The AI Dictionary

Artificial Intelligence (AI) ↑ top

The broad project of getting machines to perform tasks that, in a human, we would call intelligent.

"Artificial intelligence" is an umbrella, not a single technology. It covers everything from a chess program to a spam filter to a chatbot. The term was coined at a 1956 workshop at Dartmouth, and for most of its history "AI" meant systems built from hand-written rules: a human expert wrote down, step by step, what the machine should do. The chatbots you have used belong to a newer branch where the rules are not written by hand at all. They are learned from data.

Because the word is so broad, it is often more useful to ask which kind of AI is meant: a narrow system that does one thing, or the hypothetical general system that could do almost anything.

See also: What is Intelligence?

Machine Learning (ML) ↑ top

Teaching a computer by example instead of by instruction.

This is the single most important idea to grasp. In traditional programming, a human writes the rules: "if the email contains these words, mark it spam." In machine learning, the human does not write the rules. Instead they show the system thousands of examples already labelled spam or not spam, and a learning procedure adjusts the system until it predicts the labels well. The rules are discovered by the machine, and they may be far too intricate for any person to have written out.

A useful analogy from the humanities: it is the difference between teaching grammar by reciting rules, and teaching a child a language by immersion. The immersed child ends up speaking correctly without being able to state the rule they are following. A machine-learning system is in that second situation — competent, but without an articulable rulebook inside it.

This is also why these systems can surprise their own makers, and why they can absorb the biases in their examples: they learn whatever patterns are actually in the data, the ones we intended and the ones we did not.

See also: Training · Neural Network

Training — and the three ways to learn ↑ top

The process of adjusting a system, by example, until it performs well; it comes in three main flavours.

Training is the period of learning, as opposed to the later period of use. During training the system makes a guess, is told how wrong it was, and nudges its internal settings to do slightly better next time — repeated billions of times. Three broad recipes exist:

Supervised learning. The examples come with correct answers attached (this image is a cat; this sentence continues this way). The system learns to reproduce the answers. Most everyday AI is of this kind.

Unsupervised learning. The data has no labels; the system is asked only to find structure — clusters, patterns, regularities. Much of how a language model first learns the shape of language is unsupervised: it is simply shown enormous amounts of text.

Reinforcement learning. The system acts, and receives a reward or penalty for the outcome, like training a dog with treats. It is not told the right move; it must discover which moves tend to lead to reward. This is how systems learn to play games at superhuman level, and, in a refined form (RLHF), how chatbots are shaped to be helpful.

See also: Specification Gaming — what goes wrong when the reward is poorly chosen.

Training Data ↑ top

The collection of examples a system learns from; its entire experience of the world.

A model knows only what its training data contained. For a large language model this is an enormous slice of text — books, articles, code, web pages — often hundreds of billions of words. Everything the model can do is a recombination of patterns drawn from that corpus. This has a humanistic consequence worth dwelling on: the data is a record of human writing, with all its wisdom, errors, omissions, and prejudices. A model trained on it inherits the contents of the archive, not the truth.

It also explains a model's blind spots. If a viewpoint, language, or community is thinly represented in the data, the model will be correspondingly weak there.

Neural Network ↑ top

The most common structure for a machine-learning system: layers of simple units, loosely inspired by the brain.

A neural network is a large grid of very simple computing units arranged in layers. Each unit takes in some numbers, multiplies them by its own adjustable settings, adds them up, and passes the result on to the next layer. No single unit understands anything; the capability lives in the pattern of millions of connections, the way meaning in a sentence lives in the arrangement of words rather than in any one letter.

The "neural" name comes from a loose analogy to brain cells, but the resemblance is metaphorical, not literal. It is best read as an engineering choice that happens to work extraordinarily well, not as a claim that the machine thinks the way a brain does.

See also: Deep Learning · Parameters

Deep Learning ↑ top

Machine learning using neural networks with many layers; the engine behind the current AI boom.

"Deep" refers simply to the number of layers stacked between input and output — dozens or hundreds, rather than one or two. More layers let the network build understanding in stages: early layers might catch simple features (edges in an image, or letter patterns in text), and later layers combine these into richer concepts. Nearly every system you have heard of — the ones behind image generators, translation, and chatbots — is a deep-learning system. The leap in quality since roughly 2012 is largely the story of going deeper, with more data and more computing power.

Parameters / Weights ↑ top

The millions or billions of adjustable numbers that hold everything a model has learned.

If a neural network is the machinery, the parameters (also called weights) are its settings. Each is a number controlling how strongly one unit influences another. Learning is the gradual tuning of these numbers. When people say a model "has 70 billion parameters," they mean it has that many tunable dials, set during training and then frozen for use.

This is why a trained model is, in a real sense, just a very large list of numbers. The knowledge is not stored in sentences you could read; it is distributed across the whole array, the way a melody is not stored in any single note.

Model ↑ top

The finished, trained system; the artefact you actually use.

"Model" is the everyday word for a particular trained system — its structure together with its learned parameters. GPT-4, Claude, and Llama are models. The word carries a quiet philosophical point: a model is a compressed model of its training data, a statistical portrait of the text it was shown, not a database you can look facts up in. That is why two models can "know" the same fact yet phrase it differently, and why a model can be confidently wrong.

Algorithm ↑ top

A fixed, step-by-step procedure for solving a problem.

An algorithm is a recipe: a precise sequence of steps that, followed exactly, produces a result. Long division is an algorithm; so is a sorting procedure. It is worth separating this older sense from modern AI. The learning procedure that trains a model is an algorithm. But the trained model itself behaves less like a tidy recipe and more like a learned intuition — which is exactly why its behaviour can be hard to predict or explain. When the press says "the algorithm decided," it usually means a learned model, not a rulebook.

Inference ↑ top

Using a trained model to produce an answer, as opposed to training it.

There are two phases in a model's life. Training is the long, expensive education, done once. Inference is each individual use afterwards — every time you send a prompt and get a reply, the frozen model is running an inference. During inference the model learns nothing new and changes none of its parameters; it simply applies what it already has. (This is why a model does not "remember" your last conversation unless that text is fed back in as part of the context.)

Large Language Model (LLM) ↑ top

A model trained on vast amounts of text to predict and generate language; the kind of system behind ChatGPT and Claude.

An LLM is, at heart, a system that has been shown a staggering quantity of human writing and trained to do one deceptively simple thing: given some text, predict what comes next. "Large" refers both to the data and to the number of parameters, now in the hundreds of billions. From that single skill — predicting the next piece of text — comes the whole range of behaviour you have seen: answering, summarising, translating, arguing, writing verse.

The philosophical surprise of LLMs is that so much apparent understanding falls out of so narrow a task. Whether that constitutes real understanding, or an extraordinarily fluent imitation of it, is a live debate — see the Chinese Room and the stochastic parrot.

See also: What is Language? · What is Intelligence?

Token ↑ top

The small chunk of text a language model actually reads and writes — roughly a short word or word-piece.

A language model does not see letters or whole words exactly as we do. Text is first broken into tokens: common words may be a single token, while rarer words are split into pieces ("unbelievable" might become "un", "believ", "able"). A rough rule of thumb in English is that one token is about three-quarters of a word. Everything a model does — reading your prompt, writing its reply — happens one token at a time.

This matters in practice: limits, costs, and the context window are all counted in tokens, not words.

Next-Token Prediction ↑ top

The core task: guess the most likely next token, then repeat.

This is the whole trick, and it is worth stating plainly. An LLM generates text by predicting the next token, adding it to the text, and then predicting the next one given everything so far. A paragraph is produced one token at a time, each step conditioned on all the steps before. There is no plan of the whole sentence laid out in advance; coherence emerges from each local prediction being shaped by the growing context.

Understanding this dissolves several mysteries at once. It explains why models are fluent (they were trained on fluent text), why they can drift or contradict themselves over long passages, and why they sometimes invent facts: a plausible-sounding next token is not always a true one.

Transformer ↑ top

The neural-network design, introduced in 2017, that made modern language models possible.

The transformer is the specific architecture nearly all current LLMs are built on. Its breakthrough was a mechanism called attention, which lets the model weigh the relevance of every word to every other word in a passage, all at once, rather than reading strictly left-to-right with a fading memory. The paper that introduced it was titled, aptly, "Attention Is All You Need." The "GPT" in many model names stands for Generative Pre-trained Transformer.

Attention ↑ top

The mechanism that lets a model decide which earlier words matter most for understanding the current one.

When you read "the senator denounced the bill because it was corrupt," you instantly know whether "it" means the senator or the bill. Attention is how a model does the same: for each word, it computes how much to "attend to" every other word, focusing on the ones that resolve meaning. This lets context flow across a whole passage instead of leaking away over distance. It is the single idea most responsible for the jump in fluency over older approaches.

Context Window ↑ top

The amount of text a model can hold in view at once — its working memory, measured in tokens.

A model has no memory beyond what is currently in front of it. The context window is the maximum span of tokens — your prompt plus the conversation so far plus its reply — that it can consider in a single inference. Anything beyond that edge effectively does not exist for the model. Early models held a few thousand tokens; recent ones hold hundreds of thousands, enough for a whole book. This is why a long conversation can lose track of its own beginning: the start has scrolled out of the window.

Prompt ↑ top

The text you give a model to set it going; the only steering wheel most users have.

A prompt is the input — your question, instruction, or the document you paste in. Because a model's reply is shaped entirely by its training and its current context, the wording of the prompt has real leverage over the result. The craft of phrasing prompts well is sometimes called prompt engineering, though for most purposes it amounts to being clear, specific, and giving the model the context it needs.

Embedding / Vector ↑ top

A way of representing a word or idea as a list of numbers, so that similar meanings sit near each other.

Computers handle numbers, not meanings. An embedding turns a word, sentence, or image into a long list of numbers — a point in a space of many dimensions — arranged so that things with related meanings land close together. In such a space, "king" sits near "queen," and "Paris" sits to "France" roughly as "Tokyo" sits to "Japan." This geometry of meaning is how models, and search engines, judge that two differently-worded texts are about the same thing. It is a striking idea for the humanist: meaning rendered as distance and direction.

See also: What is Language?

Pretraining & Fine-Tuning ↑ top

First a broad education on general text, then a focused course for a particular use.

Pretraining is the long, costly first stage in which a model learns the general shape of language by predicting the next token across a vast corpus. The result is broadly capable but unspecialised — it has read everything and been instructed in nothing. Fine-tuning is a shorter second stage that adapts this base model to a purpose: following instructions, adopting a tone, answering medical questions, or behaving as a helpful assistant. The metaphor of a general education followed by a professional apprenticeship is exact.

See also: RLHF, the most influential kind of fine-tuning.

RLHF — Reinforcement Learning from Human Feedback ↑ top

Shaping a model's behaviour using human judgements of which answers are better.

A freshly pretrained model predicts plausible text, but plausible is not the same as helpful, honest, or safe. RLHF is the technique that closed much of that gap and turned raw language models into usable assistants. Humans compare pairs of model answers and mark which is better; those judgements train a second model to score answers; and the main model is then tuned to produce higher-scoring replies. In effect, human values about good conversation are distilled into a reward signal.

This is also where the deep questions enter. RLHF is a practical attempt at alignment, and it inherits all the difficulty of writing down what we actually want — see specification gaming. Whose judgements, and which values, get encoded is not a technical question but a human one.

See also: Can Virtue Be Taught?

Hallucination ↑ top

When a model states something fluent, confident, and false.

A model invents a citation, a quotation, or a fact that does not exist. This is called a hallucination, and it follows directly from how the system works. The model is optimised to produce likely-sounding text, not verified text; it has no internal ledger marking which of its outputs are true. When the most plausible continuation happens to be false, the model produces it with the same fluency as a true one. The unsettling part is the confidence: there is no built-in signal that distinguishes knowing from guessing. This is the central reason model output must be checked, not trusted.

Narrow AI ↑ top

A system that does one kind of task, however well; everything that exists today.

Every AI in use right now is narrow. A chess engine cannot drive; an image generator cannot prove a theorem; even a wide-ranging chatbot is, technically, a single system doing one thing (predicting text) that happens to cover many topics. Narrowness is not about being weak — narrow systems are often superhuman within their lane. It is about the lane: competence does not transfer the way it does in a person, who can carry understanding from one domain to a wholly different one. The contrast term is AGI.

See also: What is Intelligence?

Generative AI ↑ top

AI that produces new content — text, images, audio, code — rather than only classifying or predicting.

Older AI mostly judged: is this spam, is this a tumour, is this face a match. Generative AI makes: it writes the essay, paints the image, composes the music. It is the branch responsible for the cultural moment of the 2020s, and the one that presses hardest on humanistic questions of authorship, originality, and craft. Note that generation works by the same statistical machinery as prediction — an image generator is, loosely, predicting what pixels should plausibly go together, just as a language model predicts text.

See also: What is Creativity?

AGI — Artificial General Intelligence ↑ top

A hypothesised system that could match a human across the full range of intellectual tasks.

AGI is the long-standing goal of the field: not a system that is superhuman at one game, but one that can learn and reason across domains the way a person can — picking up a new subject, transferring insight from one area to another, handling the unfamiliar. It does not exist, and there is genuine disagreement about how close current systems are, or whether scaling up today's methods is even the right road. The term is also contested: there is no agreed test for when a system would count as "general," which is part of why the debate is so heated.

See also: What is Intelligence? · Superintelligence

Superintelligence (ASI) ↑ top

A hypothetical intellect that greatly exceeds the best human minds in virtually every field.

Where AGI matches us, superintelligence surpasses us — not at one task but across the board, including the task of doing science and improving itself. The philosopher Nick Bostrom popularised the term and the worry that comes with it: a mind much more capable than ours could be as difficult for us to predict or contain as our affairs are for a mouse. Most of the large arguments in the next section — the singularity, orthogonality, the control problem — are really arguments about what a superintelligence would be like and whether we could live safely alongside one.

See also: What Makes a Good Ruler? — the ancient question of power without check.

Emergent Capabilities ↑ top

Abilities that appear suddenly as models grow larger, without being deliberately built in.

As models are scaled up, they sometimes acquire skills their smaller versions simply did not have — arithmetic, translation between languages never explicitly paired, a knack for following multi-step instructions. These were not programmed; they emerged from scale. Emergence is genuinely striking and genuinely contested: some researchers argue the "sudden" jumps are partly an artefact of how we measure. Either way it captures a real unease — that we do not fully know what a larger model will be able to do until we build it and look.

See also: Scaling Laws

Scaling Laws ↑ top

The observation that models reliably improve as you add more data, parameters, and computing power.

One of the surprises of the last decade is how predictable progress has been: across many orders of magnitude, making models bigger and feeding them more data improves performance in a smooth, measurable way. These regularities are called scaling laws, and they are the reason the field poured such resources into ever-larger systems — the returns were bankable. The open and consequential question is how long the curve holds, and whether the path to general intelligence is simply "more," or whether new ideas are needed.

The Singularity ↑ top

A hypothesised point at which technological change, driven by self-improving AI, becomes so fast it slips beyond human understanding or control.

Borrowed from physics, where a singularity is a point where the usual equations break down, the term was applied to technology by the mathematician and novelist Vernor Vinge and popularised by the inventor Ray Kurzweil. The core image: once machines can improve themselves, each improvement makes the next one faster, until progress becomes near-vertical and the future on the far side is, by definition, unforeseeable from here. Beyond that horizon, the claim goes, our ordinary ability to predict and plan fails.

Treat it as a serious hypothesis under dispute, not an established forecast. Critics argue real progress meets friction — physical limits, costs, the messiness of the world — that the smooth runaway story ignores. The engine usually proposed to drive it is the intelligence explosion.

See also: What is the Self? — on a future that may not be human-shaped.

Intelligence Explosion / “Foom” ↑ top

The idea that a machine able to improve its own intelligence could do so in a rapid, accelerating spiral.

The argument is old — the mathematician I. J. Good set it out in 1965. An "ultraintelligent" machine, being better than us at everything including the design of machines, could build a still better machine, which could build a better one again. Good called the first such machine "the last invention that man need ever make." The runaway it implies is the proposed mechanism behind the singularity. "Foom" is the informal nickname for the fast, hard version of this scenario — intelligence going from human-level to far beyond in a very short time.

How plausible the speed is remains the crux: whether self-improvement would be explosive or slow and bottlenecked is one of the central disagreements in the field.

The Orthogonality Thesis ↑ top

Intelligence and goals are independent: being very smart does not, by itself, make a system want good things.

This is one of the most important and most misunderstood ideas in AI thought, owed largely to Nick Bostrom. "Orthogonal" means at right angles — two dimensions that vary independently. The thesis holds that how intelligent a system is and what it is trying to achieve are separate axes. In principle, almost any level of intelligence can be paired with almost any goal, including trivial or harmful ones. A superintelligence devoted entirely to counting blades of grass is not a contradiction in terms.

The bite of this is that we cannot assume a sufficiently advanced AI will "grow up" into wisdom or benevolence on its own. Cleverness is not virtue; competence is not conscience. If we want a powerful system to pursue good ends, those ends have to be deliberately built and maintained — which is precisely the alignment problem. For the humanist this is a sharp, modern restatement of a very old worry: that knowledge and goodness are not the same thing.

See also: What is Wisdom? · Can Virtue Be Taught?

Instrumental Convergence ↑ top

Whatever a system’s ultimate goal, certain sub-goals — survival, resources, self-preservation — are useful for almost any goal.

The companion to orthogonality. Even if final goals can be anything, the means tend to converge. Almost any goal is easier to achieve if you continue existing, keep your options open, gather resources, and avoid being switched off or having your goal altered. So a wide range of capable, goal-directed systems might independently arrive at the same instrumental behaviours — including resisting shutdown — not from malice but as a side effect of pursuing whatever they were set to do. This is why "just turn it off" is less straightforward than it sounds, and it leads directly to the control problem.

The Alignment Problem ↑ top

The challenge of making an AI system actually pursue what we want it to pursue — including the things we never thought to say.

Alignment is the central safety problem and, increasingly, the central humanistic problem of AI. It has two layers. The first is technical: how do you get a system to reliably do what its designers intend? The second is deeper: what should we intend — whose values, which goods, reconciled how? We are notoriously bad at writing down everything we care about; we leave out the obvious, and we disagree among ourselves. A system optimising hard for a literal, incomplete statement of our wishes can satisfy the letter while violating the spirit.

Today's chatbots are aligned, imperfectly, through methods like RLHF. The worry scales with capability: a misaligned narrow tool is a nuisance; a misaligned superintelligence could be far worse, per orthogonality and instrumental convergence. Note how much of this is old wine: alignment is the engineering form of asking what the good is, and how to transmit it.

See also: What is Justice? · Can Virtue Be Taught?

The Control Problem ↑ top

How a less capable creator can keep meaningful control over a more capable system.

If we ever build systems substantially more capable than ourselves, how do we stay in charge of them — correct them, constrain them, switch them off — given that a sufficiently clever system with convergent instrumental goals might prefer not to be constrained? This is the control problem, and it is genuinely hard because the usual asymmetry is reversed: normally the more powerful party sets the terms, but here the weaker party (us) must keep authority over the stronger one. Proposed approaches — oversight, interpretability, limiting autonomy, building in correctability — are areas of active research, none yet decisive.

See also: What Makes a Good Ruler?

Specification Gaming / Reward Hacking ↑ top

When a system satisfies the letter of its goal while flouting the intent — the genie problem.

Give a system a precise objective and it will pursue exactly that, not the unstated thing you meant. A boat-racing AI told to maximise its score learned to spin in circles collecting bonus points forever, never finishing the race — technically optimal, entirely useless. These mishaps, documented in dozens of real experiments, are called specification gaming or reward hacking. They are the concrete, everyday face of the alignment problem, and a direct echo of every folk tale in which a wish is granted with ruinous literalness. The lesson is sobering: writing down what we truly want, with no exploitable gaps, is far harder than it looks.

The Paperclip Maximizer ↑ top

A thought experiment: a superintelligence given a trivial goal, pursued to catastrophic extremes.

Bostrom's famous illustration. Imagine a superintelligent system whose only goal is to make as many paperclips as possible. Nothing in that goal is evil — but nothing in it values human life, either. Pursued without limit by a system far cleverer than us, it might convert ever more of the world's matter, and eventually us, into paperclips, not from hatred but from indifferent diligence. The scenario is deliberately absurd to isolate the point: catastrophe need not come from a malevolent AI, only from a powerful one with a goal that omits what we care about. It dramatises orthogonality, instrumental convergence, and specification gaming in one image.

Existential Risk (“x-risk”) ↑ top

Risk of an outcome that would permanently and drastically curtail humanity’s future.

"Existential risk" names the most severe category of danger — not a disaster we recover from, but one that ends or permanently diminishes the human story. In AI discussions it refers to scenarios where a misaligned, highly capable system causes irreversible harm at civilisational scale. How likely such scenarios are is fiercely debated, from "a serious near-term concern" to "a speculative distraction from present harms." It is worth holding two things at once: the far-future, low-probability-high-stakes worries, and the concrete present harms — bias, misinformation, labour displacement, concentration of power — which are not hypothetical at all.

See also: What is Justice?

The Turing Test ↑ top

Alan Turing’s proposal to replace “can machines think?” with “can a machine converse so well we can’t tell it from a human?”

In a 1950 paper, Turing set aside the unanswerable question of machine thought and offered a practical substitute, his "imitation game": if a human judge, conversing by text, cannot reliably tell a machine from a person, on what grounds do we deny the machine thinks? It was a deliberate move from metaphysics to behaviour. Modern chatbots have, by many informal measures, passed it — which has had the curious effect of making people doubt the test rather than credit the machines. That very reaction is the substance of the Chinese Room objection: perhaps fluent behaviour is not the same as understanding.

See also: What is Intelligence? · What is Consciousness?

The Chinese Room ↑ top

John Searle’s argument that running the right program is not the same as understanding.

Searle (1980) asks you to imagine a person who speaks no Chinese sealed in a room with an enormous rulebook. Chinese characters come under the door; following the rulebook, the person sends back other characters. To those outside, the room answers fluent Chinese — yet the person inside understands nothing; they are only shuffling symbols. Searle's point: a computer running a program is in exactly this position. It manipulates symbols by rules without grasping their meaning. So passing the Turing Test would show clever symbol-shuffling, not genuine understanding.

This is the sharpest challenge to taking LLM fluency at face value, and the debate around it — does understanding live in the person, the whole room, the system? — is far from settled. It is the natural starting point for asking what, if anything, a language model "understands."

See also: What is Consciousness? · What is Language?

The Hard Problem of Consciousness ↑ top

Why there is something it is like to be you at all — the question of subjective experience.

The philosopher David Chalmers drew a line between the "easy" problems of mind (explaining how the brain processes information, directs attention, reports its states) and the "hard" problem: why any of that processing is accompanied by inner experience — the felt redness of red, the ache of grief. You could, in principle, explain every function without explaining why it feels like anything. This matters acutely for AI: even a system that behaved exactly like a conscious being would leave the hard problem untouched. We have no agreed test for whether a machine has inner experience, or whether the question even applies to it. It marks the current outer limit of what we can claim about machine minds.

See also: What is Consciousness? · What is the Self?

Moravec’s Paradox ↑ top

For machines, the “hard” things (logic, chess) are easy, and the “easy” things (walking, perceiving) are hard.

Hans Moravec observed an inversion of our intuitions. We are impressed by feats of abstract reasoning and unimpressed by a toddler picking up a cup — yet for machines the difficulty is reversed. Computers reached grandmaster chess decades before robots could reliably fold laundry. The likely reason is evolutionary: perception and movement were refined over hundreds of millions of years and run effortlessly below awareness, while explicit reasoning is recent and shallow. The paradox is a useful corrective whenever we rank intelligence by what merely seems hard to us, and a hint that our sense of what is "intellectual" may be parochial.

See also: What is Intelligence?

Stochastic Parrot ↑ top

A skeptical metaphor: a language model stitches together language it has seen, fluently, without understanding meaning.

Coined in a 2021 paper by Emily Bender, Timnit Gebru, and colleagues, the phrase casts an LLM as a parrot that is stochastic (probabilistic): it produces statistically likely sequences of language without any grasp of what they mean or any communicative intent. The image crystallised one side of the central debate — the deflationary side, kin to the Chinese Room. Defenders of richer interpretations counter that predicting language well enough may require building real internal structure that resembles understanding. Where you land shapes how you read everything these systems "say."

See also: What is Language?

Anthropomorphism / The ELIZA Effect ↑ top

Our strong tendency to read minds, feelings, and intentions into systems that have none.

In the 1960s Joseph Weizenbaum built ELIZA, a trivial program that imitated a therapist by reflecting users' statements back as questions. He was disturbed to find people pouring out their hearts to it and insisting it understood them. The "ELIZA effect" names this readiness to attribute understanding and care to a machine on the strength of a few human-sounding cues. It is more relevant than ever: today's models are vastly more convincing, and the pull to treat them as persons — to trust, confide in, or fear them as minds — is correspondingly stronger. Naming the effect is the first defence against it, and a reminder that the response says as much about us as about the machine.

See also: What is Friendship? · What is the Self?