On Calling Everything AI
We keep collapsing very different technologies into one all-encompassing phrase
Why are we calling everything AI? In 1959, a group of MIT students broke into a room after hours to programme a computer. They were members of the Tech Model Railroad Club — an organisation devoted, as the name suggests, to model railways — and the computer they were sneaking time on was an IBM 704 that filled most of the room and cost several million dollars. What they were doing was unauthorised, and by any reasonable measure brilliant.
They called what they did hacking. The word meant something precise: finding a clever, unorthodox solution to a technical problem; making a system do something its designers had not anticipated; understanding something deeply enough to improve it. The hacker ethic that grew from those late-night sessions at MIT, and that Steven Levy documented in Hackers: Heroes of the Computer Revolution, was built around openness, curiosity, and the belief that information and knowledge should be freely shared. It was, Levy wrote, “a new way of life, with a philosophy, an ethic and a dream.”
Then something happened. A small number of genuinely malicious actors — phone phreakers, system intruders, people the hacker community itself called crackers to distinguish them from the real thing — became the public face of hacking in mainstream media coverage. By the mid-1980s the word had inverted. The press had turned a term of honour into a term of menace. By the time most of the world encountered the word “hacker,” it meant criminal. The community tried to reclaim it, reserving “cracker” for the bad actors and “hacker” for the creative problem-solvers, and largely failed. The battle was lost in public discourse before most people even knew it had been fought.
We are doing the same thing to artificial intelligence right now. The stakes are considerably higher.
What “AI” Actually Means
When a politician says they are worried about AI, they might mean any of the following genuinely different things:
Large language models – the technology behind ChatGPT, Claude, Gemini, Perplexity and similar tools which generate text, code, analysis, and creative content by predicting what words are most likely to follow other words, trained on enormous datasets.
Autonomous weapons systems – Military applications that make targeting decisions with varying degrees of human oversight, with specific ethical and legal implications entirely distinct from text generation.
Algorithmic hiring and credit scoring – Systems that have been in use for decades, that make consequential decisions about people’s lives, and that have well-documented problems with bias and opacity which predate the current AI conversation by years. In the UK and EU, Experian’s credit scoring models help determine whether you get a mortgage or a loan. HireVue’s video interview software analyses facial expressions, tone of voice, and word choice to score job candidates before a human ever sees them. It has been used by Unilever, Goldman Sachs, and many others. COMPAS, used in the United States to predict reoffending risk and influence bail and sentencing decisions, was found by a ProPublica investigation in 2016 to be significantly biased against Black defendants. These are not speculative futures. They are systems already in operation, already affecting real people’s lives, already the subject of serious documented harms.
Facial recognition technology – Surveillance infrastructure with specific civil liberties implications and specific failure modes, particularly for people with darker skin tones, that has been deployed by police forces and governments with minimal public scrutiny.
Deepfakes and synthetic media – Generative tools that produce convincing fake images and video, with specific implications for disinformation and non-consensual intimate imagery.
Recommendation algorithms – The systems that decide what you see on social media, what appears on your news feed, what YouTube serves next. These have shaped political beliefs and emotional states for a decade, largely without serious public debate, and are rarely mentioned in the AI conversation despite being among the most consequential deployed systems in existence.
Robotic process automation – Software that automates routine administrative tasks. It is genuinely useful, genuinely job-displacing in specific sectors, and almost entirely absent from the AI panic despite being already widespread.
These are not the same technology. They do not have the same architecture, the same failure modes, the same ethical implications, or the same regulatory requirements. Collapsing all of them into “AI” and then conducting a public debate about “AI” as though it were a single coherent thing, produces exactly what you would expect: maximum noise, minimum understanding, policy responses calibrated to the headlines rather than the substance, and a conversation that generates far more heat than light.
This is not an accident. It is what happens when a word becomes unmoored from the thing it is supposed to describe. And it is, precisely, what happened to “hacker.”
The Pattern of Semantic Collapse
The mechanism is the same in both cases. A new technology, or a new community of practice around a technology, develops its own precise vocabulary. The people inside the community know exactly what the words mean. The words carry real information, real distinctions, real ethical weight.
Then the technology becomes sufficiently visible, or sufficiently threatening to established interests, to attract mainstream attention. Mainstream coverage requires simplification. Simplification requires a single word. That word gets attached to the most alarming version of the phenomenon, the crackers, not the hackers; the autonomous weapons, not the language models, because alarm drives engagement and engagement drives coverage. Within a few years the word means the alarming thing in public discourse, and the original, more precise meaning is either lost or confined to specialists who increasingly sound defensive when they try to explain it.
The people who understand the technology correctly are then in an impossible position. They know the word is wrong. They know the public debate is being conducted against a caricature. But every attempt to correct the caricature sounds like a defence of the alarming version, because the word no longer carries the distinction. When a security researcher says “I’m a hacker” and means it as a compliment, they spend the next five minutes explaining what they don’t mean rather than what they do. When an AI researcher says “large language models don’t actually understand anything in the way humans do,” they are immediately accused of minimising risks they are actually trying to explain.
The word does the work the technology’s critics want done before anyone has examined whether the criticism is accurate.
Why it’s important
This is not a semantic argument for its own sake. The imprecision has real consequences.
When “AI” means everything from a text autocomplete function to a military targeting system, the regulations designed to govern it will either be so broad they are unworkable or so narrow they miss everything important. The EU AI Act — whatever its specific merits — was negotiated in an environment where policymakers were simultaneously worried about Terminator-style autonomous systems and about the risk of a chatbot giving bad medical advice. These are not the same problem and they do not have the same solution.
Legitimate concerns get buried under speculative ones. The documented bias in algorithmic hiring systems, a real, present, measurable harm affecting people’s employment prospects right now, gets less serious attention than scenarios about superintelligence, because superintelligence is more cinematically alarming, and cinematic alarm drives coverage.
Meanwhile, genuinely useful deployments get swept into the same conversation as surveillance infrastructure and deepfakes. The large language model helping a university restructure its curriculum, the diagnostic tool that detects early-stage cancer in a scan a radiologist might have missed, the accessibility tool that gives a blind person independent access to written information: all of these become harder to defend without appearing to defend everything.
More Precise Language
The hacker community’s attempt to distinguish hackers from crackers failed in public discourse but succeeded internally. The distinction is maintained and understood within the communities that need it. Something similar is already developing around AI, and it is worth naming it clearly.
Large Language Models (LLMs) are the text-generating systems. This is the technology people are mostly talking about when they discuss ChatGPT, Claude, or Gemini. The main concerns here involve accuracy, intellectual property, labour displacement in creative fields, and the risk of systems that sound confident when they are wrong.
Generative AI is a broader category including LLMs but also image generators, video synthesis tools, voice cloning, and other systems that produce new content rather than classifying or analysing existing content. The main concerns here include deepfakes, non-consensual imagery, and copyright.
Predictive AI / algorithmic decision systems are the systems that make or influence decisions about people: hiring, credit, bail, content moderation, medical diagnosis. Many of these predate the current AI wave by years. The main concerns here involve bias, opacity, accountability, and due process.
Autonomous systems are systems that take actions in the physical world with varying degrees of human oversight: self-driving vehicles, military applications, logistics robots. The main concerns here involve liability, safety, and the limits of machine autonomy in high-stakes situations.
These categories are not perfect and they are not fixed. But they are more useful than “AI,” in the same way that “cracker” and “hacker” are more useful than a single word that means both — if only the distinction could be made to stick.
A Personal Note
I should be transparent about something. This essay was written in collaboration with Claude, a large language model developed by Anthropic. My book, Well… How Did We Get Here?, was also co-written using Claude and Perplexity AI, and I disclosed that fact because I think disclosure matters.
I mention this not to defend the technology but to identify it accurately. Claude and Perplexity are not autonomous. They do not have goals. They generates the most statistically likely continuation of the text they have been given, shaped by training and by the conversation itself. They are tools, remarkable ones, but still tools. And like all tools, what matters most is who uses it, for what purpose, with what oversight, and under what constraints.
The hacker ethic, at its core, was a belief that the understanding of tools should be democratised, that knowledge of how systems work should be widely shared rather than hoarded by specialists or corporations. The MIT students breaking into the EAM room were not criminals. They were curious. They believed the machine should be understood by anyone who wanted to understand it, not protected behind an institutional veil.
That instinct seems right to me still. The question of what AI actually is, what these specific technologies do, what they can and cannot do, what risks they introduce and for whom, is too important to be left to the people who want to maximise alarm and the people who want to maximise adoption. Most of us live somewhere in the middle of that argument, trying to think clearly about tools we increasingly depend on but rarely fully understand.
Clearing up the language is the first serious act of thinking.