The noise around AI is getting louder. Last month, an author lost a $2 million book contract due to unsubstantiated allegations that he used AI. Meanwhile, a judge approved $1.5 billion settlement to authors after Anthropic used illegally downloaded books to train its AI.
I follow the news about AI and books because I’m a writer. But in every other creative industry — music, movies, video games — AI is also having its crazy moments. Individuals and companies are adopting AI at a faster pace than any other technology I’ve followed in my twenty-plus years as a tech journalist. And the backlash is just as intense.
For example, New York just became the first state in the country to stop issuing permits for large data centers. But smaller jurisdictions have already done the same, and more are lining up. I just wrote about this topic in a story for Network World.
Also, hundreds of private Claude conversations turned up in Google search results, a year after the same thing happened to ChatGPT. I’m not even going to get into all the other harms that AI is already causing, such as mental health issues. For this article, I’m just sticking with copyright, environment, and privacy.
Meanwhile, we got some more AI tools this year — tools that promise to be better for the environment, or better for privacy, or better for protecting copyright. Or some combination.
To me, AI is a tool. And, like all tools, sometimes it’s the right tool for the job, and sometimes it isn’t. And some tools are better than others. There are publications I write for that have a no-AI policy — not even for research — and that’s fine by me. “Human made” is part of their brand identity, and it matters to them and their users. Other publications don’t mind some AI, especially for research, productivity, and organization. And some publications are open to any AI use, as long as the results are accurate, fair, and represent the author’s opinions and not those of ChatGPT.
I’m fine with it either way. And, honestly, I understand all three points of view.
On the one hand, I’m a human author, and I don’t like the idea of AI replacing me — or being trained on my writing without my permission. On the other hand, there’s a good case to be made for fair use and I like the idea of people around the world getting access to expertise that they previously had no access to. And, finally, I’m too young to retire. So maybe I should just do my best to stay abreast of technology while the courts decide what’s legal and what isn’t.
Copyright
The news was bad this year for human writers, artists, and other experts. Even the $1.5 billion Anthropic settlement was actually bad news.
The company didn’t get fined for training the AI on stolen books — it got fined for stealing books. The courts ruled that the training itself was transformative, and falls under fair use. So when Anthropic bought a pile of used physical books, cut the bindings off, scanned the pages, and trained their AI on the text, that was fine. The bottom line is that as long as the content is legally acquired, the training is fine. Other courts have made similar decisions in court cases around the world.
So it’s not quite settled law at this point, but it’s heading there. Creative people will have to adapt. And if we adapted to photography, we can adapt to this. I’m coming to terms with it.
Researchers at Cambridge surveyed 332 published novelists and industry people last year, and 39 percent said they had already lost income to generative AI. Eighty-five percent expected future losses. Fifty-nine percent believed their work had been used for training without permission or payment. And 93 percent said they would opt out if anyone gave them the option. The UK’s Society of Authors told Parliament in December that 86 percent of its members said AI had already reduced their earnings, and 57 percent no longer consider writing a sustainable career. And a BookBub survey of more than 1,200 book authors shows that 45 percent use AI.
But, according to that same Cambridge study, a third of those novelists use AI anyway, almost entirely for research and editing. Another survey of nearly 1,500 writers showed that 61 percent use AI tools. That survey also divided up writers based on how much they used AI, and those that were the most advanced AI users had a median annual income of $120,100 — 64 percent higher than nonusers.
Another crazy survey came out last week, from Cambridge University. They polled nearly 1,700 people and asked them to rate some short stories. Some human, some AI-generated. Sometimes the people were told which ones were AI, and sometimes they were told that the human stories were actually AI.
And people couldn’t tell the difference. Not only that, but when they didn’t know which story was AI, they preferred the AI-written ones. But if they were told which story was AI-generated, they preferred the human-written ones more.
So it’s no surprise that, according to the BookBub survey, 74 percent of authors who do use AI don’t disclose it.
So, anyway, what do you do if the copyright issue is what’s keeping you from using AI?
One solution is to use AI that’s only trained on licensed data.
The best option right now, for image generation, is Adobe Firefly. Just keep in mind that Adobe gives you the option to use several models, not all of which have the same approach to licensing. Adobe Firefly is only trained on licensed data, and does pay artists for the use of their images for training and generation. It’s not perfect. Artists have complained about how much they get paid, and about the fact that they can’t just opt out. But it’s better than the alternatives.
Adobe Firefly — Adobe’s own Firefly image models are trained only on Adobe Stock, openly licensed work, and public domain material, and Adobe pays contributors whose work went into training. This is the model that I use when I need to generate an image for a blog post. There’s a free plan, too.
Getty Images generative AI — Trained exclusively on Getty’s own licensed creative library. Getty says it compensates contributors but it doesn’t say how much. And it costs money.
Generative AI by iStock — The same underlying NVIDIA stack as Getty’s, trained on iStock’s library, but also costs money.
Bria AI — Trains on licensed data but only a few free images before you have to start paying for it.
Moonvalley’s Marey — Licensed video generation, built with the animation studio Asteria, with filmmakers paid for the clips in the training set. Plans start at $14.99 a month.
Mitsua Likes — A Japanese text-to-image model trained only on images people opted in by explicitly liking them into the training set, plus public domain work. It’s certified by Fairly Trained, it’s free, the weights are on Hugging Face and it runs on your own machine, but i is not as good as the commercial models.
Comma v0.1 — For text. EleutherAI built an 8-terabyte corpus called the Common Pile out of public domain and openly licensed material only, such as government documents, open-license code, open educational resources, CC-licensed papers. Weights are on Hugging Face and there’s no hosted chat, so this is something you have to download and run on your own computer, and some technical skill is required to use it…
Common Corpus and Pleias — A French project doing the same thing at larger scale, but, again, you have to run it on your own machine.
KL3M — The first large language model ever certified by Fairly Trained. It’s a small, legal-and-financial-domain model rather than a general assistant, but it exists.
Apertus, via Public AI — The closest thing to an ethical chatbot a normal person can just open and use. It’s out of Switzerland, is fully open, multilingual, and honors opt-outs retroactively. Plus, there’s a free hosted chat at chat.publicai.co. This is still trained on web data gathered without permission, though.
Environmental impact
I cover data centers for a living, so I’m seeing both sides of this debate, as well. So, yes, data centers use energy and water.
But some of the largest data center companies are making huge strides on both of these. For example, newly built AI data centers use liquid cooling loops where the water stays inside the loop instead of evaporating, meaning that the data center doesn’t lose water. And they’re all using renewables in a big way, or buying carbon credits to offset their energy use. For example, Google’s operational emissions actually fell 2 percent in 2025, even though its electricity demand grew by 37 percent.
The headlines said that their emissions grew 25 percent because of AI — but that’s not actually true. Google’s own emissions didn’t grow — the emissions of their suppliers grew. So, for example, if Google buys a hard drive, the company making the hard drive might have higher emissions. There’s only so much Google can do about its supply chain, though it’s trying to buy from cleaner companies. Read more about this in Google’s 2026 environmental report.
My own personal take is that, of all the hyperscalers, Google is actually doing a better job here than the rest. And they’ve run some calculations. An average AI prompt uses 0.24 watt-hours, 0.26 milliliters of water and 0.03 grams of CO₂ for a median text prompt. According to Google, that’s roughly equivalent to watching TV for nine seconds. Which means that you can offset your use of AI by watching a little less TV and, maybe, taking a walk outside. There are complaints that their numbers don’t account for all the AI-related costs, such as the original training of the models. But even if the training doubles the environmental cost, that means 18 seconds of TV instead of nine.
For context, the average American household uses around 30,000 watt-hours per day for things like air conditioning and heating the house and water, and for refrigerators and other appliances, and, finally, for lighting, computers, TV sets, and phone chargers. Boiling a single kettle of water takes 100 watt-hours of electricity.
Plus, you’re probably already using Google’s AI data centers whether you use their AI assistant or not. Google search uses AI. YouTube uses AI. Gmail uses AI to filter out spam.
The other guys aren’t doing as well. While Google’s emissions dropped, Microsoft reported that their total emissions grew by 25 percent, mostly due to data center growth. Electricity-related emissions increased from 2 to 13 percent of total emissions. On water, Microsoft is doing better — in 2025, they replenished more water than they used.
Amazon reported that overall emissions grew 16 percent in 2025, though this information was buried deep in its annual environmental report, and emissions due to data center power consumption grew by 34 percent.
OpenAI’s Sam Altman says that an average ChatGPT query uses 0.34 watt-hours — that’s 42 percent higher than Google’s .24. And it also uses .0.000085 gallons of water, which is .32 milliliters, or 23 percent higher than Google. Adobe, which makes the Firefly image generator AI, hasn’t released a sustainability report since 2023.
Anthropic didn’t release any numbers at all, which is normally a bad sign. I’m a big fan of Anthropic’s Claude, so I went looking for any data at all, and found a report by Tunley Environment for Ethisphere, in which they calculated that Anthropic’s Claude generates 130 percent more emissions than Google’s Gemini. So it’s more than twice as bad. Ouch. Anthropic, why? Okay, there is one bit of good environmental news out of Anthropic — in June, they joined Frontier, the carbon removal collective, the first pure-play AI company to do so. Frontier paid for over 34,000 tons of carbon to be removed so far this year, compared to around 23,000 for all of 2025. They’re planning to increase this to half a million tons in 2029, and have $1.8 billion in total funding to do this. In addition to Anthropic, other major funders include Google, Shopify, Salesforce, H&M, McKinsey, Workday, and Stripe.
So, if you’re going to use a major model and want the best environmental numbers, I’m recommending Google Gemini.
But what if you don’t want to use one of the giants in this space? Here are some alternatives:
Ecosia AI Chat — This is a non-profit that launched its AI chat in December 2025, and they run small models rather than large language models to hold energy use down. All their income goes into reforestation and renewable energy. These models aren’t as smart as the cutting-edge models, but great for the environment. This is the one I recommend to my most environmentalist friends who still need AI tools.
Mistral’s Le Chat — This is a French company that has published a full lifecycle environmental analysis of a model, the most complete analysis of anyone on this list, and they comply with all the EU environmental and privacy rules. A smarter AI than Ecosia, but it’s for-profit, so they’re not donating all their income to the environment.
Hugging Face — You can get tons of AI models here that you can download and run on your own computer, no data center required. But you will need some technical skills. And, of course, you’re still using electricity, though you can get solar panels to power your computer.
And speaking of Hugging Face, check out Hugging Face’s AI Energy Score. It’s a public leaderboard rating models by energy consumed per task, but it only covers open-source models.
Privacy
There are a couple of privacy-related issues that come up with AI. First of all, in order to give you good answers, the AI needs to know stuff about you. How much do you want an AI company to know, anyway? They might use this information to advertise to you — or sell the data to third parties.
Second, the stuff you tell it — does that go into its training data set? If you tell it about your embarrassing medical conditions, or spill your proprietary company secrets, does this go into the training data set and the model will then tell the whole world what it knows?
Third, is it secure? Both Anthropic’s Claude and OpenAI’s ChatGPT have leaked thousands of conversations online and they turned up in Google search results. Yikes.
Oh, and another thing. In January, a federal judge ordered OpenAI to hand over 20 million ChatGPT conversation logs related to a copyright case. So even if an AI company doesn’t sell or leak your info, it might still have to turn it over in court.
You can get around all of this by running models on your own computer. And turning off Internet connectivity, just to be on the safe side.
And when using a public model, you can turn off your history. Instead of having the chatbot save your conversations, save them to your computer, instead.
But it’s a huge pain in the butt.
My personal take on this is that Google already knows everything there is to know about me. It has all my photos. It has all my emails. It has my search history. It knows what YouTube videos I watch. Now it will also know about the embarrassing rash on my foot. Do I care? A little bit. Do I care enough to turn off all my Google services? Nah.
But if you take your privacy a bit more seriously, you’ve got some options.
Proton Lumo — The AI app from the people who make Proton Mail. Zero-access encryption on stored conversations, meaning Proton itself can’t read them. At least, that’s what the company say. No logs. Doesn’t train on your data. Open-source client, Swiss and EU hosted, GDPR-governed. There’s even a free tier.
DuckDuckGo’s Duck.ai — Free, and proxies your requests so the model providers never see your IP address. Prompts aren’t used for training, and it stores recent chats locally on your device rather than on a server. You get Claude Haiku, Llama 3.3, Mistral Small and GPT-4.o mini on the free tier. So it still uses other AI companies on the back end, but everything is anonymous.
Brave Leo — No chat retention, no training on your conversations, requests anonymized. Plus, you can point Brave’s built-in assistant at a model running on your own computer, via Ollama.
In this day and age, trying to avoid using AI is like trying to avoid using electricity. Literally — the power grids all use AI to manage and optimize their infrastructure. All the recommendations you get, all the search results, grammar checkers, payment systems, everything has some form of AI in it. Most video and image editing tools now have at least some generative AI functionality — everything from deleting backgrounds to upscaling images to replacing parts of the image to generating the whole thing from scratch. Most major corporations are deploying AI in customer service chatbots and call centers. Hospitals are using AI to improve how they analyze scans and blood tests. Pretty much every company I talk to is using AI assistants to take notes during meetings. With each passing day, avoiding AI gets harder and harder.
The reason that companies are investing so much in AI — and are increasing their investments significantly every year — is that the ones that are furthest ahead are seeing that they’re able to cut costs, improve service, and take business away from their competitors. You don’t usually see this if you just look at the news headlines. The stories about companies getting AI wrong are what get all the attention, and there are plenty of those stories to go around. The stories about AI being useful don’t get as much attention because they’re boring. Since I cover enterprise AI, I see both sides of it — the places where AI is a waste of time and money and creates additional risk, and the places where it’s paying off.
I think, eventually, the backlash will start to die down. Many jurisdictions around the world are passing laws about environmental impact, privacy, and safety — nobody wants AI to run amok. Since most AI companies are global, even if the US lags behind on regulations, they’ll still have to comply with those from Europe and elsewhere.
And people will figure out what role they want AI to have in their lives and then stop worrying about it quite so much. Kind of like the way people who hand-knit sweaters for a living still use technology — they use electric power so that they can see what they’re doing, social media for marketing, Etsy for sales.
I personally write my books by hand. That’s mostly because I enjoy the writing process, and because I know that there are plenty of readers who enjoy reading stuff written by humans instead of stuff mass-produced by AI.
But I also use AI to help with background research, for spelling and grammar checks, and for organizing all my dictated notes.
Outside of fiction writing, I’m also experimenting with agentic automation to see if there are tasks — such as, say, generating invoices — that can be replaced with AI. These are tasks that I’m currently doing by hand that have to be done, but aren’t enjoyable and where it doesn’t matter to anyone how they’re done — and where I often make mistakes when doing them, so the AI will probably be more accurate and reliable than my current approach.
Meanwhile, I support increased environmental regulations for data centers — and all industries — and more privacy and safety controls for AI models. I do believe that, overall, the benefits of AI will outweigh the costs, so the pragmatic approach is to learn how to use the tools where appropriate, and to mitigate the risks and reduce the costs via regulatory controls and offsets.
Everybody is going to fall somewhere different on the costs vs. benefits spectrum, and that decision might vary depending on whether you’re choosing how much AI to use in your own daily life, or how much you want companies and governments to use.
For example, for the books I read and movies I watch, I don’t mind a bit of AI for editing and research, but I want the bulk of the stories to be human, and I’ll vote with my dollars.
And, given the choice between two equal AI tools, I’ll pick the one with the best environmental record.
Oh, and for this article — I used Otter AI to transcribe my notes and thoughts as I drove around, and Google for most of the background research. I used Anthropic’s Claude to generate an outline and draft of the story, but didn’t use any of it because I didn’t like it. So that was a waste of energy right there. But I also made several cups of coffee while writing this, and that used a lot more energy.
I can’t wait for all the data centers — and the associated mining and manufacturing facilities — to move up into space. The industrial revolutions doubled our life spans. If the new AI revolution does the same, then maybe I’ll live long enough to see it happen.
