What Does "Limiting AI" Actually Mean? Three Ways AI Can Be Constrained
"Limiting AI" is a phrase that comes up often, in news coverage, political debate and industry commentary alike, but it rarely means one single thing. Depending on the context, it tends to fall into one of three quite different areas: physical constraints, regulation of increasingly powerful AI, or regulation of access to AI, and these areas have little in common beyond the word "AI."
The first is a physical constraint rather than a policy choice: the power, water and infrastructure needed to actually build and run AI. Data centres draw large amounts of electricity and water, and expansion in a given area depends on whether local grids have spare capacity and whether water supply can support additional cooling demand. This is also the area facing the most visible public pushback. A March 2026 Gallup poll found 71% of Americans oppose the construction of a new AI data centre in their local area, with nearly half strongly opposed. Limits here tend to come from grid connection constraints, local planning decisions, or that kind of community opposition, rather than from a specific AI policy.
The second is regulation of how powerful AI is allowed to become, ongoing discussions, still largely unresolved, about whether there should be limits on the scale or capability of a single AI model, and what safeguards should apply as models grow more advanced. In practice, this usually means drawing a line based on how much computing power went into training a model, since that has generally correlated with how capable it is, with regulators in different countries setting their own thresholds for when extra safety testing and reporting requirements should apply. Major AI developers, including OpenAI, Anthropic, Google and xAI, have also signed up to a voluntary code committing them to evaluate and disclose safety risks in their most advanced models before release.
The third is regulation of who has access to AI, covering both the hardware and the software. On the hardware side, chip export restrictions determine which countries and companies can obtain the advanced processors needed to build large AI systems, most visibly in US restrictions on exporting chips to China. On the software side, governments can restrict which specific AI models or services can be used, such as banning a foreign AI model on official devices or networks, or requiring that a country's data can only be processed within certain locations.
These three areas largely operate independently of one another. A country can tighten one without touching the others at all, a chip export restriction doesn't affect data centre water use, and a local planning decision doesn't change how capable a model is allowed to be. A headline mentioning AI being limited somewhere could be referring to any one of these three conversations, and the answer changes what the story is actually about.