Everyone Hates AI Until It Cures Cancer

Introduction

Say the word AI at a dinner party and watch the room split. One side has read about data centers draining a town's water and gas turbines running next to people's homes. The other side has read about software that designs new medicines and folds proteins that stumped biology for fifty years. Here is the uncomfortable part. Both sides are right.

We keep arguing about AI as if there are only two settings. Ban it, or build it at any cost. Burn it down, or burn the world to feed it. Neither of those is a plan. They are just two ways of refusing to do the harder work, which is to keep the good and refuse to pay for it with people's health.

The headline of this post is meant to be a little unfair, because the sentiment behind it is real. A lot of people hate AI right now, and they have earned the right to. They would also take an AI in a heartbeat if it found their tumor a year before a human would have. That is not hypocrisy. That is just a normal human being asking for the upside without the harm. The whole point of this piece is that asking for both is not naive. It is the only sane position left.

The Backlash Is Earned

Before anyone defends AI, they have to be honest about why people are angry. The costs are not imaginary and they are not evenly shared.

Start with water. A peer reviewed study from researchers at UC Riverside and UT Arlington, Making AI Less Thirsty, estimated that training GPT-3 in Microsoft's US data centers directly evaporated about 700,000 liters of clean freshwater, and around 5.4 million liters once you count the water used to generate its electricity. The same paper estimated that GPT-3 consumes roughly a 500ml bottle of water for every 10 to 50 medium length responses. Those are model based estimates for a 2020 era system, and newer models are more efficient, but the order of magnitude is the point. The water is real and it has to come from somewhere.

Now electricity. The International Energy Agency reported in Electricity 2024 that data centers used an estimated 460 terawatt hours in 2022, about 2 percent of all electricity on earth, and that the combined demand of data centers, AI, and crypto could pass 1,000 terawatt hours by 2026, which is roughly the entire electricity consumption of Japan. The IEA's later Energy and AI report, released in April 2025, projects data center demand more than doubling to around 945 terawatt hours by 2030, with demand from AI optimized data centers specifically set to more than quadruple. In the United States, a December 2024 Lawrence Berkeley National Laboratory report for the Department of Energy found data centers consumed about 4.4 percent of national electricity in 2023 and could reach anywhere from 6.7 percent to 12 percent by 2028.

Then there is the part that does not show up in a chart, because it shows up in someone's lungs. In the Boxtown neighborhood of South Memphis, a predominantly Black community already living next to an oil refinery and a gas plant, xAI's Colossus supercomputer ran a fleet of methane gas turbines to power itself, with aerial and thermal imaging from the Southern Environmental Law Center documenting as many as 35 of them operating before the county issued a permit for 15. The NAACP and the SELC filed a notice of intent to sue under the Clean Air Act in June 2025 and appealed the permit a month later. In The Dalles, Oregon, a town in a multiyear drought, public records pried loose after the city dropped a Google funded lawsuit revealed that Google's data centers used 355.1 million gallons of water in 2021, close to 29 percent of the entire city's water.

The honest version: when people say AI is making them sick or drinking their water, they are not being dramatic. In specific places, with specific communities footing the bill, it is literally true. Any argument for AI that skips over this part is not worth listening to.

What AI Has Actually Solved

Here is the other half of the ledger, and it is just as real. Not promises, not demos, not a founder on stage. Published, peer reviewed, in some cases already running in hospitals.

Protein folding, the fifty year problem. DeepMind's AlphaFold predicted the 3D structures of more than 200 million proteins, nearly every protein known to science, and released them in a free public database. Working out a single protein's shape used to take a PhD student years in a lab. The work was considered foundational enough that Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry with David Baker. This is the engine under a generation of new drug and disease research.

Catching cancer earlier. This is the line in the title, and it is not a metaphor. In the Swedish MASAI randomized controlled trial, reported in The Lancet Oncology in 2023, more than 80,000 women were screened, and AI supported mammography reading cut radiologists' screen reading workload by about 44 percent while detecting more cancers, 6.1 per thousand versus 5.1 with standard double reading. The full results, published in early 2026, showed higher sensitivity and fewer of the cancers that slip through between screenings. Earlier work pointed the same direction. A Google Health and DeepMind system reduced both false positives and false negatives in breast cancer screening in a 2020 Nature study, and a 3D model from Google matched or beat six radiologists at spotting lung cancer on CT scans in Nature Medicine, hitting a 94.4 percent area under the curve.

Finding drugs and antibiotics we did not know to look for. Using a graph neural network, researchers led out of MIT discovered halicin, a structurally novel broad spectrum antibiotic, published in Cell in 2020. The same platform later screened more than 107 million molecules and surfaced eight more candidates. A follow on collaboration between McMaster University and MIT used AI to find abaucin in 2023, an antibiotic that targets Acinetobacter baumannii, one of the drug resistant superbugs the World Health Organization considers most dangerous. And Insilico Medicine's INS018_055, the first drug designed by generative AI against an AI discovered target to reach human trials, reported positive Phase 2a results for idiopathic pulmonary fibrosis, a brutal lung disease, with the trial later published in Nature Medicine.

Reading the genome for danger. DeepMind's AlphaMissense, in Science in 2023, classified all 71 million possible single letter mutations in human proteins, labeling 89 percent as likely benign or likely harmful. Before that, only about 0.1 percent had been clinically classified. That is a map for finding the genetic typo behind a disease.

Watching patients who are about to crash. At five hospitals, a machine learning sepsis warning system called TREWS monitored more than 590,000 patients, and for the alerts a clinician confirmed within three hours, in hospital mortality was 18.7 percent lower in a 2022 Nature Medicine study. Sepsis kills by being missed, so buying hours matters. A separate DeepMind model predicted acute kidney injury up to 48 hours in advance, catching roughly 90 percent of the cases severe enough to later require dialysis.

And it is not only medicine. AI discovered 2.2 million new stable materials in one shot, found the first faster matrix multiplication algorithm in 50 years, held the plasma steady inside a fusion reactor, and now beats the world's best supercomputer weather models on the large majority of forecasts while running in minutes instead of hours. It even improved a 20 year old result in pure mathematics and scored at silver medal level at the 2024 International Mathematical Olympiad.

Are any of these a finished cure you can pick up at a pharmacy tomorrow? No. Most are screening tools, early warnings, research breakthroughs, and candidates in trials. But that is exactly how progress in medicine and science has always looked, one foundational tool at a time, and AI is now one of those tools.

The False Choice

So we have two stacks of real, sourced facts sitting on the same table. AI is hurting specific people in specific places. AI is helping enormous numbers of people in ways that were science fiction a decade ago. The mistake almost everyone makes is to grab one stack, sweep the other onto the floor, and call it a worldview.

One camp says the harms prove AI is rotten and should be stopped. The other says the breakthroughs prove we should build at any cost and the rest is collateral. Both are taking a complicated thing and flattening it into a slogan, because a slogan is easier to defend than a trade off.

Key Insight: All or nothing is not a moral stance. It is a way to avoid doing the engineering. Saying "ban it" lets you ignore the cancers it would have caught. Saying "build it anyway" lets you ignore the neighborhood breathing your exhaust. The grown up question is not whether to have AI. It is how to have the good parts without making other people pay for them with their water, their air, and their power bill.

Responsible AI Is the Middle Ground

Here is the good news that the doomers and the accelerationists both tend to skip. The trend line in AI is not only toward bigger and hungrier. A huge amount of the recent progress has come from making AI smaller, cheaper, and far more efficient. The middle ground is not a compromise we are begging for. It is already where a lot of the real work is going.

Look at the cost curve. Stanford's 2025 AI Index found that the price of running a model as capable as GPT-3.5 fell from about 20 dollars per million tokens in late 2022 to 7 cents by late 2024, a drop of more than 280 times. At the hardware level, the same report found AI compute costs falling about 30 percent per year while energy efficiency improves about 40 percent per year. Capability is being pulled apart from raw power consumption, and that is the whole ballgame.

Look at model size. In 2022, clearing a 60 percent score on the MMLU benchmark took Google's PaLM and its 540 billion parameters. By 2024, Microsoft's Phi-3-mini cleared the same bar with 3.8 billion, about 142 times smaller, and it runs fully offline on an iPhone 14 at more than 12 tokens per second in under 2 gigabytes of memory. DeepSeek-V3 reached frontier class quality with a final training run of 2.788 million GPU hours, roughly an order of magnitude less compute than a comparable model, though that figure is the final run and not the full cost of research behind it. Mistral Small 3, at 24 billion parameters, runs more than three times faster than a model triple its size at similar quality, and Meta's quantized Llama 3.2 models run on the phone in your pocket. Every task that a small local model can handle is a task that never touches a thirsty data center at all.

The infrastructure is moving too, though here we should keep our skeptic hat on, because most green claims are self reported and not independently audited. Microsoft has designed a closed loop cooling system that evaporates zero water and is meant to avoid more than 125 million liters per data center each year, with pilots arriving in 2026 and 2027. Amazon says it matched 100 percent of its electricity with renewable energy in 2023, though that is annual matching through purchase agreements, not 24/7 clean power on the same grid. And Google's 2025 Environmental Report is a useful dose of honesty in both directions. The company replenished about 64 percent of the freshwater it used in 2024, up from 18 percent the year before, but it is not yet water positive, its data center electricity demand rose 27 percent in a single year, and its total emissions are up about 51 percent since 2019. The fixes are real and so is the strain. Pretending otherwise in either direction is the problem.

Finally, the rules are starting to catch up. The EU AI Act, in force since August 2024, requires providers of general purpose AI models to document the energy consumption of the model. You cannot manage what you refuse to measure, and mandatory energy transparency is how you turn vague pledges into numbers someone can check.

Key Insight: The most important question in AI is no longer "how big can we make the model." It is "what is the smallest, most efficient system that actually solves the problem." Those are very different engineering cultures, and the second one is where responsibility and good business happen to point the same way.

Conclusion: The Middle Ground Is the Only Ground

Extremism is the easy move. It always is. Burn it down or build it at any cost both let you stop thinking. The harder and better path is the one almost nobody is selling, which is to want the cancer screening and refuse the poisoned air, to chase the protein folding and the new antibiotics while insisting that no community has to give up its water to get them.

That is not fence sitting. It is a set of concrete choices. Use the smallest model that does the job. Run on the edge when you can so the work never hits a data center. Demand the efficiency numbers and the water and energy disclosures before you sign anything. Put the heavy compute where the power is clean and the community is not bearing a hidden tax. None of this requires giving up the breakthroughs. It requires engineering them on purpose instead of pretending the bill never comes due.

People do not actually hate AI. They hate paying for someone else's progress with their own health. Solve that, and the dinner party stops splitting in two.

At NeuroCore Technology, this is the work. We help organizations get real value from AI while making the early architecture choices that keep it efficient, sovereign, and honest about its footprint, across cloud, on premises, and hybrid. If you want the upside without burning the world to get it, partner with NeuroCore and let's build it the right way.

Sources & Further Reading

Everyone Hates AI Until It Cures Cancer

Introduction

The Backlash Is Earned

What AI Has Actually Solved

The False Choice

Responsible AI Is the Middle Ground

Conclusion: The Middle Ground Is the Only Ground

Sources & Further Reading

Medicine and Biology

Science and Engineering

Energy, Water, and Community Impact

Efficiency and the Middle Ground

Policy

Have a problem worth building?