Arts, Design & Communication: How AI can best tutor this subject

What you'll be able to do: Use AI for the parts of creative work it can genuinely help with (critique, context, rehearsal) while protecting the one part a good answer from the tool would cost you the most: the first pass that was supposed to be yours.

The danger this subject has and the others don't

Engineering's tool gives you a confident wrong number. History's gives you a confident invented quotation. Both are dangerous, and both are catchable, check the arithmetic, chase the citation.

Arts, design and communication has a different problem: the tool can produce the actual thing you were supposed to learn to make. Not a wrong version of the essay or the calculation, a finished, plausible, often genuinely good photograph, melody, logo, paragraph, or speech, standing exactly where your artefact was supposed to stand. In a subject where the object you make is the assessment, and is also the entire point of doing the course, that's not a variant of the usual risk. It's a different threat in kind.

This isn't hypothetical. In 2023 a German artist, Boris Eldagsen, submitted an AI-generated image to the Sony World Photography Awards, won the creative category, and then publicly declined the prize to reveal how it was made, the judges had rated it against real photographs and it won. The same year, a track built from cloned vocals styled on two named recording artists, released anonymously as "Heart on My Sleeve": spread fast enough to be pulled from major streaming platforms and was submitted for, then ruled ineligible for, Grammy consideration. Neither of those was caught because it "didn't work." Both worked. That's what makes this subject's version of the problem different: the usual tell: it's wrong, it doesn't run, the source doesn't exist: doesn't fire, because passing was never actually the danger.

Why the first pass has to be yours

There's a second, quieter version of the same danger, and it costs you even when you never submit anything AI-made.

Creative work is a search through a space of possible answers, most of them bad, and a large part of the value is in the search, the twelve sketches that didn't work, the four false starts on the arrangement, the draft paragraph you cut that taught you what you actually meant to argue. A generative tool's entire design is to short-circuit exactly that search: skip straight to a plausible, well-formed output.

See that output before you've searched yourself, and several things happen at once. You anchor on it, your own attempts start getting judged against a version you didn't make, and they drift toward resembling it whether you intend that or not. Your idea space collapses toward the statistical centre of its training data, because a "generic-shaped" suggestion is generic precisely because it's the averaged, most-probable answer to your prompt, the opposite of the odd, specific, slightly-wrong-shaped idea that usually turns out to be the interesting one. And you lose the counterfactual entirely: you will never know what you would have made, because you can't unsee what you did make.

This is why the rule in this subject has to be stricter than "use it after you try." It's: form the idea before you open the tool, every time, including the days you're staring at a blank page and it would be so easy not to.

What it's genuinely good at here

Critique in forced, specific vocabulary, once "I like it" and "it doesn't work" are pushed until they name an actual element: this is the single highest-value move in the subject, because it's the vocabulary a crit runs on.

Playing a difficult audience: the client who wants it cheaper, the festival programmer who's seen a thousand of these this year, the sceptical examiner. Rehearsing the defence before you need it live changes the defence.

Locating you in a lineage: what your approach is reacting against, and what it's quietly borrowing. Style stops looking arbitrary once you can name the alternative it's a response to.

Analysing a technique without producing the work: how a specific effect is achieved, in terms you can apply by hand. The line between this and outsourcing your own work is exact, and it's worth learning precisely where it sits.

Inventing a brief tighter than the one you were given: most creative paralysis is too much freedom, not too little, and the constraint you invent yourself is more generative than one handed to you.

Structured revision loops without the friction of finding a human free every time, checked against real exemplars at several grade bands so you can see what development actually looks like rather than guessing.

Rehearsal for a live moment: predicting where a run will fail before you attempt it, and reconstructing what actually happened immediately after.

What it's bad at here, specifically

Judging genuinely new work. It scores against the statistical centre of what it's seen. The further your work departs from convention, the less trustworthy its praise or its criticism, a tool trained on a century of recorded jazz will tend to read playing that's genuinely outside the changes as wrong rather than as outside, because it has no way to distinguish those from text alone.

Telling a deliberate choice from a mistake. Is that dissonant cluster a wrong note, or a Ligeti-style tone cluster placed on purpose? Is that blown-out highlight a metering error, or the deliberate flash overexposure Nan Goldin used for its rawness? It cannot hear or see your intention, only the artefact, and it will default to "fix" over "why."

Anything embodied. It cannot hear your intonation drift, see the tension in your bowing arm, feel your breath support fail two lines from the end of a speech, or watch your posture collapse under nerves. Everything it says about your performance is a description of your description of it.

Producing the work and it staying yours. Covered above, this is the subject's defining risk, not a footnote.

Assessing process it hasn't seen. Shown only the final piece, it can only critique the final piece. It has no way to know whether you iterated forty times or once, and in studio assessment, that's often most of the mark (see article 07).

History and attribution as fact, same as everywhere else this tool appears, movement dates, who influenced whom, what a designer actually said about their own work. Verify before it enters an essay or an artist statement.

The shape of a good session

1. Make the first pass alone: thumbnails, a rough progression, a bad draft, a scratch recording, before the tool has seen the brief.
2. Ask for critique in named vocabulary, not a verdict.
3. Ask what your approach is reacting against, and check it against a real source.
4. Compare against real, named exemplars at a band above and below your own.
5. Isolate what's actually failing and drill only that, rather than running the whole thing through again.
6. Defend the decision to a stakeholder who won't accept "it felt right".

The instruction to set

For this session you are tutoring me in [visual art / music / design / film / public speaking: name it]. Rules: never generate the piece, the passage, the draft or the image itself, even as "just a starting point"; if I ask you to, ask me why I'm stuck instead of producing anything. When I show you work, don't say whether it's good, ask what I was trying to do, then tell me where it succeeds and fails against that. Push back on vague terms: "it flows", "it doesn't feel right": until I can name the specific element and where.
Compare my work to real, named examples, never generic ones. And if I ask you to judge something genuinely unconventional, say so rather than defaulting to "this doesn't work."

The failure that looks like success

A finished piece that is competent, presentable, and could have been produced by anyone who typed the same brief in. It gets a fair grade on craft, because the craft is fine. It generates nothing in the crit, because there's nothing to discuss, no decision was made that the maker can defend, because the maker didn't make the decisions.

The subject-wide version is Eldagsen's photograph and "Heart on My Sleeve": both were good enough to clear real gatekeeping, an international judging panel, a streaming platform's algorithm and a listening public, before anyone checked how they were made.

The tell: you can't say why you chose this over the alternative, for the specific choices in front of you, or, for a performed piece, you can play it correctly but can't say what you'd have done differently if bar nine had needed to go somewhere else.

What it cannot replace

The search itself. The bad sketches, the abandoned draft, the take you threw away, that's not waste around the real work, it's a large part of how the real work gets decided.

The hand, the ear, the body's development. Built only by repetition of the physical act. A description of correct bowing technique doesn't put it in your arm; a thousand hours of your own bow does.

The stakes of showing unfinished, possibly bad work to real people, live, and living with the reaction. A screen has no social cost. A crit, a recital, a stage does, and that exposure is not incidental to this subject, It's a large part of what it's training.

Where to go next

00b has the ten methods that matter most here and why two popular study habits, generate-then-edit, and playing the piece through from the top every time, actively work against you. 07 Standard-setting comparison is the one that removes the surprise about what's actually being graded. 08 Failure-point prediction is the direct fix for practice that never touches the four bars that are actually costing you the piece.