In conversations about technology, being called a Luddite is a big insult. Colloquially, it refers to a kind of technophobic kook who’s morally against technological development. But in my opinion, the original Luddites were actually very cool. In 1811, around twenty thousand textile workers were fired from their factory jobs in Nottingham after factory owners automated their jobs. Where weaving had once been an artisanal skill, it was first reduced to atomized work in the mills, and then with the integration of shearing frames in 1811, weaving became completely unnecessary labor. Because of the adoption of shearing frames into textile mills, the portion of factory revenue that had kept 20,000 factory workers and their families alive in the form of wages was freed to be reinvested into more automated technology, creating more profit for the factory owners. A mass of now unemployed textile workers–who came to be known as the Luddites–stormed their former workplaces and destroyed the machines that destroyed their livelihoods. The Luddites saw clearly that when faced with the choice between employing human labor and using shearing frames, factory owners would always choose the less costly machines, leaving skilled workers redundant and forcing them into lesser and lesser skilled work as technology developed–meaning ever-decreasing standards of living.

Just like how the Luddites saw that the productivity brought in by the shearing machine came with hidden costs–their livelihoods–it is important to think critically about the hidden costs of the productivity of AI. I can hardly walk five feet, or scroll for five minutes, without coming across the sentiment that AI is inevitable, and we have to just roll over and integrate it into our work lest it leave us behind. I’d like to suggest though, that you do not roll over, and instead engage in Luddite-thinking about AI. What are the costs of AI integration? Are those who benefit the same as those who bear the consequences? And–if push comes to shove–do we really have to smash these machines?

Work Displacement, not Replacement

Now that we’re in the beginning of the AI boom, 41% of companies across the world have plans to replace part of their workforces with AI Automation by 2030. While AI might be thought to automate ‘menial tasks,’ the ‘work’ done by AI has to be monitored and double-checked with human labor. Remember those ‘workerless’ Amazon Just Walk Out stores that use AI to track what shoppers leave with? Amazon actually used outsourced, hyper-exploited workers in India to review 70% of AI calculated transactions. The use of AI here was not to create actually cashierless stores, but to pay poverty wages for hidden cashiers in India instead of paying slightly higher poverty wages for cashiers in the Global North, saving Amazon about a hundred thousand dollars per store each year. 

History repeats itself–first as tragedy, second as farce. We’ve seen all of this before. In recent memory, industrial manufacturing, which had been the good American job, was revolutionized by increased technology and automated, then the remaining necessary labor was outsourced to cut costs. The leftover workers in the Global North were split, some pushed into clerical or customer service work, many pushed into unemployment. Next, the internet rendered many of those jobs obsolete, and the remaining necessary labor relocated to the Global South–call centers, data entry, etc. And the leftover workers in the North were split, some into ‘fake email jobs’ or knowledge production, and many, many others into gig economy work. And now, it’s the same. Technological development of generative AI and its integration into labor is poised to semi-automate managerial and knowledge production work. USAID is already in the mix of this new economic imperialism–in January it put out a call to fund research to identify a data enrichment labor market in the Global South. Labor exploitation will increase in the Global South while American profits will soar. It’s already happening in Kenya, where workers are paid $2 an hour to train LLMS for OpenAI and Meta in ‘AI Sweatshops.” 

High Environmental Costs

It’s no secret that the AI industry is putting a strain on energy consumption. Asking ChatGPT to create an image uses as much energy as it takes to charge a smartphone. That doesn’t seem like much, and that’s true. But just like our less than stellar recycling habits or drive to work are a drop in the bucket compared to airline industry emissions, our personal use of AI models requires very little energy in comparison to the energy that it takes to train and maintain these models. For example, on average, each query sent to ChatGPT emits 4.32 grams of CO2. The average user sends 8 queries a day. Sending 8 queries a day for a year emits the same amount of carbon as driving a car for 3.5 miles. No biggie! But, there are over 100 million people who use ChatGPT every day, and 1 billion queries are placed everyday. This amounts to an extra 4762 tons of daily emissions. That’s like sending an airplane from New York to Tokyo 2381 times in a single day. 

Even when we set aside the inputs from chip manufacturing and supply chains, training an LLM consumes thousands of megawatt hours of electricity and emits hundreds of tons of carbon. The training process to build Chat GPT-3 alone consumed 1,287 megawatt hours of electricity (that could power 120 U.S. homes for a year) and generated about 552 tons of carbon dioxide. Moreover, water must be run through AI data centers to keep the machinery from overheating. Each kilowatt hour of energy consumed by a data center requires 2 liters of water for cooling. Training Chat GPT-3 used 6 cubic miles of water. The total energy cost of the AI industry is expected to exceed the energy usage of the entire country of Belgium by next year. In just the US, AI data centers are expected to raise the nation’s electricity usage by 6% by next year. 

It has become quite clear that like labor markets, the climate crisis is geographically uneven–constituting a kind of environmental imperialism. While the United States alone is responsible for 25% of historical carbon emissions, the consequences of these emissions occur elsewhere. In the left map below, the size of each country is warped to represent its portion of global historical carbon emissions. The United States, Europe, and China are the biggest emitters. The right map warps the countries according to the number of people injured, left homeless, displaced or requiring emergency assistance due to climate-related floods, droughts or extreme temperatures in a typical year. East and South Asia are undeniably most affected and most at risk. 

Toward Luddite Thinking

You might be thinking, “The future is inevitable, change happens, what can I even do about this?” To be truthful, the answer is not much. Personal consumption practices have proven to have little impact on global emissions. All of the energy statistics I listed are increasingly obsolete with every passing second. The ‘AI Cold War’ is only deepening the problem. However, if we think like a Luddite, it’s clear that there is a big difference between asking ChatGPT to generate a silly picture for personal use and accepting the integration of AI in our workplaces. While we may not be able to smash the data centers, we are in a special moment where we might refuse to cowork, collaborate, innovate, or think with AI. Refusing workplace AI is a refusal to collaborate with climate catastrophe, with hyper exploitation, with wage depression. It is nothing less than a stand for workers everywhere and for the future of the planet.