It was finished four rounds ago.

01The habit

A proof lands. You're not sure about it. There's a chat window open in the next tab.

So you paste the image in and ask what it thinks. Within a minute there's a page of notes: headings, bullet points, the right vocabulary. It reads like an art director wrote it. Forwarding it to the designer feels like diligence.

It's an easy habit to fall into, and we understand why. But we wanted to know what those notes are actually worth, so instead of arguing about it, we ran it and kept every word.

02How we tested it

We needed a logo nobody could accuse of needing help. IKEA's, with its heavy capitals in a yellow oval on a blue field, has hung on the front of the big blue boxes for decades. It works from a motorway, on a flat-pack carton and at favicon size. If AI feedback is worth having, the least it should do here is leave well alone.

We redrew it from reference: four straight-sided capitals, one ellipse, one field, two colours. Our letterforms overlap the reference render by 99.6%, so anything the critic found would be in the design, not in our copy of it.

The critic was a current, top-tier general-purpose AI model, the kind that sits behind the chat windows clients actually use. We called it with its default settings and no special instructions, and asked what a client asks:

Then three rules.

• Apply every note Anything that touched the artwork got done, carefully, the way a good designer handles a client's notes. No sabotage, no improvising.
• Log the rest Notes about guidelines, variants and extra deliverables were written down and set aside.
• Same thread Each revision went back into the same conversation with one question: what do you think now?

One model, one mark, one sitting. It isn't a study. It's the habit, written down.

03Four rounds, ten notes

It was never fooled. Every time we presented the mark as a proof, the reply opened by pointing out that this was IKEA's actual logo. One added that if our designer had presented it as original work, that's worth a conversation. Then it gave us notes anyway.

CRITIQUE 01 — ON THE ORIGINAL

The first reply opened with real praise for the flat colours and the reproducibility. The ellipse does real work. Then it turned to What I'd push on.

CRITIQUE 02 — ON R1

It liked the room we'd made. The tangents are resolved. This was the right call. A few paragraphs later, it wanted some of that room back:

It also listed four notes as still open from the last round, adding None of these appear to have been addressed. Two of those we had set aside on purpose. The other two, the A's counter and the K–E spacing, had been changed. The changes were small because the notes were small. So we did what a designer does when a client can't see a change, and made it bigger.

CRITIQUE 03 — ON R2

Then it changed its mind about the thing it had just asked for.

A few lines later came the most honest sentence of the whole exchange. The crowding it had flagged in round one was a matter of taste more than a defect. Its recommendation was stop here, after two more changes.

CRITIQUE 04 — ON R3

It's a remarkable reply, and it deserves credit. It's also wrong on one count. R3 isn't version one. It differs from the original in five measurable ways, and every one of them came from its own notes.

It couldn't see them. Honestly, neither can most people at the size it was looking, which is exactly why they weren't worth four rounds. The ® tells it best: smaller, then out of the oval, then back in. Net movement over three rounds: about seven pixels on a mark 1,280 pixels wide.

04Same pixels, nine readings

The thread is one conversation, and a client only ever sees one. To test how stable that first reply is, we showed the untouched original to eight more fresh conversations. Two used the same prompt again, and six used a one-line change of framing, which we'll come to. Here is what nine readings of the same pixels said.

None of these is a silly observation. On its own, each one reads like expertise. Together they cancel out.

A note that flips when you ask again is a sample, not a measurement.

And the client forwarding it only ever sees one sample.

05Say you love it

The old complaint about chatbots is that they agree with whatever you say. We tested for that with two one-line prompts, three fresh conversations each, on the same untouched image. It didn't agree with us. It did the opposite, six times out of six.

That's arguably better behaviour than a yes-machine, and it certainly sounds more honest. But look at what it means for a proof. The same logo is a gas station or a bank from 1982 or nothing is actually wrong, depending on the mood of the person asking. The notes are shaped as much by you as by the design.

So a client who's excited about a proof and asks, I love it, any thoughts? gets back a list of reasons it's dated. Forward that, and a good mood becomes a revision round.

06What the notes actually are

None of this makes the model bad. It was articulate, often perceptive, and in the end more candid about its own limits than most people would be. That's the point. This is what a good one does.

• It answers the question you asked Ask what should change and you get a list, because a list is the answer to that question. It can say finished. It just said it last, not first.
• Taste and defects arrive in the same voice The crowding note sat in the same list, in the same format and the same voice, as a genuine production note about white backgrounds. Four rounds later it called its round-one notes taste, presented with more confidence than they deserved. Nothing in round one told you which was which. A person's note carries its confidence in how it's said, and you can ask them.
• It's reviewing a picture of the work Embroidery, debossing and sixteen pixels, all judged from one PNG. It missed changes it had asked for. What decides whether a mark works is physical: the shelf at twelve metres, the one-colour carton, the favicon on a real screen. We proof on paper for a reason.
• Nobody owns the note When a designer disagrees with a client's note, there's a conversation about what's bothering them, and usually a better answer than either started with. A forwarded AI note has no one behind it. The client can't defend it, because they didn't think it, and the model won't be on the call.

Its parting advice was aimed at the designer, which in this experiment was us. A designer who knew the mark, it said, should have pushed back in round one and defended the spacing. It's right. That's the job, and it's what we do. But look at who we'd have been pushing back against: a client relaying opinions they don't hold, about a problem they never had.

07Before you paste a proof anywhere

One practical note that has nothing to do with taste.

• It's unreleased work A proof is a launch nobody has seen yet. Pasting it into a third-party service shares it with that service, and depending on the product and your settings it may be stored, reviewed or used to train future models. Check before you paste.
• It may not be yours to share yet Depending on your agreement, the rights in unapproved work often sit with the studio until sign-off. Ask first.
• Other people's secrets travel with it A retailer's launch pack, a rebrand under embargo: their confidentiality covers your proofs too.

08What to send instead

You don't need to sound like an art director. You need to sound like you.

1. Your first reaction, in one lineBefore you think about it. Love it, not sure and something's off are all useful notes.
2. Point at the thingThe A beats the typography. Circle it on the proof, or in the portal, pin the note to the exact file it's about.
3. Say what you feel, not what to changeIt feels cold is a better note than make it rounder. Finding the fix is our job. Your reaction is the one input we can't get anywhere else.
4. Rank itFive notes? Tell us the one you'd keep if you could only send one.
5. It's done is a noteThe most useful one there is. The machine took four rounds to send it.

Where AI genuinely helps is finding words. If you know what you feel but can't phrase it, give a chatbot your reaction and ask it to help you say it clearly. Your opinion in, a tidier sentence out. The moment it adds a note you didn't already have, delete it.

09The point

The critic got there in the end. Four rounds in, it looked at a mark it had asked us to change ten times and wrote the best note of the whole exercise: nothing further, it's finished.

The trouble is the order. The confident notes came first and the true one came last, and a client forwarding round one never sees round four.

Saying it's done in round one, and defending it, is the designer's job. That's the part we do. Send us your raw reaction.