Understanding how models are trained and generate their output helps us distinguish between training on existing work, reproducing it, and committing plagiarism. There are still legitimate questions about consent, copyright, and compensation. There is a difference between a model reproducing an existing text and one drawing on patterns learned from millions of texts to compose a new answer. With code, things get even more interesting. After all, developers also use existing solutions, libraries, documentation, examples, and code written by others. Of course, we cannot simply copy everything. Licences and copyright still matter. But we generally see nothing strange about developers learning from other people’s code.

Rather than asking whether AI “steals”, perhaps we should ask where to draw the line.

Many companies use open-source projects or libraries to build commercial products while keeping their own source code private. And that is not automatically wrong. Open-source licences differ in what they allow and what they require. Some licences permit commercial use without requiring derivative software to be made public. Others impose additional conditions. But it does raise an interesting question:

This may be one of the trickiest points, because people often lump several different debates together. An outspoken critic—or supporter—can bring up privacy, copyright, concentration of power, misinformation, environmental impact, and the effects on employment all in one breath. I think it is more helpful to address those issues separately. I would expect different safeguards for an AI application handling medical information than for one helping me write a regex or recreate Grandma’s apple pie. Still, we should not overlook the ethical questions AI raises.

These are questions we need to take seriously. They call for clear agreements and, where necessary, guidelines and regulation.

Training large AI models requires a lot of energy. Using those models also consumes energy. And those energy costs come on top of hardware manufacturing, data centre construction, and any water needed for cooling. But context matters here as well. A trained model can then be used millions of times. We should therefore consider the energy needed for training alongside the energy consumed over the model’s lifetime. Running a model can consume substantial amounts of energy in its own right, especially at scale.

Fortunately, work is also being done on more efficient models and hardware. Not every application needs the largest model available, and using smaller or specialised models can sometimes produce substantial savings.

But how much difference does that actually make? We need to look at the bigger picture. Data centres, cloud computing, and other digital services also consume energy. AI is therefore not the only technology drawing on our energy supply. The additional demand from AI still deserves attention and may be substantial.

Still, modern data centres and cloud providers are increasingly using renewable energy sources. Research into better algorithms and hardware also aims to make AI models more energy-efficient. Reusing existing models can also help reduce energy consumption. If an existing model meets an application’s needs, training a new one from scratch may be unnecessary. This is where access to model weights matters. Several large AI models are available as open-weight models, allowing others to run, adapt, or further train them. This can encourage reuse and prevent unnecessary duplication of effort. Not every model is fully open-source, of course, and running, adapting, or further training existing models still consumes energy. Still, making model weights available could help us use resources more efficiently.

An older model is not necessarily worse than a newer one, either. If it is good enough for your intended application, why automatically choose the newest and largest model? Techniques such as retrieval-augmented generation (RAG) can also give an existing model access to current or domain-specific information without retraining. This can keep an older model useful for longer in certain applications. Of course, RAG does not make every model equally suitable for every task. It provides access to information from outside the model; it does not automatically overcome limitations in reasoning or other capabilities.

The idea of AI taking over the world is a popular theme in clickbait articles and films. It conjures up images of a dystopian future in which AI rules over humanity. But reality is rather more nuanced. AI is a technology developed and deployed by people. Today’s AI systems do not need human-like will or consciousness to be useful—or to pose risks. Even without the ability to make independent decisions about their own goals, these systems can take on an increasing number of tasks and influence decisions that were once made by humans. As those systems become more powerful and are deployed more widely, their potential impact grows. Perhaps we should ask how much influence we are willing to give AI over society—and how much control we want to keep. Ultimately, people decide how AI is developed, who has access to it, and what the technology is used for. This makes our responsibility all the greater: AI’s power demands careful thought about the limits we set and the control we want to retain.