A scary text deserves a plain answer before anyone taps the link.
ScamScanner is my concept for an AI helper that reads a suspicious text, email, flyer, or voicemail and tells people in plain words how risky it looks and why. I designed it so the AI suggests and the person decides, with a risk score, the reasons behind the score, an honest "I'm not sure," and a way to push back.
Client
ScamScanner, my own concept
My role
Designer and researcher, from concept to screens
Focus
Plain‑language scam checks for people at risk
Result
A trust‑first concept, tested with a generative agent
Scams got better at sounding real, and help stayed hard to reach.
Scams arrive as texts, emails, robocalls, voicemails, and even paper flyers. They borrow a bank's name, set a deadline, and push for action now. The people they hurt most often have the least technical help nearby: older adults, teens, and busy parents reading their phones between everything else.
My fraud work keeps teaching me the same lesson. When the path to help is confusing or inaccessible, the harm compounds for the people who are already vulnerable. A person who cannot tell whether a message is real may pay the scammer or may freeze and ignore the real bank as well.
For my capstone in Stanford's design for AI products course in 2025, I asked a narrow question. What would a scam check look like if it were designed for the person who is scared, unsure, and holding a phone?
Four trust decisions shaped every screen.
An AI that sounds certain when it is guessing can do real harm, so each of these decisions limits what the AI is allowed to do. Click or tap a card to turn it over.
People can check a message in the form it reached them.
The home screen offers three large choices: Snap a Pic, Paste a Message, and Share Recording. A flyer on a windshield, a text from an unknown number, and a voicemail all get the same plain check.
I also designed a chat version of the same helper for vague or emotional questions, such as "I got a weird text from the IRS. What should I do?" The chat answers with reassurance, asks follow-up questions, and offers a hand-off to a human advisor or emergency services when the stakes are high.
- Snap a Pic checks flyers, letters, and screenshots.
- Paste a Message checks texts and emails.
- Share Recording checks voicemails and robocalls.
- The chat helper takes questions in the person's own words.
- A person can always ask a human for a second opinion.
I reviewed the risks of a scam-checking AI and planned a guard for each one.
A helper built for vulnerable people has to be checked for the ways the helper itself could hurt them. Click or tap a card to turn it over.
Check a sample message and see how ScamScanner explains its answer.
This prototype rebuilds my three screens as a working flow in a fresh, current look. The scan results are scripted with sample messages, and no real AI runs in this prototype.
Things to try
- Choose Paste a Message, then pick a sample or paste your own text.
- Read the Scam Risk Score and the reasons under Why we flagged this.
- Try Share Recording and the new number sample to see how ScamScanner admits doubt.
- Choose Was this right? and override a result you disagree with.
A dark midnight and violet palette keeps the answer in front.
These are the colors, type, and parts the prototype uses. I kept the midnight, slate, violet, and lavender family from my original ScamScanner screens and rebuilt it as a dark, square-cornered app that follows Apple's dark mode layering: a Midnight base, lighter grouped surfaces, and clear separators, with every text color checked for WCAG AA contrast.
Color
- Midnight
#0B0C1CThe base background behind every screen and the dark end of the street light glow. - Slate
#1F2240Grouped list cells, reason panels, sample messages, and the tip, one clear step lighter than Midnight. - Slate hairline
#3D4168The separators between rows and the edges of cards, fields, and the tab bar. - Violet
#6E4FF6The solid avatar circle, the square icon tiles, the start of the score dial, and the outer glow in the greeting. - Violet mist
#A796FFThe middle of the dial and button gradients, step numbers, and the scanning ring. - Lavender
#DCD4FFThe bright core of the greeting's street light, the selected tab, back links, the primary button, and the focus outline. - Mist text
#CDCEE3Secondary text such as descriptions and section headers, at 9.9:1 contrast on Slate. - High risk red
#EF8A80Only the High risk chip, the red flag icons, and the marker on the score dial. - Unsure amber
#E2B565Only the Not sure chip, its question mark icons, and the dial marker when context is limited. - Low risk green
#6FC99COnly the Low risk chip, its check icons, and the dial marker when a message looks normal.
Type
Screen headlines
Got a message that feels off?
SF Pro Display, Inter fallback
Row and panel titles
Paste a Message
SF Pro Text, Inter fallback
Body text
Do not tap the link or reply.
SF Pro Text, Inter fallback
Section headers
Check it the way you got it
SF Pro Text, Inter fallback
Risk score
83%
SF Pro Display, Inter fallback
Components
Large title that collapses
Each screen opens with a large title. As you scroll, the title moves away and a compact bar with a small title and a frosted background stays on top, so the content gets most of the screen.
Street light greeting
The home greeting sits under one soft point of lavender light high on the right that fades into Midnight, like a street light at night, so the dark screen feels safe instead of empty.
Grouped way-in list
The three ways in sit in one inset grouped list with a solid violet icon tile, a plain label, the kinds of messages each accepts, and a chevron, separated by visible hairlines.
Scam Risk Score dial
A circular gauge fills with the violet to lavender gradient up to the score, and a small marker in the risk color sits at its end, while a dashed arc and a question mark show when ScamScanner does not have enough context.
Risk chip and verdict
A small square chip pairs an icon with words such as High risk, so the meaning never depends on color, and one plain sentence sits under it.
Reason sections and You decide banner
Grouped sections hold why the message was flagged, the red flags, similar reported scams, and numbered next steps, and a banner with a gradient side bar confirms that nothing was blocked, deleted, or reported.
Buttons, feedback, and tab bar
The primary button uses the lavender to violet mist gradient with Midnight text, the feedback tiles and checkmark list let the person override the result, and the tab bar marks the current tab in Lavender.
Principles
- Every score comes with its reasons in everyday words.
- Color never carries a meaning by itself, so every verdict is also written out.
- The screen tells people what ScamScanner did not do, so nobody wonders whether a message was blocked or deleted.
- When ScamScanner is unsure, the screen says so and offers a person to ask.
The trust decisions held up when I put the helper to the test.
ScamScanner is concept work, so there are no launch numbers. The results are the trust decisions above and what I learned when I ran the chat helper as a generative agent against a sample persona. Click or tap a card to turn it over.
The same trust rules guide my AI-native product research and design.
I bring these decisions to teams that are putting AI into fraud, payments, and safety tools. The AI suggests and the person decides, the reasons are visible, uncertainty is stated, and people have a way to say the AI got it wrong. If your team is putting AI in front of people at a frightening moment, I would be glad to review that flow with you.
There is plenty more where this came from.
Names, brands, and data in my case studies are changed to protect my clients.
Tell me about your project.
Tell me a little about it on my contact form, and a real person will get back to you soon.