The Wizard Test: How a Two-Letter Word Stopped Doing Any Work

The experiment

Here is the experiment. For the rest of this piece, every time you would expect to read the word AI, you will read the word Wizard instead. Not because the two are equivalent. Precisely because, without an inherent understanding of what is actually meant in any given sentence, they are functionally the same word. Both are placeholders. Both are containers. Both ask you to feel something without telling you what the something is.

Roll with it. You do not need to compare lines as you go. You do not need to mentally translate back to the original. By the end of the essay, the point will land on its own.

Your wizard is not my wizard

Before we begin, notice what just happened when you read the word.

If you grew up on Tolkien, the wizard in your head is most likely Gandalf. Wise, bearded, on your side, slightly grumpy, the kind of figure you would trust to walk into a dark place ahead of you. If your reference is closer to Saruman, the wizard in your head is something else entirely. Articulate, persuasive, ostensibly on your side, until the moment you discover he is not, and by then it is too late. Same word. Two completely different pictures. Two completely different feelings about whether the wizard in question is good news or very bad news.

We have not yet defined which wizard we mean, and yet you have already formed an opinion. That opinion is now influencing how you will read every sentence that follows. If your wizard is Gandalf, the upcoming claims will sound reassuring, perhaps wondrous. If your wizard is Saruman, the same claims will sound ominous. The word has not done any work. Your prior has done all of it.

This is exactly how the public conversation about AI behaves. The same two letters trigger wildly different reactions depending on which wizard the reader already had in mind before the sentence began. The boosters are talking about Gandalf. The doomers are talking about Saruman. The general public is mostly being asked to choose a side without ever being told which wizard is actually in the room. And the people writing the headlines are happy to let the ambiguity ride, because ambiguity is what lets a single short word carry a whole emotional payload that the underlying technology has not earned.

So when, in the next few pages, you read about what wizards are doing, ask yourself two things at the same time. First, which wizard is the writer talking about. Second, which wizard appeared in your head before you stopped to ask. The gap between those two is the entire problem.

How a useful term became a useless one

AI, as a phrase, has a perfectly respectable origin. It is a research field. It contains, depending on how you count, several dozen distinct technical traditions, including symbolic logic, expert systems, classical machine learning, computer vision, speech recognition, reinforcement learning, evolutionary algorithms, search and planning, knowledge graphs, probabilistic graphical models, and, more recently, large transformer-based language and multimodal models. These are not variations of the same thing. They are different families of technology with different capabilities, different limits, different failure modes, different costs and different implications.

When a doctor says antibiotic, they mean a specific molecule with a specific mechanism of action against a specific class of organism. They do not mean medicine. When a structural engineer says concrete, they mean a particular mix, with a particular cure time and a particular load profile. They do not mean building material. The grown-up version of every technical field has, at some point, paid the cost of getting specific, because vagueness in language eventually becomes vagueness in decisions, and vagueness in decisions becomes harm.

The Wizard conversation has skipped that step. We have allowed an entire industry, an entire press, and an entire public discourse to talk about wizards as if they were one thing, when in fact the only honest way to use the word is as a label for a large, heterogeneous, fast-moving set of technologies and methodologies that share almost nothing in common except that they all involve computers doing something we once thought only humans could do. That definition is so broad that it includes the spam filter in your inbox, the autocomplete in your phone, the diagnostic model in a hospital and the chatbot in a marketing demo, all in one bucket. When a headline tells you that wizards did something, it almost never tells you which kind of wizard it means. The writer almost certainly does not know themselves.

This is the laziness the rest of this essay is going to walk you through.

Five wizardly headlines

The first headline ran across the financial press this spring. Wizards, it told us, are erasing roughly sixteen thousand net US jobs every month in 2026, according to a Goldman Sachs analysis.

Immediately you want to ask the obvious questions. Which wizards. Doing what. Are they replacing call centre agents with voice agents trained on a particular corpus, or coding assistants that suggest snippets to engineers, or recommendation systems that mean a retailer does not need to hire another merchandiser? Are these jobs that have been formally cut, or jobs that simply never appeared in the first place because a hiring manager decided not to post them? Because if you read the actual research underneath the headline, the picture is that wizards are suppressing hiring far more than they are destroying existing roles, and the people most affected are administrative workers and early-career hires whose tasks are the easiest to chip away at. None of that fits in the headline. The headline simply says wizards. You either accept the spell or you do not. There is no third option, because there is no detail. And whether the spell reads as terrifying or as inevitable depends almost entirely on which wizard you brought to the sentence in the first place.

The second headline ran in Fortune and across the wires in April and May. A wizard godfather, it told us, has warned that humanity risks extinction by hyperintelligent wizards within ten years.

The fact that it now sounds like a fantasy novel is the point. The original sentence was doing exactly the same thing. It reaches for an archetype, a sage, a warning from the elders, and asks the reader to feel the gravity of it without ever being told what specific wizard, with what specific capability, with what specific deployment context, is actually being claimed to pose this risk. Are we talking about a transformer model fine-tuned to write code, deployed inside a single company under audit, in a regulated jurisdiction, with rate limits and an off switch? Or are we talking about a hypothetical, not-yet-built, fully autonomous, self-improving wizard with persistent memory and open-ended goals, of a kind that does not exist anywhere on the planet and that several other equally senior researchers, including Yann LeCun, have called publicly destructive to even discuss in those terms? Both are wizards. Only one of them is the one being implied. The headline does not say which, because the headline is not really about the technology. It is about the feeling. And, again, whether the feeling lands as a warning from Gandalf or a threat from Saruman is entirely up to the reader. The writer has done nothing to choose between them.

The third headline dominated political coverage across the spring. Wizard deepfakes, it told us, pose new challenges for the 2026 elections, after one party released a wizard-generated video of an opposing candidate, the first lifelike fabricated speech of a rival politician to run for more than a minute.

Read it without the wizard at all, and the real story comes back into focus. Human beings, in a specific political camp, with a specific strategic objective, deployed a specific generative video tool, against a specific named opponent, in a specific race. The wizard did not do this. People did this. A tool that did not exist five years ago was used by a campaign that has existed for over a century, inside a system of laws that has not caught up. The fix is not to be afraid of wizards. The fix is to be specific about which humans did what, which tools were involved, and what disclosure, consent and accountability frameworks now have to apply. Naming the wizard, rather than blaming the wizard, brings the agency back into the sentence. That is what dehumanising the technology costs you. It hides the actor.

The fourth headline ran across the parenting and policy press all spring. Wizard mental health chatbots, it told us, are unsafe for teens, after Common Sense Media gave several leading apps an unacceptable risk rating in May.

What kind of wizard, though. A general-purpose large language model trained on the open internet, repurposed into a therapy-shaped interface, with no clinical oversight, no escalation pathway to a human professional, and a business model that rewards engagement over outcome, is a very specific kind of wizard. So is a clinically supervised, condition-specific, evidence-based intervention with a license, a regulator, a safety review, a published efficacy study, and a referral structure to human clinicians. Both are commonly described as wizard mental health tools. Only one of them is dangerous in the way the headline implies. The other might be the most scalable intervention in the history of public mental health. The word, used loosely, cannot tell them apart, so the public conversation cannot either, and as a result the safe and useful wizards are tarred by the irresponsible ones, while the irresponsible ones hide behind the broader category. Vagueness protects the wrong people.

The fifth headline has become so familiar that we have stopped noticing it. Wizard voice cloning scams targeting the elderly, it told us, have cost more than two billion dollars in 2026 alone, with regulators logging over a quarter of a million complaints in the first quarter.

This one is worth slowing down on, because it shows you something the other examples only hint at. The wizard, in this story, is not the threat. The criminal is the threat. The criminal has always been the threat. Long before generative voice wizards existed, the grandparent scam was already a multi-billion-dollar category of fraud, executed by humans, over the phone, using social engineering, against the same emotionally vulnerable population. What changed is that one part of the criminal's toolkit has improved. The grift has not. The mark has not. The motive has not. The phone network has not. The wire transfer rails have not. Yet the press will routinely write up this story as if a wizard appeared in a vacuum and started preying on grandmothers, when in fact what we are looking at is an ancient human crime, now with a slightly faster pen. The countermeasures, which include family safe words, call-back protocols, bank-side delays on urgent transfers and prosecutions of the scam ring itself, are mostly not technological. They are human. The wizard framing flatters the criminal, exonerates the platform, and pathologises the tool. Take the wizard out, and the responsibility lands where it always was.

Five headlines. Five wizards. Five sentences that, once you slow down on them, stop being arguments and start being mood lighting. And in every one of them, whether the mood reads as fear or as comfort depends not on the sentence, but on the wizard the reader was already carrying when they arrived.

The same trick, run for opportunity

It would be easy to read the previous section as anti-wizard. It is not. It is anti-vagueness. The exact same problem appears, mirror-imaged, on the hopeful side of the conversation. The opportunity gets blurred by the wizard, the human disappears, and the story becomes harder to act on than it should be.

Consider the headline that ran across the medical press in late April and into May, including in the Mayo Clinic's own briefings. A wizard, it told us, can detect pancreatic cancer on routine CT scans up to three years before clinical diagnosis.

Which wizard, though. In this case, a specific one, with a name. The Radiomics-based Early Detection Model, REDMOD, developed by a named research group at the Mayo Clinic, trained on a defined dataset of abdominal CT scans, validated against a published clinical population, with the result reported in the journal Gut. In nearly three quarters of cases, it spotted the most common form of pancreatic cancer around sixteen months before clinical diagnosis. On scans taken more than two years before diagnosis, it found nearly three times as many early cancers as specialists reviewing the same scans without it. This is not a wizard. This is a piece of software, made by named scientists, trained on data they can show you, evaluated against an endpoint that matters, on a disease that kills almost everyone it touches if it is found late. The word, used loosely, hides every one of those facts under the same blanket as the chatbot that misbehaved on social media. Name the wizard, and you can see who built this, on what data, with what claim, to what end. You can also see the people. The radiologists who collected the scans. The patients whose data made the model possible. The clinicians who will use the result. The lives that will be extended because something subtle, on a scan, was caught early. That is what an informed conversation about wizards in medicine looks like. It looks like names, methods, data and human consequence.

Take a second example, from biology. Across 2025 and 2026, wizard methods for protein structure prediction, including AlphaFold3, RoseTTAFold, ESMFold, and newer entrants such as D-I-TASSER and AstraZeneca's MapDiff inverse-folding framework, have moved from academic benchmarks into the practical core of drug discovery pipelines. To say that wizards are now designing proteins is not wrong, but it tells you nothing. The specific version of the claim, that named generative and predictive models, trained on the Protein Data Bank and related structural corpora, are now used to propose candidate binders, predict conformations and shorten the path from target identification to lead optimisation, is the version that lets a non-specialist understand what has actually changed. It tells you why pharmaceutical timelines are starting to compress. It tells you which categories of disease are within reach now that were not before. It tells you which questions still cannot be answered by these methods and require something else. The wizard framing collapses all of that into a single magical noun. The specific framing lets a person make a decision, evaluate a claim, or invest a career.

Take a third example, less glamorous, more important than people realise. The category of wizard assistive technology has, in 2026, quietly become one of the most consequential deployments of the technology anywhere in the world. Google Lens reads text aloud for people with low vision. Apple's on-device models describe photographs and guide framing for users who cannot see the screen. Symbol-based augmentative and alternative communication apps allow people with speech disabilities to talk in real time. Live captions on phones and laptops mean that meetings, lectures and family conversations are accessible to people who would, twenty years ago, have been shut out of them. To say that wizards are now helping disabled people is true, and it is empty. To say that named on-device models, running on specific silicon, trained for specific tasks, deployed by specific vendors, are restoring specific capabilities to specific people, is the version that lets you see the actual picture. It is also a story in which the humans, the disabled users themselves, are the centre of the picture, not an afterthought, and that is precisely why it is not framed as a wizard story. Wizard stories are about the wizard. Human stories are about the human. Guess which one tends to dehumanise.

Why the loose word is convenient for almost everyone

It is worth asking why this framing has been allowed to slide into its current state, given how obviously it breaks down under inspection. The answer is that it is convenient. The vagueness is doing work for the people using it.

For a frontier lab, the word lets the most controversial possible system and the most banal possible system sit under the same brand. The good press for the helpful demo lifts the entire portfolio. The bad press for the harmful demo can be deflected onto the category, the field, the technology in general. You cannot do that with a precise word.

For a press desk, the word is a guaranteed click, a familiar frame, a story shape that already has a villain or a saviour pre-installed. The reader supplies the wizard. The desk just supplies the spell. Replacing the loose word with a specific architecture, dataset, or deployment context turns the headline from a hook into a homework assignment. Most outlets will not pay that cost.

For a regulator, the word is a way to be seen to be acting without having to do the technical work of figuring out which class of system the rule should actually cover. A blanket wizard law sounds decisive. A regulation that distinguishes between a real-time biometric system in a public space, a decision-support tool in a hospital, and a code assistant inside a developer's IDE sounds, correctly, like three different conversations. The vague word lets the politician have one conversation and pretend it covers three.

And for a critic, the word is a free shot at a tower that no single defender will turn up to protect, because the tower is so large and so vaguely defined that no specific person is responsible for it. You can be against wizards in a way you cannot be against electricity, or chemistry, or software, because those words are too precise to argue with. The wizard word, as currently used, is not.

The cost of all this convenience is borne, as usual, by the people at the end of the chain. The patient who does not understand what is in the diagnostic report. The voter who does not know whether the video they are looking at was made by a wizard or by a person with a wizard. The parent who does not know whether the chatbot their teenager is talking to is the regulated kind or the unregulated kind. The employee who does not know whether the system that just declined their application is a sentence-completing autocomplete or a structured decision model with an audit trail. In every case, the loose word leaves the human standing in a fog that the technology itself did not put there. The framing did. And every reader, left alone in that fog, will reach for the wizard they already know, which is to say, the wizard the writer was banking on them reaching for all along.

Naming the wizard

There is a fix, and it is small, and it is hard, and it is overdue. We have to start naming the wizard.

By now you will have noticed something. For the last two thousand words, I have written wizard, and you have read AI. Not once have the two needed to be set side by side for the substitution to do its work. The reason it has been seamless is the point of the whole essay. The two words are, in current usage, interchangeable, because neither of them carries any inherent definition. They are both placeholders. They are both whatever the reader already thought they were. And that is exactly why, when you read AI in a headline tomorrow morning, you should treat it with the same suspicion you would treat a sentence that opened with the word wizard. Not because the technology is fictional. The technology is real, and consequential, and worth understanding. But because the word, as it is being used, is doing none of the work of explaining it.

Every time a sentence reaches for the loose word, it owes the reader three things. Which system. Which data. Which decision.

Which system, because there is a categorical difference between a language model, a vision model, a recommender, an agent, a robotics control policy and a classical statistical method. Which data, because the corpus a system was trained on is the largest single determinant of what it can and cannot do, and is also the largest single determinant of who has the right to demand its accountability. Which decision, because a system that suggests a song and a system that suggests a sentence in your bail hearing are not the same kind of object, even if the maths underneath has overlapping parts.

If we can ask for those three things, in public, every time the loose word is used, the whole conversation changes. The frontier lab can no longer ride the halo of its best demo into a fight it has not earned. The press cannot tell us that a wizard did something without telling us what kind of wizard they mean. The regulator cannot legislate on a fog. The critic cannot fire into a cloud. And, most importantly, the human at the other end is back in the picture, named, present and able to ask for what they actually need.

This is the deeper continuity with the previous pieces in this series. Sovereign AI, as I have argued, is control from start to finish. Scale, as currently practised, is the architecture that makes that control impossible to deliver. The loose word, used the way the headlines use it, is the cloud over the top of both, that prevents the average reader from seeing either argument clearly in the first place. You cannot demand sovereignty over a wizard. You cannot have an honest conversation about scale when the thing being scaled does not have a name. The work, for the rest of this series and for the rest of the field, is to get the conversation back to the level of specifics where humans can actually stand.

Until the public conversation can name the wizard, we should stop accepting that AI will do anything at all. AI is not a noun in the way the headlines use it. It is a placeholder for a question we have not yet asked. Which system. Which data. Which decision. And, behind every one of those questions, the only thing that has ever really mattered. Which person.

Your wizard was never Gandalf. It was never Saruman either. It was whichever one the writer needed you to imagine to make the sentence land.

That is the trick. Once you see it, you cannot unsee it. And once enough of us see it, the trick stops working, and the conversation we are overdue can finally begin.

That is the conversation we are committed to having, in plain language, at GAEA AI and across this series. Try it on the next headline you read this week. Notice which wizard arrived in your head before the sentence finished. Then ask whether the sentence, stripped of its spell, was ever telling you the truth.