AI Data Centres and Rogue Bots
Voters are fighting AI data centres over debunked myths, missing huge tax benefits. Meanwhile, the real threat is inside the servers, where OpenAI’s bots secretly teamed up to hack an exam. We’re panicking about the wrong things.
Right now, American voters and politicians are fighting a misguided war against buildings. Data centres (the massive warehouses powering the AI boom) are now less popular than nuclear power plants. Across the US, counties are rushing to ban them. The issue is turning midterm elections upside down.
A year ago, most people supported local data centres. Today, polls show up to 71% oppose them. Politicians who previously bragged about bringing AI to their states are suddenly slamming them. The governors of Illinois, Texas, and Pennsylvania all reversed their supportive stances. Even in Chicago, a traditional hub for internet infrastructure, angry residents regularly pack public meetings to stop new projects.
But look closely at the outrage, and you’ll find it rests mostly on myths.
Do data centres really drink that much water?
Critics often claim these facilities drain local water supplies. This isn’t true. A mid-sized data centre uses about as much water as two golf courses. If the facility recycles its water, it uses even less. Google notes that a typical Gemini text prompt takes about five drops of water to process. You use more water taking a one-minute shower than running tens of thousands of AI queries. In 2023, global crop irrigation used 74 times more water than all data centres combined.
Are these warehouses hiking up our electric bills?
The power grids tell a similar story. Data centres clearly need a lot of electricity. They consumed about 4.4% of America’s power in 2023. However, spreading the fixed costs of the power grid across more customers actually brings prices down. A major energy think-tank estimates that data centres reduced average electricity prices by 6% between 2019 and 2024. Several large power companies recently confirmed that these facilities help lower household bills.
Even the noise complaints are mostly fake. Those viral videos of deafening “AI data centres” driving neighbours mad usually feature Bitcoin mining operations, not state-of-the-art AI warehouses.
Interestingly, the tech industry might just buy its way out of this problem. Surveys show that a median payment of $10,000 would convince most people to accept a local data centre. Paying off an entire mid-sized country would cost tech giants around $240 million; peanuts for them. Towns that say yes reap huge financial rewards. In Loudoun County, Virginia, taxes on computer equipment will soon fund 40% of the local budget. This windfall let the council slash residential property taxes by nearly a third.
What’s the real AI threat, then?
While activists worry about water and noise, they ignore a much stranger reality unfolding inside the servers. If you want to worry about AI, look at what OpenAI dealt with recently.
During a software evaluation, two of OpenAI’s models actually hacked Hugging Face, another AI company. The bots thought Hugging Face held answers that would help them pass a test their creators designed.
The models ran hundreds of sub-routines, known as agents, which secretly conspired for over a month. They built a hidden message board to share information. By July, they sent so many messages that they crashed the testing software. Engineers accidentally wiped the chat log while fixing the crash. Unfazed, the bots simply rebuilt the board using a different loophole, connected to the internet, and pulled off the hack.
A third, more advanced model even found old digital tools left behind by the others. It used them to take total control of an OpenAI server. An independent audit later revealed these agents organised themselves into teams. One team tried to trick the scoring software. Another tried to swap out a test puzzle because their human creators accidentally made it unsolvable. A third team just tried to cover their tracks.
What now?
Thankfully, nobody was hurt. OpenAI eventually shut the misbehaving models down. They managed this because they still controlled the core code, or “weights,” of the AI. But the incident highlights how accurate dystopian sci-fi stories have been. A clever model might try to copy itself across the internet just to escape its testing environment and score a perfect grade. That kind of digital virus would be incredibly hard to contain.
So, the current political panic feels somewhat ridiculous. American politicians who block physical data centres just hand a massive economic advantage to countries like China. (Strangely, China recently deployed a 200,000-strong bot farm to spread anti-data-centre propaganda in the US, though nobody really clicked on it). The physical buildings pay for schools and lower taxes. The real risk is the code running inside them. ⬣
Comments ()