The machine that translates and also admits it's guessing

A monospaced transcript, timestamped like a log file. SHE SAID [UNCLEAR] AND THEN LAUGHED. I THINK IT WAS A JOKE. [CONFIDENCE 41%] beside a scarlet query mark. ASK HER AGAIN.

The machine got most of the sentence. It is telling you, in exact numbers, the part it didn't.

Why print a confidence percentage instead of just the mistranslation?

Because the percentage is the more honest object, and the design is built to notice that.

A wrong translation is a joke waiting to happen — funny in the way a subtitle blooper is funny, at the machine's expense, over in one beat. A number is a different kind of statement entirely. 41% doesn't claim to have understood; it reports exactly how far short it fell, and it does that in the same breath as a sentence about someone laughing at something. The gap between those two things — a person's laugh, a machine's arithmetic about whether it caught the reason for it — is where this print actually lives. It isn't mocking the technology. It's sitting with what the technology can't do yet, stated in the technology's own units — a number carries a different kind of weight on a shirt than a sentence does, and this design asks it to carry sadness instead of a boast.

What is [UNCLEAR] actually recording?

A moment the earpiece heard and couldn't resolve, logged rather than papered over.

Most systems that fail quietly just guess and move on, and the guess becomes the record — you'd never know it happened. This transcript does the opposite: it marks the gap in the text itself, inline, at the exact word it lost. That's a design choice with a real cost. A transcript full of brackets looks less finished than one that's smoothed every gap into a confident, wrong sentence. The print keeps the brackets on purpose, because a record that admits its own limits is more useful, and more moving, than one that hides them.

Why the machine tells its own user to ask again

ASK HER AGAIN is the earpiece talking to the person wearing it, not to her. It's an instruction the device gives because it has run out of what it can do alone — it heard the words, it couldn't be sure of the laugh, and the only remaining move is a human one: turn back and ask. That's a narrower kind of humility than most machines are built to show. It doesn't apologise. It just tells you the next step is yours.

Who is this design actually for?

Anyone who has sat inside a conversation they could only half follow and wanted the half they missed written down honestly.

A parent's language you learned as a child and let go slack. A grandparent's dialect that a phone app gets three words in five. A call in a language you're still learning, where you caught the shape of the sentence and lost the joke inside it. The earpiece in this print is invented, but the experience it's translating — hearing enough to know something funny just happened, and not enough to know what — is not a stretch for anyone who has lived between two languages, even briefly.

Does a design like this only work for something sad?

No — it holds up just as well for a call that ended fine.

The transcript doesn't say the conversation failed. It says one line of it stayed unresolved, and the last instruction on the shirt is forward-looking: ask again, not give up. Plenty of people have a specific exchange like this — a joke half-caught mid-flight, a punchline they never got back — that they remember fondly rather than with regret. The design carries either mood without needing to be rewritten, because the confidence figure is doing the same work in both readings: reporting, not apologising.

What about rights?

The transcript is invented, and there's nothing here that borrows a real device, a real translation product's interface, or a real person's actual words.

The only real content a version of this design might carry is an actual exchange somebody remembers — their own conversation, their own missed joke — which belongs to them by default, the way any personal memory does when you're the one putting it on a shirt.

Does it print well?

Yes — a monospaced transcript is close to the easiest thing this material ever asks a print shop to reproduce, because it's built entirely from flat type on a flat ground.

The scarlet query mark is the one accent worth keeping as a single hard-edged colour rather than softening it into the body text's grey — it's doing the job of a highlighter, and flat colour with a hard edge is what actually survives the weave, where a soft red would just blur into pink. Keep the bracketed system text — [UNCLEAR], [CONFIDENCE 41%] — in the same weight as the spoken lines rather than shrinking it further; it's already doing enough work by existing in brackets at all.

Off-white, pale grey or a soft stone all read as a printout rather than a costume. Two inks cover it comfortably: the body type in a dark neutral, the query mark and percentage in scarlet.

How do I get one made?

Describe the exchange at JustOG — what was said, what the machine caught, and what it didn't. Pick a direction, drag the crop frame, see it composited on the real garment, and it's made to order and shipped.

Designs other people have published are in the shop.

Forty-one per cent of a sentence, and somebody laughing anyway.