Two questions get asked about AI in the same breath and deserve straight answers: who owns what comes out, and what does running these things cost the planet. Both have real answers and a lot of noise around them.
The output: mostly nobody, and it varies by country
Copyright in most systems protects human authorship. Push a button and the machine produces an image, and in several major jurisdictions that image has no copyright owner at all.
The US Copyright Office has been explicit: purely AI-generated material is not copyrightable. In *Thaler v. Perlmutter*, US courts upheld the requirement of human authorship. Where a human contributes meaningfully — selection, arrangement, substantial editing — those human contributions can be protected, but the machine-generated parts are not.
The UK is unusual: its 1988 Act provides for computer-generated works with no human author, giving 50 years of protection to the person who made the arrangements for creation. This provision is contested and has been under review.
India's position is unsettled. Its Copyright Act allows registration for computer-generated works listing a human author, and the Registry has handled such cases inconsistently, including granting and then questioning registrations naming an AI as co-author.
The practical upshot for you: if you generate a logo and someone else uses it, you may have no infringement claim. If a client asks you to assign copyright in AI-generated deliverables, you may be assigning something that does not exist. Say so rather than discovering it later.
The input: genuinely unresolved
Whether training on copyrighted work without permission is lawful is being fought in court right now, and the honest answer is that nobody knows yet.
Cases are live in several countries — newspapers, authors, artists and music publishers against major AI companies. Some early rulings have found training itself can be transformative fair use while treating the acquisition of pirated copies as a separate and serious matter. Others have not reached judgment. Several have settled. Japan has a broad statutory exception permitting text and data mining. The EU allows TDM with an opt-out for rights-holders, and the AI Act requires providers to publish a summary of training content.
What you can rely on: reproducing a recognisable copyrighted character or a distinctive artist's signature style commercially is risky regardless of how it was produced. Some vendors now offer indemnification to business customers, which tells you they consider the risk real enough to price.
The environment, with actual numbers
This is where hype runs in both directions. Some figures worth holding.
A single text query to a large model uses electricity on the order of a few watt-hours — Google reported a median of about 0.24 watt-hours per Gemini text prompt in 2025, roughly nine seconds of a television. That is small. Multiply by billions of daily queries and it stops being small.
Training is a large one-time cost: training GPT-3 was estimated at roughly 1,287 megawatt-hours. But over a model's life, serving it to users generally exceeds training. Inference is the bigger number.
Image and especially video generation cost far more per output than text — image generation has been measured at hundreds of times a text query's energy.
The honest framing is at the level of data centres, not your chat window. The International Energy Agency projected global data centre electricity consumption roughly doubling to around 945 terawatt-hours by 2030, close to Japan's total consumption today, with AI the fastest-growing part. Water matters too: data centres consume water for cooling, and several are sited in water-stressed regions, which is a local justice problem rather than a global one.
And the emissions depend enormously on where the electricity comes from. The same computation in a coal-heavy grid and a hydro-powered one differ by an order of magnitude in carbon terms. This is why "how much CO2 does one prompt emit" has no single answer.
The usable conclusion: your individual chat use is a rounding error against your flights and your diet. The decisions that matter are corporate and governmental — where data centres are built, what powers them, whether efficiency gains are reported honestly, and whether a video model is used where text would do.
Before you move on