Back to Explore
GuideVibeFix editorialUpdated Nov 16, 2026

Say It Was Made by AI: Labeling AI-Generated Content in Your Vibe Project

If your product shows something AI made, say so yourself — before users have to guess. This guide covers the three labeling layers, what a good label looks like, mixed human-AI content, and keeping the label attached when content is shared. All example projects, people, and numbers in this guide are fictional.

A hand holds a smartphone showing an AI image generator app with a prompt about friends on a beach at sunset, six style buttons, and a generated group photo below

Chen Mo used "Lightcatcher Album" to turn a cloudy family photo into a "sunset beach" version with one tap: evening glow on the horizon, warmer sand, the four of them smiling by the sea. Pleased with the result, he dropped it into the family group chat. His grandmother was delighted, and the next day she called to ask: "When did you go to the beach? Why didn't you take me along?" Only then did Chen Mo realize nobody in the group could tell the photo was AI-edited — at first glance, even he had almost believed they had really been to the beach. He hurried to explain in the group that it was an AI edit, not a real trip. His grandmother half believed him, and his cousin asked, "So did you go or not?" A warm little share turned into a family-wide correction.

Two things to make clear first. First, the walkthrough project "Lightcatcher Album", the operator "A-Guang", the user "Chen Mo", and every time and number in this guide are fictional, invented only to make the process concrete. Second, the boundary: this site has already published several AI guides — "Don't Let AI Sound Confident While Making Things Up: Hallucination UX for Vibe Products" is about what to do when AI states wrong facts, a question of whether the content is true; "Generative UI Design" is about interfaces generated by AI in real time, a question of how the interface looks; the watermark and anti-scraping guide is about stopping other people from stealing content. This guide repeats none of them. It covers one thing only: when your product shows something AI generated, how to tell users plainly that it was made by AI. Hallucination defenses govern truth; labeling governs whether you said so clearly. They are different jobs.

A hand holds a smartphone showing an AI image generator app with a prompt about friends on a beach at sunset, six style buttons, and a generated group photo below

1. First, the reality: users cannot tell, and that is not their fault

Many teams building AI features carry a quiet assumption: users can tell at a glance that something is AI-generated. The label says "AI generated" right under it — who could miss that? In reality, users do not read every small word on the screen. They scroll fast, spending a second or two on an image, and their brains default to "a photo is a photo, text is written by a person." AI output is now good enough to look real: the light is right, the expressions are natural, the writing reads smoothly. Unless you go looking for it, you cannot tell. Chen Mo's "sunset beach" is the proof — even the person who made it almost believed it at first glance. Expecting his grandmother to spot it instantly was never realistic.

The fictional numbers in this guide are: of 100 users who saw an AI-generated image, 78 noticed the small "AI generated" label in the corner, 61 understood from the label that the image was made by AI, 44 felt the app was "at least honest" because of it, and 29 kept the label attached when they shared the image. These numbers are fictional, and they make one point: labeling is not a step you tick off once and forget. It is a funnel, and people drop out at every step. Only 78 out of 100 noticed the label, so the label first has to solve "being seen"; 61 understood what it meant, so it also has to solve "being understood"; only 29 shared it with the label intact, so it also has to solve "traveling with the content." Skimp on any step and the funnel leaks right there.

Here is another fictional comparison: without labels, only 12 out of 100 users could tell which images were AI-generated; with clear labels, 71 out of 100 could. The numbers are fictional — the point is not 12 and 71 themselves, but that "cannot tell" is the normal state and "can tell" only happens because the product spoke up. When users cannot recognize AI content, they are not slow. The product simply did not say so. Hide the fact, whisper it, or phrase it vaguely, and users are left guessing; when they guess wrong, they are the ones embarrassed, and the product is the one that loses trust. The first rule of labeling is that simple: when something in your product is AI-generated, the product should say so first — not wait for users to ask, guess, or correct the record afterward.

2. Where to label: three layers, none optional

The most common labeling mistake is doing only one layer: a small badge in the corner of the image, or one line of small print on the detail page, or a marker buried in the file metadata. One layer fails because users do not always reach that layer. People scrolling the feed never see the detail page's small print; people reading the detail page never check file metadata; people who download the image may have the corner badge cropped off. All three layers are needed, and each covers a different situation.

The first layer lives on the content itself: the label sits right next to what it describes. In the corner of an image, in the first seconds of a video, at the start or end of a piece of text — users should see the label at the same moment they see the content. The further the label sits from the content, the more likely it is missed; a label at the bottom of the page or hidden under a "More" menu is the same as no label. The second layer is the detail page: one sentence that says it plainly — "This image was generated by AI, based on a photo you uploaded" — plus the original for comparison, one tap away. The detail note does not need to be long; three lines covering "who generated it, based on what, and what changed" is enough. The third layer is metadata: a machine-readable marker inside the file recording the tool, the time, and how much was changed. Users never see this layer, but platforms, search engines, and other tools can read it — and after the content is forwarded, downloaded, or reprocessed, it is the only evidence still traveling with the file.

Remember what each layer is for: the first is for people scrolling past, the second is for people who want to check, the third is for machines and platforms. Do only the first, and the label is lost the moment someone screenshots and forwards the image. Do only the second, and everyone scrolling the feed never sees it. Do only the third, and you have told the truth to machines and said nothing to people. Only all three together make the label hold. If Chen Mo's photo had all three — a badge in the corner in the group chat, a detail page saying "AI edit based on a cloudy original" with the original beside it, and a marker inside the file — his grandmother would never have thought they had really been to the beach.

3. What the label looks like: visible, readable, and not dismissible

With placement settled, the label itself is next. Of the three label styles you see most often, two are wrong. The first is "too small to see": tiny gray text in a corner, smaller than a watermark, blending into the background, findable only when the designer zooms in three times. The second is "written like a riddle": "assisted by intelligent technology," "algorithmically enhanced," "contains synthetic elements" — after reading it, users still do not know whether the thing was made by AI or touched up with a filter. Only the third style is right: visible, readable, and not dismissible.

Visible has three hard standards. First, contrast: the label background and text must stand out clearly against the content — a light label on a dark image, a dark label on a light image, never semi-transparent gray text on a busy background. Second, a fixed position: every AI-generated item uses the same spot — always top-left, or always bottom-right, never top-left for one image and bottom-right for the next. Once the position is fixed, users learn where to look after seeing it twice. Third, a minimum size: the label must be legible on a phone screen without zooming. The fictional number in this guide is that the label should be noticed within 1 second of the user seeing the content. The number is fictional; the point is "no searching required" — a label users have to hunt for is a label that does not exist.

Readable also has three hard standards. First, plain words: write "AI generated" — four words everyone understands — not "intelligently synthesized" or "algorithmically enhanced," phrases that need translating. Second, state the degree: fully AI-made is "AI generated"; human-shot and AI-edited is "AI edited"; human-written and AI-polished is "AI assisted." Different degrees get different wording — do not call everything "AI generated," and do not blur everything into "AI involved." Third, offer a way in: the label should be tappable, opening a one-sentence explanation and the original for comparison. Users may never tap it, but "I could tap and understand" is itself part of trust.

Not dismissible is the rule most often "optimized" away. Some teams, wanting a cleaner interface, add a little X to the label so users can close it; others hide labels by default in a "minimal mode." Once a label can be closed, hiding it gains a respectable excuse: a user closes it once, and every AI item after that ships without a label; a product ships one "minimal mode," and labeling is gone. Write the rule down in stone: an AI label is part of the content, not interface decoration — users cannot close it, and the product cannot hide it by default in any mode. The interface can reserve space for the label and match its colors, but it cannot offer it a way to disappear.

4. How much human editing makes something "human-made"?

The hardest line to draw is mixed content: a person took the photo and AI swapped the sky — is that AI-generated or human-taken? A person wrote a draft and AI rewrote it three times — human-written or AI-written? Without a line, users and products argue past each other forever. The way to draw it is not to count "what percentage AI changed," but to ask three questions. The answers decide the label.

First: would AI's changes alter a user's judgment about the facts? Swapping in a sunset changes whether "they went to the beach that day" is true — label it "AI edited" and show the original beside it on the detail page. Adjusting brightness or smoothing skin changes no factual judgment; leave it unlabeled, or mention it on the detail page. Second: would this thing exist at all without AI? An article written start to finish by AI gets "AI generated"; a human outline expanded by AI gets "AI assisted"; a human draft with only typos fixed by AI gets "AI proofread" or no label. Third: will users act on this content? A resume photo, a product shot, a news image, a health explainer — users make decisions from these, so label them one notch stricter. A chat sticker, a decorative avatar, an inspiration sketch — users will not take them as fact, so the label can be lighter, but it cannot be absent.

There is also a backstop rule: when in doubt, label stricter. Labeling strictly costs the user four extra words at most; labeling lightly costs a trust incident when someone is misled. Many teams worry that strict labels make everything feel "too AI," hurting the feel of the content — that worry is misplaced. What users resent is not "this was made by AI," but "this was made by AI and you didn't tell me." If Chen Mo's photo had been labeled "AI edited" with the cloudy original beside it, his grandmother would have seen not "my grandson lied to me" but "my grandson made the photo look lovely with AI." Same photo, labeled or not, and the mood at home is completely different.

One more line to draw in advance: do not label content your users upload on their own say-so. A user uploads a painting they drew themselves, and the product's detector flags it as "probably AI-generated" and slaps a label on — the user feels wrongly accused. Detection can prompt, never convict: if the system thinks an image looks AI-generated, it can ask, "This looks AI-generated — want to label it?" and let the user decide. If the user says no, do not label it. A wrong label hurts just as much as a missing one; only the victim changes, from the person viewing the content to the person who made it.

5. After sharing: the label has to travel with the content

The easiest place to lose a label is sharing. Inside your own app the label is crisp and clear; the user taps share, the image lands in the camera roll and goes to another platform, the corner badge gets cropped off, the detail-page note never follows, the metadata gets stripped by a compression tool — the label is gone the moment the content leaves home. The funnel's last step, "29 people shared with the label intact," is where those people drop out. The rule for sharing is one sentence: the label is part of the content, and what gets shared must be the labeled version.

Three things make that true. First, burn the label into the shared image: draw "AI generated" into the pixels themselves, instead of floating it as an overlay that only works inside your own app. A floating label works at home; a burned-in label travels wherever the image goes. Second, make shared links carry the detail page: when someone taps the link, they should land on a page with the label, the original for comparison, and the one-sentence explanation — not on a bare image. Third, do not let your own compression pipeline strip the metadata: exporting, transcoding, and compressing must each be checked to confirm the marker survives. Many teams do not lose metadata to other people; they lose it to their own image pipeline. Before launch, run one marked image through the full export flow and check whether the marker is still there. That is far cheaper than fixing it afterward.

There is one more kind of sharing that is easy to miss: screenshots. A screenshot drops the metadata and the detail-page link, leaving only the picture itself. What saves the day then is the burned-in label and the fixed position: the label sits top-left, and people rarely crop off the top-left corner on purpose; big and clear enough, it stays readable in a screenshot. The goal is not to make screenshots impossible to strip — it is to make a casual screenshot carry the label anyway. If someone is determined to crop the label out and repost, that is their choice; the product has done its duty to inform. What the product controls is "labeled by default," not "impossible to remove."

6. Four clever ways to hide a label — don't

Once labeling has been around a while, someone always tries to have it both ways — compliant and pretty — and invents clever ways to hide the label. The first is "hide it on the detail page": the feed looks spotless, and only the detail page carries one line of small print. Users skip the detail page nine times out of ten, which means nine times out of ten there is no label. The second is "write it as a riddle": "intelligently enhanced" or "algorithmically created" instead of "AI generated." Users finish reading none the wiser, while the product can claim "we labeled it." The third is "label only half": fully AI-made items get labeled, heavily AI-edited ones do not; text gets labeled, images do not; new content gets labeled, old content never gets backfilled. The fourth is "time-limited labeling": version one ships with labels, a redesign finds them ugly, and they quietly disappear, justified with "users are used to it by now."

All four share the same bet: that users will not notice, will not push back, and nothing will go wrong. Win the bet and the interface looks a little cleaner; lose it and you have a trust incident. The fictional numbers in this guide are: without clear labels, complaints about "fake photos and fake content" averaged 23 a month; after the three-layer labeling was done properly and the labels written in plain words, they fell to 4 a month. The numbers are fictional — the point is that complaint volume is the most honest measure of labeling quality. Label clearly and complaints fall; complaints rise, and the first thing to check is the labeling, not the users for being "too picky."

There is a simple test for whether labeling passes. Hand the product to an older relative who knows nothing about AI, let them use it for half an hour, then ask three questions: of the things you just saw, which ones were made by AI? How do you know? If you sent one to a friend, how would you describe it? If they cannot answer, the labeling fails. The test costs nothing — no user recruitment, no survey, just a relative at home. The reason many teams' labels keep getting vaguer is that the people testing them already know the answers.

7. Pre-launch checklist

Go through this list item by item before shipping a product with AI generation features. First, all three layers are in place: a visible label on the content, a one-sentence explanation plus the original for comparison on the detail page, and a machine-readable marker in the file metadata — and all three agree, with no "AI generated" on the content and "AI assisted" on the detail page. Second, the label itself passes: it says "AI generated," "AI edited," or "AI assisted" in plain words, not riddles; the position is the same across the whole product; dark images get light labels and light images get dark labels; it is legible on a phone without zooming; users cannot close it, and no mode hides it by default. Third, mixed content has a written rule: did it change the facts, would it exist without AI, will users act on it — ask the three questions before choosing the label; when in doubt, label stricter; user-uploaded content gets a prompt, never an automatic verdict. Fourth, sharing carries the label: what leaves the app is the labeled version, the label is burned into the pixels, shared links open the detail page, and the export pipeline does not strip metadata; a casual screenshot still shows the label. Fifth, complaints have an owner: look once a month at how many "fake photo, fake content" complaints arrived (in the fictional example, from 23 down to 4), and if the number climbs, check the labeling first; spot-check 30 AI-generated items a week to confirm the labels are present and worded right. Sixth, the older-relative test passes: hand it to someone who knows nothing about AI for half an hour, and only ship when they can answer the three questions.

And six pitfalls to repeat once more. First, do not bet that users "can tell by themselves": without labels only 12 out of 100 could tell (fictional number) — not telling is the normal state. Second, do not do only one layer of labeling: feed, detail page, metadata — miss one and the label vanishes in some situation. Third, do not write the label as a riddle: "AI generated" in plain words is the clearest wording there is. Fourth, do not leave the label a "close" button: a label that can be closed will eventually be closed. Fifth, do not label your users' own uploads for them: when detection suspects AI generation, ask the user instead of passing judgment. Sixth, do not label only on launch day: old content needs backfilling, redesigns must not quietly drop labels — labeling is ongoing maintenance, not one line on a launch checklist.

Labeling funnel (fictional sample): 100 saw an AI-generated image, 78 noticed the corner label, 61 understood it was AI-made, 44 felt the app was honest, 29 shared it with the label

Back to Chen Mo. He still edits photos in "Lightcatcher Album", and the little "AI edited" label still sits in the top-left corner of every image. Tapping it shows the cloudy original and one sentence: "Based on the photo you uploaded, AI replaced the sky and the light." He sends the "sunset beach" version to the family group, and his grandmother spots the label at first glance and replies with a smile: "It looks lovely — take me along next time you really go to the beach." His cousin taps through to the original and says the cloudy version has its own charm. Same photo: labeled clearly, it is a warm little share; left unlabeled, it is a family-wide correction. That is what labeling AI-generated content does — it lets anything AI made be recognizable as AI-made at a glance, so users never have to guess, ask, or explain afterward. The product says it first, and trust has somewhere to land.

Views 5Comments 0

Comments (0)

ME
0/1000
Loading comments...
Browse projectsPublish your project

Related articles

Real photo: a laptop, tablet, and phone on a wooden table showing the same green-leaf website, the phone view with navigation in one column, a potted plant nearby
Guide
Squashed on phones? Responsive layout and media queries, explained plainly

One page should stand up on a phone, a tablet, and a computer. This guide follows MDN's Responsive web design, Using media queries, and the viewport meta reference: fluid HTML, the viewport tag, fluid grids and max-width, images that never burst their container, media query syntax, breakpoints set by content, mobile first, responsive type, testing capability instead of width, and a pre-launch checklist.

Mobile AppsDesign ExperienceFrontend Engineering
Real photo: a bank of light gray metal filing cabinets with one drawer pulled open, revealing rows of hanging file folders inside
Guide
A real database inside the browser: IndexedDB local storage, explained plainly

localStorage only stores strings and has no indexes. IndexedDB is a real database in the browser: structured data, far larger capacity, indexes, and transactions. Following MDN's Using IndexedDB, IDBDatabase, and IDBObjectStore: the five-step routine, opening, upgrades, keys, indexes, transactions, add versus put, get, cursors, ranges, and multi-tab upgrades.

Tool TipsData WorkflowsFrontend Engineering
Real photo: hands typing on a MacBook Pro showing JavaScript code that changes HTML content, with a white mouse, papers, and binders on a white desk
Guide
Functions without the confusion: JavaScript functions, parameters, and arrow functions, explained plainly

You already call functions every day; this guide names what you have been doing. Following MDN's Functions, Arrow function expressions, and Rest parameters: declarations versus expressions, parameters and arguments, default values, rest parameters, spread syntax, return values, arrow function shorthand, the this trap, callbacks, and scope.

Tool TipsFrontend EngineeringAI Coding