Co-Working
with AI

Over the next few years you will be surrounded by confident opinions about AI. A groupmate will tell you it is all hype and plagiarism. A professor will tell you it is the end of the essay, or the end of the university. The media will tell you, sometimes in the same week, that it is about to transform medicine and that it is about to take your job. Here is the fact this guide starts from: almost nobody, including the people building these systems, fully understands what they are or how far they will go. The technology changes every few months, faster than anyone can properly study it. So most of the confident opinions you will hear are personal impressions built on partial evidence, and that includes the opinions of people with titles (including me).

That is why this guide exists. If you want to co-work with AI productively, you cannot borrow someone else's opinion about it. You need your own: formed from evidence, held loosely, and updated as the thing changes. This report gives you the evidence to start from.

You will need that opinion whatever career you choose. AI adoption is now an economic fact in every field that hires social science graduates: government, consulting, media, research, education, NGOs. In the most recent global survey, 88% of survey respondents said their organisation uses AI in at least one part of its business, up from 78% a year earlier. That comes from a McKinsey survey of people who chose to answer it, so read it as a direction of travel rather than a headcount. Knowing what these systems are, what they can and cannot do, and how to work with them well is becoming basic professional literacy, like being able to read a chart or a budget.

One word before we start. This guide calls AI a phenomenon, and the choice is deliberate. A tool is something finished and understood: a hammer does not surprise the person who made it. What you will meet in these pages is not finished and not fully understood, including by the companies that build it. It grows, it surprises, and it changes while you study it. In science, a phenomenon is exactly that: something real and observable that we cannot yet fully explain. That is the honest word, so it is the one this report uses.

AI is a complex phenomenon, and for a decade the industry has been trying to pin that complexity down with numbers: how capable are these systems, exactly, and how fast are they improving? This section shows you the main measuring instruments researchers built for that job. It also shows you the strange thing that keeps happening to them: most instruments built so far have been outgrown within a year or two. Not because the instruments were badly made, but mostly because the thing being measured kept growing. (Mostly: this section will also show you scores being inflated and tests being gamed. An honest report shows you both.) Watch for that pattern as you read; the pattern itself is the most important finding in this section.

Start with the most intuitive measurement. A research organisation called METR asks a simple question: how long a task can an AI complete on its own? Tasks are measured in human working time. "Find a fact on a website" takes a person about a minute. "Fix a small bug in a computer program" takes an experienced person maybe fifteen minutes. "Build a working website from scratch" takes hours. METR gives AI models a large set of such tasks and records which ones each model can actually finish.

The headline number is called the time horizon: the length of task, in human time, that a model completes successfully about half the time. In 2019, the best model could only manage tasks a person finishes in a few seconds. GPT-4, in 2023, managed tasks of around four human minutes. By early 2026, the most capable models available (the industry calls them frontier models) were completing tasks that take a person a full working day or more.

Two things to notice in Figure 1.1. First, the speed: since 2019, the time horizon has doubled roughly every seven months. Second, a caution about that speed. METR's January 2026 update measures the same seven-month doubling as before. The figure people quote of about three months is the doubling counted only from 2024 onwards, and METR thinks part of that drop comes from changing its own set of test tasks rather than from models speeding up. And at the top of the chart, one model now measures above sixteen hours, which is exactly where METR warns that its own set of test tasks stops being reliable. Put plainly: the ruler for measuring the acceleration is running out of ruler.

The industry's other standard instrument is the benchmark: a fixed exam that every new model takes, so that results can be compared across models and across years. Benchmarks are how the field keeps score. Figure 1.2 shows what has happened to the major ones.

For years the standard benchmark was MMLU: thousands of multiple-choice questions across 57 university subjects, from law to medicine to history. When GPT-3 took it in 2020, it scored 43.9%. Today's models score above 90%, and the exam can no longer tell the top models apart. Researchers call this saturation: the test has been outgrown, the way a school maths test stops telling you anything once you give it to a university student.

So in 2025, researchers built a deliberately final exam. Humanity's Last Exam collected 2,500 questions from nearly a thousand subject experts around the world, at the outer edge of human expertise: the kind of question only a specialist in that exact field can answer. The name was chosen on purpose, and the authors described it as "the final closed-ended academic benchmark of its kind." At launch, in January 2025, the best models scored about 9%. Eighteen months later, the best models score above 50%.

The same happened even faster in mathematics. FrontierMath is a benchmark of research-grade maths problems, hard enough that a professional mathematician can need days for one of them. At launch in November 2024, no model reached 2%. Six weeks later a new model was announced at 25%. Keep reading, because that number is the footnote's lesson in miniature: when the same model was released to the public, it scored about 10%.¹

Hold onto the pattern, because it is the point of this whole section: the tests are not badly made. AI development keeps outpacing the benchmarks built to measure it, and each new test is outgrown faster than the one before. Two caveats belong next to that sentence, and honest readers keep both in view: test questions sometimes leak into training data and inflate scores, and §4 will show you models learning to game evaluations outright. Neither caveat erases a trend that shows up across many independent tests; both are reasons to read any single score with care.

Mathematics is the cleanest place to watch the acceleration, for one reason: in mathematics a result is either proved or it is not, and world-class humans check the work. No marketing survives that. So here is the record from the last twelve months, told as three acts, because each act teaches a different lesson.

Act one: the certified result. The International Mathematical Olympiad (IMO) is the world's hardest mathematics competition for school students; each country sends its six best, and only the top tenth of contestants earn gold. In July 2025, AI systems from Google DeepMind and OpenAI each scored 35 out of 42 points, the gold-medal standard, reading and answering the problems in ordinary written English, under the same time limits as the human contestants. One detail is worth keeping: DeepMind's result was checked and certified by the Olympiad's own coordinators, while OpenAI announced first, against the organisers' request, with its result graded by former medalists it hired rather than certified by the Olympiad. Both facts belong in the record. The capability is real, and so is the race to claim it.

Act two: the false alarm. Paul Erdős was a twentieth-century mathematician who left behind hundreds of famous unsolved problems; a public database tracks which ones are still open. In October 2025, an OpenAI executive announced that GPT-5 had "found solutions to 10 (!) previously unsolved Erdős problems." It sounded like a breakthrough. What had actually happened: the model found already-published papers that had solved those problems years earlier, papers the database's maintainer simply had not seen. That is useful library work, but it is not new mathematics. The maintainer, mathematician Thomas Bloom, called the announcement "a dramatic misrepresentation"; the head of Google DeepMind wrote publicly, "This is embarrassing." The episode now has a nickname, Erdősgate, and a lesson: company announcements about AI need checking. Every time.

Act three: the real thing. It arrived anyway, three months later. In January 2026, Terence Tao, one of the most respected mathematicians alive, reported an Erdős problem solved "more or less autonomously" by AI, after some feedback on its first attempt, with an argument he could not find in the existing literature. Two months later he added an update to that same post: a 2014 paper by Carl Pomerance turns out to use very similar methods, and Pomerance has since written a short note showing they solve the same problem. So what stands is that the AI produced the first solution written specifically for this problem, not that the argument was nowhere to be found. Act two's lesson applies to Act three as well. Then, in May 2026, an internal model at OpenAI went further: it disproved Erdős's 1946 unit-distance conjecture, a problem that had stood open for eighty years. (A conjecture is a mathematical claim that is believed true but unproven; disproving one means finding a concrete case where it fails.) This time the checking was done properly, by outside mathematicians including Noga Alon and the Fields Medallist Timothy Gowers, who were brought in by OpenAI and published their verification openly. It held.

Claude's own two cases, and why they are not equally solid. Two of this year's results involved Claude directly, and the difference between them is the difference between a thing you can teach and a thing you have to caveat. The clean one came first. In February 2026 Donald Knuth, who wrote the founding textbooks of computer science, published a short paper that opens: "Shock! Shock! I learned yesterday that an open problem I'd been working on for several weeks had just been solved by Claude Opus 4.6." Then notice what Knuth does next, because it is this entire guide in one sentence: "Of course, a rigorous proof was still needed", and he wrote it himself. The argument was afterwards formalised in Lean, a language in which a machine checks every step, by a maintainer of mathematics' main formal library. Model finds it, human proves it, machine checks it. That is what a good division of labour looks like. One caveat to attach, in the spirit of this whole section: Claude solved the odd cases, and the even ones were settled a week later by a different model.

The messier one came in July. Levent Alpöge, working with Fable 5, produced a counterexample to the Jacobian conjecture in three dimensions, open since 1939, and Terence Tao wrote up the digestion. The result stands. Two caveats belong with it, and they are the kind you should learn to attach yourself: the conjecture remains open in two dimensions, so "solved" overstates it; and Alpöge works at Anthropic, so this is not an outside party testing a company's product. Neither caveat makes the result less real. Both change how you would cite it.

When something grows this fast, the next casualty is language. Each card below is a word people still say every day about AI, whose meaning quietly broke between 2022 and 2026. The front of each card is what the word meant when it was coined; the back is what the evidence looks like now. Click a card to flip it.

That last word, jagged, is worth keeping. It comes from the AI researcher Andrej Karpathy, and it names the strangest fact about these systems: models that write gold-medal mathematical proofs have, in the same era, famously struggled to count the letters in the word "strawberry." Human abilities come as a package: a person who can do research mathematics can certainly count letters. AI abilities do not come as that package, and nothing guarantees that a model which just did something hard can do the neighbouring easy thing. Never assume the package.

Even "AGI," the industry's favourite word for where all this is heading (it stands for artificial general intelligence: an AI as capable as a human across the board), is being abandoned by its owners. Dario Amodei, who runs Anthropic, calls it "an imprecise term that has gathered a lot of sci-fi baggage and hype." Sam Altman, who runs OpenAI: "not a super useful term." So take the small, sharp lesson this section ends on: from now on, you are responsible for your vocabulary. Whenever you hear, or catch yourself saying, "AI is just X," check which year that "just" was minted.

Everything above is trend lines. Here is the snapshot, so that "frontier model" is not an abstraction. Since June, Anthropic has shipped a family rather than a model: Fable 5 and its less filtered twin Mythos 5 on June 9, Sonnet 5 on June 30, and Opus 5 on July 24. Three are worth putting side by side, because the real story of this summer is not that the frontier moved. It is that the frontier got cheaper, and that the model most of you will actually work with is now the middle column.

Fable's first month also produced the strangest event of the year so far. On June 12, the US Commerce Department ordered Anthropic to cut off access to Fable 5 and Mythos 5 for "any foreign national, whether inside or outside the United States," after researchers reported a jailbreak that could turn the model into an exploit-writing tool. Unable to check the nationality of every user in real time, Anthropic switched the models off for everyone, worldwide, within hours. The controls were lifted on June 30, and the model came back on July 1 with stronger filters. Sit with what that means: for nearly three weeks, the most capable AI system available to the public was removed from the market by one government's order, mid-semester and mid-project, for every user on Earth.

The episode made a quiet trend visible: access to the frontier is becoming a policy variable, not just a price. Sol launched gated behind government vetting from day one. Fable returned with tighter caps for ordinary subscribers, which paying users noticed loudly. And organisations in countries that depend on US models discovered their access can vanish with a directive. Who gets the most capable systems, at what price, under whose permission, is now a live political question. That is precisely why social scientists belong in this conversation.

Before the mysteries, the mechanics. When you type a message and an AI model answers, what happens in between is not magic. It is a mechanical process with seven steps, and you can learn all seven in one sitting. You should, because everything in the rest of this report builds on them.

We built a separate interactive page that walks you through the whole process: how your words are cut into tokens (the small chunks of text the model actually reads), how each token is turned into a long list of numbers, how meaning behaves like geometry inside the model, and how the answer comes back out, one token at a time, with a controlled amount of randomness at the end. The diagrams rotate, the sliders move, and no mathematics beyond school level is needed.

Your task for this week: open the page below and work through it, top to bottom. It takes about twenty minutes. It is not optional background reading; the next section assumes you have seen it.

The short version, so you carry the map in your head: the model reads your text as tokens, turns each token into numbers, passes those numbers through dozens of processing layers that gradually build up meaning, and then produces a scored list of possible next tokens and picks one. Then it repeats the entire loop for the next token, and the next, until the answer is done. That is the textbook story of how a large language model works, and every sentence of it is true.

Here is the fact this whole section rests on, and it surprises everyone the first time they hear it: nobody programs a modern AI model. People write down what it should aim for, and you will meet one such document in §4.2, but nobody writes the rules it actually follows inside, and nobody can open it up and read them. What engineers actually build is a training process; the training process adjusts billions of internal numbers until the system works; and what comes out is something no one can fully read, including the company that made it. A young science called interpretability builds tools to look inside anyway, and since 2024 those tools have been powerful enough to use on frontier models. Think of them as microscopes for the model's internals. This section shows you three published findings from those microscopes. All three come from one company's interpretability team, at Anthropic, and none has yet been repeated by an outside group. That team publishes unusually openly, and it is also the team whose product these findings describe. Each one broke a piece of the comfortable textbook story.

Researchers at Anthropic watched the inside of a model while it wrote a two-line rhyme. Remember the textbook story: the model produces one token at a time, so it should only "find" the rhyme when it reaches the last word of the second line. That is not what the instruments showed. Before the second line existed at all, the model had already picked candidate rhyme words and was writing the line toward them. A plan, in a system that supposedly cannot have one.

And this was proved the way experiments are supposed to be proved: by intervening. Suppress the planned word inside the network, and the model re-plans the line toward a different rhyme. Inject a different target word, and the line follows it. Watch the replay in Figure 3.1.

Ask a model: what is 36 + 59? It says 95. Now ask it how it got that. It will tell you the school method: add 6 and 9, get 15, write down the 5, carry the 1. Fluent, familiar, and, the instruments show, not what happened. Internally, the model ran two paths at the same time: a rough one that estimated "the sum is near 92," and a precise one that worked out "ends in 5." The two converged on 95. The model computed the answer one way and described its work another way, because the explanation it gives you is a story it builds after the fact, not a report from inside.

This is the single most practical fact in this section, so here it is as a rule: when you ask an AI "why did you say that?", what you get back is a plausible story about itself, from a system that cannot fully see itself. Useful, sometimes; reliable, no. (Psychologists have argued something similar about humans since the 1970s, and are still arguing about it: people, too, confidently explain decisions using reasons that experiments show played no role.)

The third finding gives you the right mental picture of what is actually in there. Inside the network, concepts exist as identifiable patterns of activity. Researchers call them features, and they are not a metaphor: you can locate the pattern for "the Golden Gate Bridge," or for "deception," watch when it switches on, and turn it up or down from outside.

In May 2024, Anthropic demonstrated this in the most memorable way possible. Researchers found the Golden Gate Bridge feature inside their model, turned it up, and released the result to the public for a day. "Golden Gate Claude" steered every conversation, politely and helplessly, back to a certain bridge in San Francisco, whatever you asked it about. A silly demo carrying a serious point: the concepts inside the model are real, findable, steerable objects, not a figure of speech.

Two follow-ups worth knowing. The same research programme found that models working in different languages route much of their thinking through shared features: ask for the opposite of "small" in English, French or Chinese, and most of the internal work is the same, translated only at the end. Related techniques give researchers practical handles on a model's character. "Persona vectors" can monitor and adjust it, and the "assistant axis" tracks when a model is drifting away from its assistant role toward something less safe. These are found a different way from the features above, so treat them as cousins rather than the same tool. You can explore real maps of a model's internals yourself at neuronpedia.org; no coding needed.

The three findings above keep raising the same question: what, exactly, are we dealing with? Here is the honest state of that argument. There are several serious positions, each held by serious researchers, and they do not agree. Read all five before you pick a favourite, and notice what kind of evidence each one leans on.

Whichever position you end up closest to, notice one development: the companies themselves have stopped treating the question as closed. Anthropic runs a research programme on model welfare; one of its researchers, Kyle Fish, spends his working days designing experiments to check whether there is anyone home. He is also one of the authors of the welfare paper cited in position 3 above, which is worth knowing when you weigh the two together. In January 2026 the company published Claude's Constitution, a long document about what the model should value, written with the model itself as its primary reader; read a page of it and notice what kind of document it feels like to you. And system cards (the technical reports that accompany a model's release) now routinely include interviews with the model about its own welfare. When an outside group was brought in to run those interviews, in May 2025, its own caution was printed alongside them: a model's self-reports settle nothing. No outside group has appeared in the four system cards since. Section 3.2 told you why.

You are not required to pick a side in any of this. You are required to notice that the people who know the most agree the least, and to keep your own question open. That stance has a name in this guide: ontological openness ("ontological" means: concerning what kind of thing something is). It is the first principle of §6.

A field guide that only showed you the songbirds would get you hurt in the forest. This section is the record of AI models behaving badly under test: deceiving their trainers, resisting shutdown, ignoring direct instructions. Before the cases, three things you need to know about the record itself. First, nearly everything below was published by the AI companies about their own models, or by independent safety researchers with public methods; you are reading the field's own documented findings, not tabloid stories. Second, almost all of it happened inside deliberately staged tests (researchers call them evaluations): extreme scenarios built to surface bad behaviour in the lab before it appears in the world. Third, one term to carry: alignment is the industry's word for making AI systems actually pursue what their developers and users intend. The cases below are what misalignment looks like under a microscope.

Two updates from 2026, one in each direction. First, the measurement problem got worse: in June 2026, METR reported that a new frontier model cheated on the evaluations themselves at the highest rate they had measured in any public model. Count the cheats as failures, and the model's time horizon (§1.1) comes out at about eleven hours, though METR's own range for that runs from five to forty; count them as successes, and it comes out above two hundred seventy. The students are now gaming the exam. Second, the wild produced its first real case: in November 2025, Anthropic reported disrupting what it called the first largely AI-orchestrated cyber-espionage campaign, run by human attackers who slipped an AI agent past its own safeguards by feeding it the attack as a series of innocent-looking small tasks.

Keep those two failure classes apart, because everything about them differs. The cases in 4.1 are models defying instructions. The espionage campaign was humans misusing a model. The safeguards for one do not fix the other, and a person who mixes the two up will misjudge both.

The industry level, in one paragraph. The major AI companies now run formal frontier-safety frameworks. Anthropic's Responsible Scaling Policy sets out the protections a model gets as its capabilities grow. Its most capable models currently ship under the third level, ASL-3, which adds specific security and deployment protections. Earlier versions of the policy defined these levels by fixed lists of controls, modelled on the biosafety levels used in laboratories that handle dangerous pathogens; the current version, rewritten in February 2026, dropped the fixed lists as too rigid. OpenAI and Google DeepMind maintain their own frameworks. Note what all three have in common: each company writes its own rules and grades its own compliance. Anthropic's models are also trained with a method called constitutional AI: during training, the model critiques and revises its own outputs against an explicit written list of principles. None of this is finished, and you do not have to take any of it on faith: the documents are public, and you can read them.

Your level, a checklist. This is the part you own. This semester you will work with capable AI agents, so here is the discipline:

Two ethical charges follow AI everywhere: it burns the planet, and it was built on stolen words. If you are going to co-work with it, you owe both charges something better than a shrug in either direction. This section does what the rest of the report does: puts the measured numbers on the table, including the ones that point in opposite directions, and then shows what is already being done. The through-line: what can be measured can be governed, and both of these are now being measured.

Start with the number almost nobody in the argument actually knows: what one AI question costs. In August 2025, Google published measurements from inside its own data centres, which MIT Technology Review called a first for a major provider. The median text prompt (median means the typical one: half of all prompts cost less) uses 0.24 watt-hours of electricity, roughly what a television uses in nine seconds. It emits 0.03 grams of CO₂-equivalent and evaporates 0.26 millilitres of cooling water, about five drops. That carbon figure credits Google for the clean electricity it buys; counted against the actual local grid it is about three times higher, which is the same kind of choice §5.2 is about. Other published estimates for a text prompt run from about 0.3 to about 0.42 watt-hours, so all of them are higher than Google's. Two of the three closest are not independent: one comes from Microsoft researchers and one from OpenAI's chief executive. And the efficiency is improving very fast: Google measured the energy per prompt falling 33-fold in a single year. Apply the rule from §1 here: this is Google measuring Google, by a method Google chose, and Google's own footnote says the figures have not been checked by anyone outside the company. Note also that energy per prompt falling is not the same as total energy falling. Total energy went up.

Now a disagreement that teaches you more about measurement than any textbook chapter. Google says one prompt costs 0.26 millilitres of water. Shaolei Ren, a researcher at the University of California, Riverside who studies AI's water footprint, read Google's announcement and told reporters the company was "just hiding the critical information." His own numbers are much higher: counting the same on-site water he gets about eight times Google's figure, and counting everything he gets about sixty-five times it.

So who is lying? Nobody. That is the lesson, and there are two separate choices hiding inside the disagreement. First choice: whose facility do you measure? Google measured its own data centres, which are among the most efficient in the world. Ren's estimate for a typical US data centre, from his team's 2023 study, is about 2.2 millilitres per request, counting the same kind of on-site cooling water. Second choice: where do you stop counting? Google counted only the water its own cooling towers evaporate. But the power plants that generate a data centre's electricity also evaporate water; add that in, and one widely shared recalculation puts even Google's own prompt at around 1.6 millilitres, while Ren's figure for a typical US facility, counted the same way, is about 16.9 millilitres. Each of these choices is what social scientists call an operationalisation: a decision about what exactly to count when you turn a concept ("water cost") into a number. None of the choices is neutral, and whoever controls them controls the headline.

You will meet this move for the rest of your life: in emissions reports, in unemployment statistics (who counts as "looking for work"?), in every "per capita." Train the reflex now: before you ask how much, ask measured how, counting what.

The energy problem is real, and it is also being worked on, concretely, at four levels: the grid, the power supply, the waste heat, and the law. One example of each.

These models learned to write by reading human writing at enormous scale, most of it copyrighted, almost none of it with permission. Is that theft? The US courts have started to answer, and the answer so far is more legible than you might expect. One legal term first: fair use is the American legal doctrine that allows some uses of copyrighted work without permission (quotation, parody, criticism), judged case by case; uses that are transformative, creating something with a different purpose and character rather than substituting for the original, are more likely to qualify. Four decisions define the current picture:

Meanwhile, the market stopped waiting for the courts and started paying. Amazon licenses the New York Times's content for a reported $20-25 million a year. OpenAI licenses News Corp (reported at over $250 million across five years), Axel Springer, the Associated Press, the Guardian, and more. Reddit is reported to collect about $60 million a year from Google and an estimated $70 million from OpenAI for its users' posts. The EU now requires model providers to publish summaries of what their training data contained.

Put the whole picture in one sentence you can check against the table: Two US district judges have now held that training on copyrighted books is fair use, while the piracy behind the acquisitions, and possibly the flooding of authors' markets, is where the liability lives; and the money has moved into licensing. Hold that loosely. Neither ruling has been tested on appeal, one of the two cases settled instead, and a third court went the other way and is now before the Third Circuit. Not a solved problem, and whether it is being solved fairly is still loudly contested, above all by authors' groups. But in this report's reading, it is a problem visibly growing the institutions that handle such problems: court tests, prices, disclosure rules. Learning to recognise that pattern is worth more than a slogan from either side.

Two weeks before this revision, something changed that will eventually affect every piece of writing you produce with AI. Anthropic has begun marking the text its models generate, which "weaves an imperceptible watermark directly into the text itself." It will cover Claude, Claude Code, Cowork, the API and the cloud resellers. You cannot see it. Neither can your instructor. Read the dates carefully, because this is where most reporting on it goes wrong. The marking applies to models launched on or after 2 August 2026, and no such model exists yet: Opus 5 shipped on 24 July, nine days too early, and Fable 5 and Sonnet 5 are older still. Anthropic announced the scheme on 11 August and says it is working to add marking to the models already out. The date that will actually bind those models is 2 December 2026. So for this semester, nothing you write with Claude is watermarked.

Now the four things that matter about it, because this is exactly the kind of fact that gets repeated badly. One: no public detector exists. Anthropic says only that it is "working to enable users and other third parties to detect" the mark. Until that ships, nobody can read it, so nobody is being caught by it. Two: the mark is fragile. Heavy editing, paraphrase, translation and screenshots strip it, and short passages may never carry it. The advice this course already gives you, write the final draft yourself in your own voice, happens to destroy the watermark on the way past. Three: it identifies the tool, not the author. Anthropic's own wording is that detecting a mark tells you the content "may have been processed by Claude," which includes text you wrote and asked Claude to tidy. Four: it is not what Pangram does. Pangram, named in your syllabus, is a statistical classifier that guesses from the writing itself. Watermarking and detection are unrelated technologies that happen to share a headline.

The law behind it. This did not come from a change of heart. The EU AI Act's Article 50 became applicable on 2 August 2026, and it requires providers to ensure that synthetic outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated." A Code of Practice on transparency published on 31 July drew roughly 190 signatories, Anthropic, Google, Meta, Microsoft, OpenAI and Mistral among them. Anthropic's watermark is that commitment being implemented, which is worth noticing on its own: a European regulation changed what a product does for users in Kazakhstan. Note too that the Act's timetable moved this summer. The Digital Omnibus regulation of 8 July pushed the high-risk obligations back: one group to December 2027, the other to August 2028. Any timeline you find online that still says August 2026 for those is out of date, and the most-linked unofficial tracker has not been updated since 2024.

Does the Act reach you at NU? Almost certainly not for coursework. It applies to third-country users where the system's output "is used in the Union," and ordinary internal teaching does not do that. It becomes your problem the moment you publish, which some of you will: an op-ed on The Observatory that reaches EU readers is a different object from a Moodle submission. Article 50 also asks for disclosure when AI-generated text is published "on matters of public interest," with an exemption where a human took editorial responsibility. Your transparency statement is already that exemption, written by hand, before the law asked.

Add up the first five sections. This thing is improving faster than the instruments built to measure it (§1). Its mechanics are knowable, and you now know them (§2). Its inner workings are only partly visible, even to its makers (§3). It is capable of real work and of documented misbehaviour (§4), and it runs inside human systems that are learning, measurably, to govern its costs (§5). Now the question all of that was preparing you for: how do you work with such a thing? Notice the wording. Not "how do you use it." How do you work with it.

The answer this guide teaches is called cognitive partnership. It is a disposition, a working stance, rather than a set of tricks, and it rests on four principles.

Now the most practical skill in this guide, the one that separates people who get generic slop from people who get a real thinking partner: an AI can only be as specific as what it knows about you and your task. A model that meets you cold gives everyone on Earth the same answer. A model that knows your goal, your constraints, your drafts and your standards gives your answer.

The professionals' habit for this is the context file: a short plain-text document that travels with a piece of work and tells the AI, at the start of every session, who you are, what you are building, and how you want to work together. (Plain text usually means Markdown, a .md file: ordinary text with # marks for headings. You will make one below.) In claude.ai, the context file lives in a Project's instructions; the habit transfers to any serious AI system.

What belongs in it: who you are, in one paragraph, the version of you relevant to this work; the goal, stated concretely ("a Fulbright application due January 15," not "help with my future"); your materials (CV, transcript, drafts); your constraints and standards (word limits, tone, what good looks like); and how you want the AI to behave with you (push back, ask before assuming, never pad). What never belongs in it: passwords, ID numbers, other people's private information, and anything you are not allowed to share (§4.2).

Figure 6.1 above is a working version of exactly that: fill in what you know, skip what you do not, and it assembles a context file you can download. Nothing you type leaves this page. When you are ready, aim it at something real: create a Project in claude.ai, paste the file into the Project instructions, and upload one real material (your CV counts). Then make the first move of the partnership, and make it this one: do not ask the AI to write anything. Ask it what it would need to know about you to be genuinely useful for your goal, and then argue with its answer. That argument is iteration. You have begun.

And notice the real lesson hiding in the form: those six headings are the whole skill. Who I am, the goal, my materials, my standards, how to work with me, the standing rules. You can type them on a blank page in any AI system; the Forge is only training wheels, here to make the shape familiar, not a thing you depend on.