What visual patterns signal that content was AI-generated?

Aldermarsh Community College's bridge-year writing program uses Penmark, an AI tool that drafts margin comments on student essays for instructors to review before sending. Thandiwe Marechera is the program's Writing Director. She has taught composition for fourteen years.

How to answer this, stage by stage

Nobody is grading whether you can name five UI patterns off the top of your head. They're grading whether the pattern you pick can survive someone trying to lean on it wrongly.

Let's learn

Here's what a real classroom looks like once an AI drafting tool has been running quietly for a semester.

Before Penmark, Thandiwe hand-wrote margin comments on roughly 150 essays a semester, about 12 minutes each, close to 30 hours of writing feedback by hand. Penmark drafts a first-pass comment in under 10 seconds. Once she started reviewing and editing those drafts instead of writing from scratch, her time per essay fell to about 4 minutes.

Five ways a screen can show its work. Penmark, at launch, used none of them.

Here's the turn: the extra minutes Penmark saved were never the real story. The real problem showed up the moment an unread AI draft and a hand-rewritten comment started rendering in the exact same box, same font, same weight, with nothing on the page to tell them apart.

A pencil draft and an inked page hold the same words sometimes. Only one of them should look finished.

At its worst, a comment box that can't tell a reader whether a person looked at it doesn't just waste a parent's trust once. It quietly tells every student in the program that the personal attention they were promised might not have happened for any of them, and there's no way to check.

What I would leave alone: a spelling fix or a comma correction doesn't need this same tagging treatment. Nobody assumes a red squiggly underline required a person's judgment, and tagging every small mechanical fix would just teach readers to stop noticing tags where they actually matter.

The lesson: a comment box that looks the same whether a person or a model made it isn't simpler. It's a hidden decision about who gets the credit, made without anyone choosing it on purpose.

Now here is the same thing as a story

The short version above is what you'd say defending this design to Aldermarsh's department chair. Read this one for how close the actual near miss came.

For fourteen years, Thandiwe's margin comments were the thing students kept. Not graded feedback, real sentences, written in her own hand in the gap beside a paragraph that needed one. She could tell a genuinely confused essay from a merely lazy one by the second page.

Penmark arrived in August, and the first semester was good. She'd open a draft comment, read it fully, rewrite half of it in her own words, and send it. The tool saved her real time. Nothing about the page looked any different than it always had.

By October, she was reading every draft fully. By January, she skimmed the ones that looked reasonable at a glance. By March, she'd taken to approving a batch of ten or twelve at a time between classes, trusting that Penmark's drafts were, by now, usually fine.

Both panels used to render as one identical box. The redesign is what finally tells them apart.

The trigger was small. A student teacher named Mateo, shadowing her for the week, pulled up two essays side by side and asked which comment was hers and which was Penmark's draft. She looked for a long moment and admitted she genuinely couldn't say without opening her own memory of writing it.

That evening she pulled the last sixty comments she'd sent and checked each one against her own edit history. Forty of them had gone out with zero changes from Penmark's first draft, unread in any real sense, sitting in the exact same box as comments she'd rewritten from scratch.

Six good months hid a habit that had already changed. The question in week 25 is what made it visible.

Aldermarsh didn't pull Penmark. They rebuilt the comment box: a small amber tag for anything Penmark drafted and Thandiwe hadn't yet touched, switching to full color the moment she made a real edit, with a one-click diff showing exactly what changed. Batch-approving without opening a comment now leaves it visibly amber, forever, until someone actually looks.

SPARK, in one screenNot a UI style guide. SPARK is what decides which signal earns a place on the page at all.

Four small facts turn a vague tag into something a reader can actually check.

Three states, three signals. The branch on the right is the one worth protecting: no tag where none is owed.

The recap, one line per letter: situation is Thandiwe writing every comment by hand, payoff is students trusting that a marked comment reflects real judgment, anchor is the persistent amber-to-full-color state tag with an openable diff, risk is a batch-approved comment staying visibly unreviewed instead of blending in, and keep out is leaving mechanical fixes untagged so the tag still means something where it matters.

And if you want to be sure it really works, try it somewhere elseSame five letters, a logistics company's customer support inbox instead of a classroom. A different anchor risk breaks the second story.

Ferro Logistics runs customer support for freight delay claims. Kenji Osei, Support Operations Lead, oversees a team of twelve agents using ReplyDraft, an AI tool that drafts replies to delay complaints for an agent to send or edit. Mapped onto SPARK: situation is agents typing every reply by hand, about 6 minutes each, across roughly 300 tickets a day; payoff is customers trusting that a reply actually addresses their specific shipment, not a template; anchor is the same persistent state tag idea, an amber marker on any reply sent without an agent edit, full color once an agent has changed at least one sentence.

The risk here is different from Thandiwe's. It isn't a teacher losing track of her own hand. It's a customer, already annoyed about a late shipment, receiving a reply that reads as generic and assuming nobody read their complaint at all, whether or not that's true. The old decision being taken back: ReplyDraft's replies were designed to match the company's support voice exactly, on purpose, so customers wouldn't feel like they were talking to a bot. That made sense when reply volume was low enough that agents edited nearly everything. It stopped making sense once volume tripled and unedited sends became the majority.

The same five signals apply to a support reply as to a margin comment. Only the stakes changed.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "a persistent state tag, amber until a person edits it, full color after, with a diff behind it," and stop.
Cost: there's no engineering time this quarter for a diff view. Say so honestly, and ship the two-state color tag alone first, since even a plain amber-or-not signal beats no signal at all.
The model gets better, for real: if Penmark's drafts get good enough that teachers barely need to edit them, that's still not a reason to remove the tag. It's a reason to watch the unreviewed-and-sent rate even more closely, since a better model makes it easier to stop looking at all.

Where people run it wrong.
They ship a one-time watermark that says "AI-assisted" and call the trust problem solved.
They tag everything, including mechanical fixes, until readers learn to ignore every tag on the page.
They let the tag's color depend on the model's own confidence score instead of on whether a person actually made a change.

How to use it live. When someone asks what visual pattern signals AI-generated content, ask yourself first: does this signal answer "what made it" or "has anyone looked yet"? Only the second question is the one worth designing for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0