AI is powerful. Ask ten people on the street and seven would likely agree, if not all. But not many know how much goes into building and running it: from the ground up, the minerals (chips require 20+ different minerals, much of it mined in politically unstable regions with documented labor abuses), to manufacturing the hardware itself (cases, chips, wires), to the electricity and grid it leans on, to the water footprint running through every one of these processes, to the collection and processing of data, to the compute burned in training and then again every single time it runs, every response it generates. And underneath all of it, human labor at every stage and the damage that comes with it. Kenyan workers paid as little as $1.32 to $2 an hour to label OpenAI's training data were fed a steady diet of graphic text on child abuse, bestiality, and torture, and came away with lasting trauma. AI is powerful, costly, and consequential, all at once.
Looking at the common AI applications, I want to ask why are we pointing such precious technology at mundane things that might not really be that much value add and also displace human labor and disrupt economies? Setting aside my other question for society ( how much faster is fast enough, how much more productive is productive enough), we were told AI was going to push us forward - cure cancer, model proteins - actually move the needle on the human well-being (to which I also wonder how old is old enough, how healthy is healthy enough). But look at what the usage data actually shows. OpenAI's own research puts non-work messages at over 70% of ChatGPT traffic. Sorted another way, about half of all messages (49%) are “asking” - seeking advice or information, while 40% are “doing," meaning task-oriented work like drafting, planning, or programming, and the remaining 11% is “expressing": personal reflection, exploration, and play. I observed this too at my university reading group where the common usage among eleven people of different academic backgrounds was using chatbots to navigate difficult interpersonal situations. Useful, clearly. But closer to daily errands than moonshots. And by traffic, some of the most-used consumer AI apps are beauty filters that reshape a face in a selfie, companion apps that simulate a relationship, and background-removal tools for a product photo. Not exactly cancer.
Let’s also not forget that the world is not equal. Both the access and benefits from technology, and the suffering from resource extraction are not distributed equally. Most AI applications users are sitting in homes with air conditioning and clean water. The cost of building and running the models that serve them accelerates climate change. But they don't have to worry about it, because that cost gets spread out to somewhere they can't see. One thing that’s clear is that the surest beneficiaries of these AI labs and companies are the shareholders invested in these companies. Beyond that, I haven't researched the user groups enough to say more here. But let's put the technology to use where it counts.
Where it counts… what do I mean by that? Here's one way I think about it. Machine learning is good at recognizing patterns in historical data (that's the basis of both classical statistical models and today's transformers), including the gaps and biases baked into it. Left uncontrolled, it detects and amplifies them. Used with the right intention, pattern-recognition capability can tell us what's wrong with our society today, and the answer to these identified problems usually isn't AI at all.
Let’s read the usage patterns as point to unglamorous but important work. The AI beauty filter is a signal that tells us we have a beauty-standards problem. AI therapy chatbots are a signal about our barrier to real mental healthcare - cost, access, stigma, and more. AI companion apps is signal of our loneliness epidemic, and a shortage of real community and trust between people. Reaching for a machine when we want to turn someone down, deliver news the other person won't want, say the hard thing without blowing up the relationship is signal that our education might have been focusing on producing containers of knowledge, rather than decent, emotionally-regulated, understanding humans with communication skills and confidence. Let's stop building gold plumbing for a broken toilet when the real issue is something's wrong with the whole infrastructure.
But that's just me. What about others? This pulls up a question I have about AI alignment in AI Safety, and frankly, a well-debated classic question. Values are not universal. I'm pessimistically optimistic about the alignment problem, because if we can't get humans to align their values with each other, I'm not sure what confidence we're supposed to have about aligning a machine's values with ours, when "ours" is so fuzzy to begin with. You can't align anything to fog. I'm an optimist at heart, but climate change is the case study sitting right in front of us - we've made real improvements, and I still don't see incentives shifting fast enough to matter. Canadian climate activist David Suzuki said something close to this recently: the fight against climate change is lost.
Speaking of value alignment, Isaiah Berlin put it this way: "The world that we encounter in ordinary experience is one in which we are faced with choices between ends equally ultimate, and claims equally absolute, the realization of some of which must inevitably involve the sacrifice of others." There's no common currency to weigh values against each other. More justice means less mercy, more equality less liberty, more efficiency less spontaneity, and no clear rule that tells you how to trade one off against another. How are we supposed to encode that into AI models when humans have struggled with it for our whole history? Why are we adding this challenge on top of everything else? Which is exactly what Berlin pointed out, the balancing act. More innovation means more complexity means more problems.
This reminds me of the law of conservation of mass. Take two quick non-digital technology examples. First one is heat transfer: to generate x amount of energy, we will always need x + y input, as there's always waste in doing the work and there's a byproduct (heat, usually). Then another example is waste management. Incineration is a standard solid-waste treatment method, and burning organic solid waste converts carbon into gaseous CO2 (and other combustion gases like NOx). As for wastewater treatment, the process generates a solid byproduct called sludge. When cleaning air, particulate matter that would’ve stayed airborne ends up as solid ash or dust that has to be collected and landfilled, or it gets absorbed or chemically reacted to a liquid spray, producing a wet byproduct slurry and then get treated back to solid.
The physical law of conservation of mass is unbiased. The treatment doesn't destroy waste, it moves it from one environmental medium (solid, liquid, gas) to another, and usually changes its form in the process. You can concentrate, dilute, or convert waste, but the underlying material has to end up somewhere. And I feel it’s the same with technology. You create convenience but somewhere there will ultimately be new problems caused by convenience. Any promises that claims otherwise are worth examining closely. If you are experiencing convenience upstream, there are problems downstream, or further upstream than you can see. I feel confident about this idea. Hydropower is another example for another day.
This is where I feel we should leave it to the community to decide, what's worth the risk, worth the trade-off, and which values to align with. Perhaps one way is to leave the big tech behind and do what Māori communities have already done, build from the open source models and model the community’s value. Are open source models really open? Do people trust the capability of open source models? Or are we going into the same dilemma as search engines where the cost of starting one is insanely high (indexing is expensive for search) and it’s really hard to have a meaningful alternative? I have heard enough times of academics saying to do cutting edge research, you have to go into the industry. Which is quite shocking. Research used to sit at the top of academia. Now it sits inside labs run by profit-driven companies.
Even though I think the value should be left to the community to decide, I still think there should be some coherence around the fact that we have one Earth and its ecosystems are interconnected. Some would say the answer to our scarcity problem is to explore a different planet, and they want to build ships to get there but that’s at the expense of the other people living on the earth. The resources and emissions of space programs are extracted from everyone's shared Earth while the escape option belongs to a few. They are actually violating other people’s consent.
I will write more about my thoughts on AI and technology. A lot of my thoughts above require further examination, research and reflection. I am glad I have finally started writing to think out loud. This is just a start.