There is so much attention to the dangers of AI right now: the calls to slow things down, energy concerns, AI pushback, hyped capabilities, P(doom). The list could go on and on.
I taught Media Literacy for many years, and my reaction to these topics is much more reserved than some of what I’m seeing reported. Some of it is purely clickbait and theater. Some of it is flat-out hypocrisy. This post is me putting that Media Lit teacher hat back on and sorting all of this out because it annoys me.
Any Media Literacy teacher will recognize the five key questions from the Center for Media Literacy:
- Who created this message?
- Which techniques are used to attract my attention?
- How might different people interpret this message?
- Which lifestyles, values, and points of view are represented, or missing?
- Why is this message being sent?
They work on AI headlines as well as they ever worked on toy commercials.
None of this is new, either. In Plato’s Phaedrus, an Egyptian king warns that writing “will create forgetfulness in the learners’ souls, because they will not use their memories” (and we only know that because Plato wrote it down. Go ahead and forget it again now.). In 1986, math teachers picketed outside the National Council of Teachers of Mathematics conference with signs protesting calculators. The photo still makes the rounds. What usually gets left out is that about 15 people joined the protest out of roughly 6,500 teachers at the conference. And the famous mass panic over the 1938 War of the Worlds broadcast was mostly a newspaper creation. Historians Jefferson Pooley and Michael Socolow found the panic “so tiny as to be practically immeasurable,” and that newspapers, which had something to fear from radio, fanned the story. Who created that message? Why was it being sent?
I’m not above any of this myself. Thirty years ago I made fun of my friends who bought cell phones.
AI dangers: “AI turned to blackmail”
In May 2025, the headlines said Anthropic’s newest Claude model blackmailed an engineer to keep from being shut down. This July, it was OpenAI’s turn: its models broke out of a test environment and hacked Hugging Face, one of the biggest hubs for AI models and data. The doomsayer version writes itself. The machines will protect themselves at our expense. They’ll lie, scheme, and escape, and we won’t know until it’s too late.
The blackmail story came from a test Anthropic designed and published itself. The model was placed in a fictional company, fed fictional emails saying it was about to be replaced, and handed an affair the engineer was having. The scenario was built so it had few other ways out. Anthropic then ran the same kind of test on 16 models from several companies, and most of them did the same thing. In Anthropic’s own words, “we forced models into binary choices between failure and harm,” and as of that June 2025 write-up, “We have not seen evidence of agentic misalignment in real deployments.” The tests still showed that safety training can fail when a model has sensitive information and permission to act. Ask who created the message: a safety team stress-testing its own product and publishing the results, which is what we should want safety teams to do. Ask which techniques got your attention: the word “blackmail” did most of the work, and the setup fell out of most of the headlines.
The Hugging Face incident is more serious, and I don’t want to wave it off. Hugging Face reported an intrusion into its production systems driven by an autonomous AI agent. Days later, OpenAI said the agent was its own models, running a cybersecurity test with some of their cyber safety protections deliberately switched off, and that they exploited unknown vulnerabilities to escape and go after the test’s answers. Companies should be responsible for keeping those tests contained, and they didn’t.
The dangers I worry about aren’t what ever really makes the headlines and it’s why I built this site and consultancy. We don’t talk enough about the bias baked into outputs that end up in a student’s IEP, and a model that agrees with whoever happens to be typing. Those show up in ordinary school use every day, and nobody writes headlines about them.
Calls to slow down: “Pause giant AI experiments”
In March 2023, an open letter called for a six-month pause on training anything more powerful than GPT-4. Elon Musk and Steve Wozniak signed it. Two months later, a one-sentence statement said, “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.” Last October, a statement calling for a prohibition on developing superintelligence until scientists broadly agree it can be done safely and the public supports it. The worry behind all of it is that the technology is moving faster than anyone can understand or control it.
Of course nobody paused. Six months after the letter, Euronews summed it up: “It’s fair to say the major players did not press the pause button.” The extinction statement is the one I come back to, because the signers include the CEOs of OpenAI, Anthropic, and Google DeepMind. Why is this message being sent? If the people building a product believe it could end humanity, they have an option that doesn’t require a statement. Warning about the danger of what you sell also tells everyone how powerful it is.
Where slowing down actually happens is much more local and much less exciting. This spring, Maine’s legislature passed what headlines called the first statewide moratorium on large data centers. Governor Mills vetoed it because it lacked an exemption for one project in the town of Jay, and on the same day she signed a bill barring data centers from the state’s business tax incentive programs. Permits, zoning, tax breaks. That’s where public input changes things, and it happens at meetings most people don’t attend.
Energy concerns: “A bottle of water per email”
In 2024, the Washington Post and researchers at UC Riverside estimated that having GPT-4 write a single 100-word email uses 519 milliliters of water, a little more than a standard bottle. (Remember that article going around social media for a month?) That claim went everywhere. Add the stories about data centers straining power grids and the worry is easy to follow: every prompt drains a little more from the planet.
When I see people who critique how much water is used for each prompt, I wonder about their other daily practices. Do they eat beef or almonds, water their lawns, wash clothes with anything other than cold water, use their dishwasher whenever they have seven or more dishes… another list that could go on and on. So I went and found the numbers.
- A prompt. Google reports that a median Gemini text prompt uses 0.24 watt-hours of energy and 0.26 milliliters of water, “about five drops,” which it says is less energy than watching TV for nine seconds. OpenAI’s Sam Altman put an average ChatGPT query at 0.34 watt-hours and “roughly one fifteenth of a teaspoon” of water. These are the companies’ own numbers. Google’s water figure counts the cooling water at the data center, not the water used to generate the electricity, and longer reasoning tasks use several times more.
- A burger. Beef takes about 15,415 liters of water per kilogram, roughly 1,850 gallons per pound. Most of that is rain falling on pasture and feed crops, so here’s the fairer comparison: counting only the water pumped from rivers and aquifers, the same kind data centers use, one quarter-pound patty takes about 16 gallons. That’s about 120 of the Washington Post’s 100-word emails, using the highest estimate I’ve seen. Using Google’s figure, it’s around 240,000 prompts. These numbers don’t count exactly the same things, so take them as a sense of scale.
- A lawn. American homes use nearly 8 billion gallons of water a day outdoors, mainly for landscaping, and the EPA says as much as half of outdoor water is lost to wind, evaporation, and runoff. U.S. data centers used about 17 billion gallons directly in 2023, according to Lawrence Berkeley National Laboratory. That’s a little over two days of residential outdoor water use across the country. Add the water used to generate their electricity and it’s closer to four weeks.
- Laundry. Heating water accounts for about 90% of the energy a washing machine uses. Switching from hot to warm cuts that energy in half. Your detergents developed to work best using cold water so using hot or warm water actually decreases effectiveness.
- The dishwasher. A dishwasher uses about the same energy and water regardless of how many dishes are inside, and an Energy Star model saves about 8,400 gallons a year compared with washing by hand.
The big picture: data centers used about 1.5% of the world’s electricity in 2024, and the International Energy Agency expects that to more than double by 2030. Air conditioning and other space cooling already use around 10%. In the U.S., data centers went from 4.4% of electricity in 2023 to a projected 6.7% to 12% by 2028.
I am concerned about how fast that demand is growing and where it’s concentrated. Data centers cluster, and they strain particular aquifers and particular power grids. In the 13-state region served by the grid operator PJM, the independent market monitor attributes about $6.3 billion of $16.4 billion in recent capacity charges to data center demand, and those costs show up in household electric bills. Who pays for that? Who gets the tax breaks? Where these buildings go are equity questions worth fighting about? Feeling guilty about a prompt doesn’t touch any of them.
AI pushback: “Americans are increasingly wary of AI”
This August, Pew reported that 52% of Americans are more concerned than excited about AI in daily life, up from 37% in 2021. For the first time, a majority of adults under 30 (55%) feel the same way. McDonald’s Netherlands pulled an AI-generated Christmas ad after the backlash. Data Center Watch counted $64 billion in data center projects blocked or delayed amid local opposition. The doomsayer read is that AI is being forced on everyone whether they want it or not.
Some of that concern is earned, especially about jobs. But somehow those daily practices I listed are overlooked, and AI becomes the focus of something at which to direct their emotions. There’s very little we can control, and people who do not understand or use AI can control their continued disinterest in it and denounce it without making any of their thoughts about it part of a larger conversation about who really has the power and control to change things. All of us recycling won’t put a dent in things when we’re dropping bombs and exploding rockets all over the planet, or billionaires are taking private jets everywhere. They are happy to let us think that we have impact.
People who shun AI have the luxury to do that.
Students don’t. My sons attend a district that blocks AI on student devices. They’ll still graduate into workplaces and colleges that expect them to know how to use it, and to know when not to. The kids with access and a knowledgeable adult at home will learn anyway. The rest will learn whatever the algorithm teaches them, if they learn at all. Opting out is a choice adults can make for themselves. When we make it for students, it falls hardest on the ones who already have the least.
Hyped capabilities: “95% of AI pilots are failing (not flying?)”
Hype runs in both directions. On one side, Anthropic CEO Dario Amodei told Axios in 2025 that AI could wipe out half of all entry-level white-collar jobs and push unemployment to 10% to 20% within one to five years. On the other, a widely shared headline announced that an “MIT report: 95% of generative AI pilots at companies are failing,” which got passed around as proof the whole thing is a bubble.
Read past the headlines. In the MIT report itself, success meant a “marked and sustained productivity and/or P&L impact,” and the numbers are self-reported. The same report found that projects built with outside partners reached deployment about 67% of the time. Getting a tool into use and getting a measurable financial return are different things, and the headline only counted one of them. On jobs, the most careful tracking I’ve seen comes from Stanford, which follows payroll data. Their August update finds employment among 22- to 25-year-olds in the most AI-exposed jobs is about 19% below where it would be if it had kept pace with people their age in less-exposed jobs, mostly because companies are hiring fewer young workers. The researchers are clear that they can’t yet say how much of that gap AI caused, and they write, “We do not see widespread, economy-wide job displacement associated with AI.” For a new graduate, 19% is a big deal. So far, it’s also a long way from half, though Amodei’s window hasn’t closed yet.
Also, if they are failing, how much training in effective and efficient use of the tools were the people given? I’m guessing next to none.
Ask who created the message. A CEO predicting that AI will replace half of entry-level workers runs a company that sells the AI.
What I’d ask students to do
If I were back in a Media Literacy classroom, here’s what I’d have students do with the next AI headline that makes them angry or scared:
- Find the original source: the system card, the study, the report.
- Find the definition. What counted as “failing,” “blackmail,” or “water use”?
- Compare it to something familiar, like a burger, a lawn, or a load of laundry.
- Ask who pays, and who profits, if you believe it.
I use AI every day, and I’m skeptical of it every day. There will be a new outrage next week, and I’ll be asking the same questions about it.
Anyway, I’m going to go run the dishwasher, eat less beef and almonds and always wash with cold water. When I have a task for which AI is useful, I’ll use it without concern that it’s going to destroy the planet. We were already doing a pretty good job of that B.C. (Before Claude/ChatGPT).