you find anything ma’am?” show up on my glasses, and could turn around and reply! (And the surprise on the clerk’s face because I am a regular there and he probably says that to my back alllllll the time and I “ignore” him!) AND when I went to pick up my prescription, instead of maneuvering in the line so I can get the clerk without a face mask, I just stepped up to the one with the mask on.
Game changer! I can’t tell you how excited I am about this. It’s amazing how the smallest interaction can make such a difference in one’s outlook.”
Imagine having glasses that change your life by removing some of the obstacles created by a physical difference like deafness. This is a use of speech-to-text functionality that I can absolutely get behind! A truly wonderful application of large language models. (Not all that large, by the way. You don’t need systems trained on stolen data, using billions of parameters, and chewing through power and water supplies like some kind of electronic tiddalik to create a system that does this effectively. Speech-to-text is solvable with much smaller systems, and has been available for years.)
Sure, it won’t be 100% accurate, but in this case there are built in humans in the loop – both the wearer of the glasses, and the person talking to them, who can communicate and figure out any glitches. It’s certainly a huge improvement over not even knowing there’s someone behind you trying to talk to you!
For another positive use of Large Language Model technology, look no further than New Zealand, where Te Hiku Media is using LLMs to build a speech recognition model for te reo Maori that has achieved 92% accuracy by being trained on 30 years worth of archival footage and requested audio clips from community members – a consent-based opt-in model for the collection of training data that proves that an ethical approach to building LLMs is both possible and effective. Their system enables transcription and translation of te reo Maori, which will help preserve the language. The authors of that system have important things to say about Indigenous Data Sovereignty, particularly as it relates to language.
Of course, Artificial Intelligence is more than just Large Language Models, despite the frenzy of hype that surrounds them. One of my favourite applications of AI technology – more accurately known as Machine Learning – is this project at Pawsey Supercomputing Research Centre, that uses machine learning to predict crises in patients with traumatic brain injuries (TBIs). Much of the damage from TBIs is caused some time after the initial injury, when the brain starts to swell, and the resulting intracranial pressure causes severe and irreversible brain damage. ICU doctors at Royal Perth, the Alfred, and Royal Melbourne Hospitals partnered with Data Scientists and Pawsey Centre to develop a system to predict the onset of intracranial hypertension, enabling medical staff to intervene before brain damage occurs, and have a good chance of preventing it.
These are just three examples of ethical, sustainable, and impactful AI research and applications. There are many, many more. AI technology is diverse and potentially high impact The AI Hype industry wants us to believe that megascale LLMs developed and operated unethically and unsustainably are the only possible path to the future, but these examples clearly show that argument to be nonsense. Smaller systems designed to do one thing really, really well can be game changing. We can choose what direction we take AI, what impact it has, and which games we want it to change.