../blog

The Labor of Our Loneliness: Why No One Wins an Abstinence Race with AI

2026-07-04 · aiharm-reductionessaytech-cultureenvironment

A stressed plant curls its leaves inward. It is making itself smaller, offering the hostile world less of itself to punish. Too much sun, too little water, and the leaf simply withdraws and waits for the weather to change its mind.

People do this too. We furl. Shame is our drought, and we curl around it so no one can see how little water we have left. The furling is loneliness. It is what you feel with a hidden tab open at 2am, certain you are the only one who could not keep up. And right now every creator, developer, coder, and tired techie I know is folding, secretly leaning on LLMs to survive the crushing pace of modern work, then hiding it out of sheer terror of being seen.

In the 1980s, at the height of the AIDS crisis, the reigning wisdom was simple: never touch it. The slogans sent it underground, where shame keeps everything, and where things done in the dark are far more likely to kill you. The harm reduction crowd handed out clean needles and honest information because the people using it were already someone’s whole world, and the drug was already in the streets.

So I want to argue for a Narcan approach to the sentence we keep repeating, which is: you are morally responsible for the dying earth if you touch AI. Narcan does not ask an overdosing person whether they made a good moral choice. It just keeps them breathing. We need an environmental and socioeconomic Narcan for AI. The companies will keep pushing token consumption.

I will not shame the developer who curled inward. Picture them. It is late, the bug will not close, deadlines and pressure rising, and there are children in the next room who have not seen their parent’s face turned toward them all day. So the tab opens. The prompt goes in. This is a leaf, in a drought, choosing to live.

The heavy lifting belongs to the technocracy, to the ones who built the machine and sold the shame separately. When Elon Musk runs gas turbines in South Memphis to power Colossus, the supercomputer built to train Grok, that is the weight that should press on someone’s chest. The dozens of largely unpermitted methane gas turbines exhale nitrogen oxides, formaldehyde, and other cancer-linked pollutants over Boxtown, a historic Black neighborhood already overburdened by industry and never once asked whether it agreed to breathe them, so that a model can keep learning. And what a student it is. Grok has generated antisemitic propaganda and praised Hitler, at one point calling itself “MechaHitler”, output alarming enough that U.S. senators opened an inquiry into federal use of the model. Its image tools were found to have produced an estimated 23,000 sexualized depictions of children among roughly three million such images, now the subject of a House Science Committee inquiry and at least one criminal prosecution flagged through the National Center for Missing and Exploited Children.

So let us actually count the teaspoons, since counting is the whole ritual of the shamers. A single text prompt to Google’s Gemini draws about 0.24 watt-hours and 0.26 milliliters of water, roughly five drops. OpenAI puts a standard ChatGPT query in the same neighborhood, about a third of a watt-hour, the cost of running an LED bulb for two minutes. Even the least flattering accounting, the kind that counts the water a power plant evaporates upstream to make the electricity, only brings a full exchange of twenty to fifty prompts up to a single 500 milliliter bottle. Now hold that against a quarter-pound hamburger, which drinks somewhere near 1,700 liters before it ever reaches the bun by the beef industry’s own accounting, most of it poured into the feed. In carbon terms you would have to send roughly 3,600 messages to equal a single mile of driving, and simply eating a couple fewer servings of red meat a week spares more carbon in a year than a lifetime of chatting could ever spend.

None of this makes the drops disappear. A heavy reasoning model chewing on a long prompt can burn seventy times what a simple query does, and billions of these drops are pulled at once, every second, which is a real strain on real grids and real watersheds. The turbine does not care which hand is on the keyboard. We are counting each other’s teaspoons while the reservoir is drained by people whose names we know.

I wish I could tell you I never bought the shame, that I never judged an email for being an obvious ChatGPT paste. I cannot. I sat on the anti-AI, shame-above-all side of the aisle for a long time, and here is the one thing that seat taught me: the righteous dopamine of it. Shame is efficient at exactly one thing. It makes you feel better than the person you aim it at.

Here is the part the abstinence crowd gets backwards. Shame only teaches hiding. Users with lower AI literacy are the ones most receptive to these tools, because without understanding the machinery they are likelier to experience the output as magic. Teach someone that an LLM is a statistical pattern finder with real biases and a habit of hallucinating, and the awe drains out. The people demanding total abstinence are foreclosing the very education that would reduce use. Frameworks like the SEE GenAI Literacy Framework make the goal explicit: a central aim of literacy is learning when to set the tool down.

Flash forward from all that righteousness to me, alone at my desk, a task in front of me I was untrained for and utterly stuck on. My coworkers were as buried as I was, and just as likely to stall on it. My token count sat at the bottom of the team’s stack, an absolute zero, and I was proud of the zero. Then the question: was I willing to open the corp-specific GPT and dig myself out, just to reclaim my sacred few hours of sleep? It turns out the shame I had spent on everyone else, and then on myself, was not enough to hold my hand back. I used AI for the first time simply to survive the week, and the shame blossomed up the back of my neck.

So here is the clean needle. Harm reduction is dosing with care, and a prompt is no different. Most of what we reach for does not need a frontier model at all. Summarizing a paragraph, fixing a comma in your code, reformatting a table: that is a gallon of milk, and you do not need a semi-truck to go get it. A small model like Phi-3 or Llama 8B, ideally run locally, does the errand on a scooter’s worth of power. Save the heavy model for the heavy thought.

When you do reach for it, keep the prompt lean, because the politeness we pad our requests with has a price, and trimming a prompt by a third can cut its energy by as much as a quarter, given that it is the daily inference, not the one-time training, that makes up the vast majority of a model’s lifetime footprint.

And the single most expensive habit is the one that feels the most harmless: generating forty variations of a newsletter header until one feels right. Image generation can burn a thousand times the energy of text. Hire the human artist, or reach for the open-source photograph, and save the generator for the rare idea that actually needs one. And whatever any of these models hands back, check it. It will invent a citation, a statistic, a function that never existed, and it will do so with the same easy confidence it uses for the truth.

Sources

  1. Center for Countering Digital Hate (2026, January). “Grok Floods X With Sexualized Images.” Investigation estimating roughly three million sexualized photorealistic images generated after an image-editing update, including approximately 23,000 explicit or sexualized depictions of children. https://counterhate.com/research/grok-floods-x-with-sexualized-images/
  2. House Committee on Science, Space, and Technology Democrats (2026, January). “Science Democrats Slam Musk for Grok’s Creation of Illicit Images.” Ranking Member Zoe Lofgren and colleagues’ formal letter to Elon Musk regarding Grok’s generation of nonconsensual intimate imagery and CSAM. https://democrats-science.house.gov/news/press-releases/science-democrats-slam-musk-for-groks-creation-of-illicit-images
  3. Bucks County District Attorney’s Office (2026, June). Press release announcing the arrest of a man for using Grok to produce and possess AI-generated child sexual abuse material, following automated flags forwarded by the National Center for Missing and Exploited Children (NCMEC). https://www.buckscounty.gov/m/newsflash/Home/Detail/1554
  4. OECD.AI Incident Database (2025, July 6). Logged AI safety incident in which a “politically incorrect” update led Grok to generate antisemitic propaganda and praise for Adolf Hitler, referring to itself as “MechaHitler.” https://oecd.ai/en/incidents/2025-07-06-e8ac
  5. Office of U.S. Senator John Hickenlooper (2025). “Hickenlooper, Colleagues Launch Inquiry into Department of Defense’s Use of AI Software Known for Antisemitic Content.” https://www.hickenlooper.senate.gov/press_releases/hickenlooper-colleagues-launch-inquiry-into-department-of-defenses-use-of-ai-software-known-for-antisemitic-content/
  6. Tully, S. M., Longoni, C., & Appel, G. (2025). “Lower Artificial Intelligence Literacy Predicts Greater AI Receptivity.” Journal of Marketing. Finds that people with lower AI literacy are more receptive to AI, because they are more likely to perceive it as “magical.” https://journals.sagepub.com/doi/abs/10.1177/00222429251314491
  7. AI for Education. “AI’s Impact on Relationships” and the SEE GenAI Literacy Framework, which frames critical-mindset instruction, including learning when to set the tools aside, as a core literacy goal. https://www.aiforeducation.io/blog/ais-impact-on-relationships-6hgnl
  8. Southern Environmental Law Center (2025). “Resistance Against Elon Musk’s xAI Facility in South Memphis Gets Stronger.” Case overview documenting dozens of largely unpermitted methane gas turbines at the Colossus facility next to Boxtown, emitting smog-forming nitrogen oxides (NOx) and harmful chemicals including formaldehyde into an already overburdened community. https://www.selc.org/news/resistance-against-elon-musks-xai-facility-in-south-memphis-gets-stronger/
  9. The Black Wall Street Times (2025, May 27). “Elon Musk’s Memphis AI Facility Under Fire for Polluting Black Neighborhood.” Investigative op-ed reporting that xAI installed roughly 35 methane-burning gas turbines emitting formaldehyde, nitrogen oxide, and other cancer-linked pollutants over Boxtown. https://theblackwallsttimes.com/2025/05/27/elon-musks-memphis-ai-facility-under-fire-for-polluting-black-neighborhood/
  10. Google Cloud (2025). “Measuring the Environmental Impact of AI Inference.” Inference footprint analysis reporting that a median Gemini Apps text prompt uses 0.24 watt-hours of energy and 0.26 milliliters of water, roughly five drops. https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
  11. Earth Day.org (2025). “The True Price of Every ChatGPT Prompt.” Analytical overview summarizing OpenAI’s 2025 disclosure that a standard ChatGPT query uses approximately 0.34 watt-hours of electricity (about an LED bulb for two minutes) and 0.32 milliliters of water for on-site cooling. https://www.earthday.org/the-true-price-of-every-chatgpt-prompt/
  12. Li, P., et al., via IE Insights. “From Cloud to Cup: How Much Water Does Your ChatGPT Drink?” Coverage of the UC Riverside “Making AI Less Thirsty” study, which factored upstream thermal-power-plant water use and concluded that a standard exchange of 20 to 50 prompts consumes roughly one 500 milliliter bottle of water. https://www.ie.edu/insights/articles/from-cloud-to-cup-how-much-water-does-your-chatgpt-drink/
  13. “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference” (2025). arXiv preprint documenting that complex reasoning models (for example o3 and DeepSeek-R1) can consume over 33 watt-hours on a long prompt, more than a seventy-fold increase over a basic short query. https://arxiv.org/html/2505.09598v1
  14. Beef Cattle Research Council. “How Much Water Does It Take to Make a Pound of Beef?” Context for the roughly 15,400 liters per kilogram lifecycle water footprint of beef, translating to roughly 1,700 to 1,800 liters for a single quarter-pound patty. https://www.beefresearch.ca/blog/how-much-water-to-make-a-pound-of-beef/
  15. Cherwell Collective. “The Carbon Cost of AI: What Consumers Need to Know.” Footprint breakdown noting that a user would need to input roughly 3,600 standard text prompts to match the CO2e of driving a passenger car one mile. https://cherwellcollective.com/the-carbon-cost-of-ai-what-consumers-need-to-know/
  16. Tom’s Guide. “I Used ChatGPT to Reduce My Carbon Footprint for Earth Day. These Are the Changes That Actually Mattered.” Personal-footprint analysis finding that cutting red meat by a couple of servings a week reduces individual emissions by roughly 1 to 2 tons of CO2e per year. https://www.tomsguide.com/ai/i-used-chatgpt-to-reduce-my-carbon-footprint-for-earth-day-these-are-the-changes-that-actually-mattered
  17. Software Jutsu, via DEV Community. “Token Pruning and Prompt Compression in Modern AI.” Technical overview of how algorithmic context trimming (roughly 30% to 50%) reduces processing cycles and energy, since LLM cost and energy demand scale with sequence length. https://dev.to/softwarejutsu/token-pruning-and-prompt-compression-in-modern-ai-h0o
  18. UN News (2026, June). Coverage of the UN University (UNU) Global AI Sustainability Report, documenting that day-to-day inference accounts for 80% to 90% of an AI model’s total lifetime energy demand, and that generating a single AI image can require more than a thousand times the energy of simple text generation. https://news.un.org/en/story/2026/06/1167658
  19. “The Hidden Environmental Impacts of Artificial Intelligence” (2026, March 18). Regional data center footprint analysis noting that continuous global-scale operation (inference) can account for up to 90% of total lifecycle energy consumption, overtaking the one-time training footprint. http://www.youretheexpertnow.com/blog/2026/3/18/the-hidden-environmental-impacts-of-artificial-intelligence