Arguments about AI summarisation
This is from an W3C discussion thread, where AI summarisation and minuting of meetings was proposed, and it lays out some interesting issues with LLM summarisation:
Sure I’m excited about new tech as the next person, but I want to express my concerns (sorry to point out some elephants in the room):
-
Ethics – major large language models rely on stolen training data, and they use low wage workers to ‘train’ at the expense of the well being of those workers.
-
Environment – Apart from raw material usage that comes with increase in processing power, LLMs uses a lot more energy and water than human scribes and summarisers do (both during training and at point of use). Magnitudes more, not negligible, such that major tech cos are building/buying nuclear power plants and areas near data centres suffer from water shortages and price hikes. Can we improve disability rights while disregarding environmental effects?
-
Quality – we’ve got a lot of experts in our group: who are sometimes wrong, sure, but it seems like a disservice to their input, knowledge and expertise to pipe their speech through LLMs. From the couple of groups I’ve been in that used AI summaries, I’ve seen them:
- a. miss the point a lot of the time; it looks reasonable but doesn’t match up with what people said/meant;
- b. ‘normalise’ what was said to what most people would say, so it biases towards what’s more common in training data, rather than towards the smart things individuals in this group often bring up. Normalising seems orthogonal to innovation?
- c. create summaries that are either very long and wooly, with many unnecessary words, or short but incorrect.
If we’re considering if it’s technically possible, I’d urge us to consider the problems with these systems too, including in ethics, environmental impact and quality.
The “normalising” risk is one that hadn’t occurred to me, but it makes perfect sense given how LLMs operate.
Tags: llms ai summarisation w3c discussion meetings automation transcription
-