The consent problem
Generative AI is unethical. There’s no way around that.
I don’t expect this to radically change anyone’s mind about using it. We live inside unethical systems our whole lives, and we’ve all got blind spots – life is complicated. But we should at least be aware of the impact.
Building generative AI models starts with ingesting absolutely any data that can be found: the whole web, all the books (including pirated copies), television, podcasts, you name it… It all gets scraped, regardless of whether the authors or creators have consented or not.
The AI companies say this is just like a search engine spidering the web, but it’s not. Search engines point people back to the original content, giving the authors traffic and readers. Large language models summarise or, worse, plagiarise the content, rarely crediting the original source or pointing to it. This can only result in many publications eventually shutting down as the AI ouroboros slowly kills the web.
Then you’ve got all the exploitative, neo-colonial, sometimes trauma-inducing labour. Behind the clean, futuristic-looking magic of AI, there are scores of humans in a network of low-wage digital sweatshops, mostly in the Global South, sorting through and labelling all the data to ensure models seem smart and safe.
The environmental aspect is probably the worst. The AI companies are far from transparent about their energy and water use (never a good sign). But conservative estimates put the consumption of AI at five times the energy of standard computing. Training a new model takes data centres running for months on overdrive, literally burning through chips as they work.
I don’t want this to be too long, so I’ll end here. But I haven’t even mentioned the psychological harm, the use in military kill-chains (despite their posturing, Anthropic are still very much involved too), collaboration with the current US administration, the so-called open source models, etc.
There’s a lot more to say here, but that’s for another time.
Colin