DEFINITIONS
WHAT DOES IT MEAN FOR AN AI TO BE “SAFE”?
It is Sunday, the tenth day of this series, and the end of the week for most people. Which is why, today, instead of writing about the issues surrounding AI and the risks and dangers involved, I thought about writing instead, on the main question I’ve been writing around.
It is the question that every research paper, every alignment project, every legislation, and every government policy is ultimately geared at answering.
What does it actually mean for AI to be safe?
One speaker on AI once noted that in trying to make AI safer, we must be careful not to constrain it. That is true, and it points to the core difficulty: safety sounds like a simple goal until you try to define it. Because to define is to limit.
Thus, before you create a set on “AI Safety”, you must understand that it isn’t just one set, but a collection of intersecting sets. (Forgive me, I should draw up a Venn diagram here, but you will have to make do with your imagination.)
Considering that, below are the three main definitions of AI safety that are commonly considered in this context:
1. That AI Does Not Harm the User
The most common definition of AI safety is the one that affects most of us directly. A safe AI is one that does not harm the person using it. That through its actions or outputs, it does not bar people from getting opportunities due to bias and discrimination, does not give a depressed teenager suggestions for self-harm, and does not manipulate a lonely adult into an emotional dependency that separates them from human attachment.
This is the definition that guides most of the governmental legislation I have written about over the past week: the GUARD Act in America, China’s Interim Measures, Article 14 of the EU AI Act. All of them, in one way or another, are working from a consumer protection framework. They are asking: Is this product safe for the person using it?
This definition is necessary. But on its own, it is insufficient, because it only looks in one direction.
2. That AI Is Not Used as a Tool for Harm
The second definition seems similar to the first but is almost the mirror image of it. Here, safety means that the AI cannot be weaponised by one person to cause harm to others. Can someone use this system to generate targeted propaganda? To produce synthetic videos designed to destroy a person’s reputation? To automate manipulation at a scale no individual human could achieve alone?
Get Shimnom_Will’s stories in your inbox
Join Medium for free to get updates from this writer.
Earlier on, I read about sonic weapons, for instance, which already exist as instruments designed to induce illness in human beings through the use of sounds, with some variants documented to cause severe physical injury in animals. Now imagine AI being paired with such weapons. In science fiction, we have long imagined AI being deployed for mass violence precisely because it lacks the human inhibitions that slow such decisions down such as concience, empathy and indecision. AI safety, in this second sense, means that serious effort is made to prevent those fictional scenarios from becoming our realities.
This definition is to apply or put into perspective than the first, because almost any capability can be turned into a weapon. A system that writes persuasively can write disinformation. A system that processes information at scale can process the wrong information with total confidence. Thus, in defining AI safety this way, we must be careful not to blur the line between regulation and constraint.
3. That AI Is Safe for the World
This is the definition that receives the least attention in mainstream coverage, partly because it does not produce dramatic headlines, and partly because the harms it describes are diffuse, slow-moving, and shared by everyone, which makes them easy to ignore.
Even if an AI system never harms an individual user, and even if it is never weaponised for targeted violence, its existence at scale still costs the planet something.
The carbon footprint of AI systems alone could amount to between 32.6 and 79.7 million tons of CO₂ emissions in 2025, while the water footprint of those same systems could reach as high as 764.6 billion litres. To put that in perspective, research suggests that AI systems may have a carbon footprint equivalent to that of New York City in 2025, while their water consumption could approach the scale of the world’s total annual consumption of bottled water.
This problem is hardly discussed, even in AI research circles, with the exception in circles with environmental advocates.
. AI data centres consume enormous volumes of water primarily for on-site cooling, water used by the power plants which supply electricity to the data centres, and water consumed during the manufacturing of the hard drives for the AI systems. A study by the Houston Advanced Research Centre found that data centres in Texas alone will use 49 billion gallons of water in 2025, and as much as 399 billion gallons by 2030, an amount equivalent to drawing down the largest reservoir in the United States by more than 16 feet in a single year.
This definition has been adopted by environmental and human rights activists. Between March and June 2025, community opposition led to $98 billion in data centre projects being blocked or delayed, and at least 25 projects were cancelled in 2025 in response to local objections from the activists.
Butis the hardest one to apply, or even consider. Because it raises questions such as: do we give up advancement for the sake of present preservation? Do we praise the communities that banned data centres from being built in their community due to rising electricity costs and water consumption?
Well, the answer to that will have to come from environmental protection leaders and technological advancements leaders who must sit together to look for more effective ways to balance technology and humanity.
In the Arts and Humanities, before giving a definition, we say, “there is no generally acceptable definition.” This is because people view things differently due to the differences in environment and situations.
Similarly, the definition of AI Safety is not generally acceptable or applicable. It is a question that must be viewed from different perspectives, with each perspective coming together to form a collective whole. Something like Solitaire pieces.
A policy that answers the first question without touching the third has not solved the problem. It has only addressed the part of it that is easiest to see.
See you tomorrow.
Hi! I am Will Simon, and this is day 10 of writing on AI Safety. Thank you so much for reading once again. I hope you had a lovely weekend!
