Is This Text a Scam? Check It With the AI on Your Phone

The text says your package can’t be delivered and you have 24 hours to confirm your address. Maybe you are waiting on a package. Maybe you aren’t. Your thumb is already hovering over the link, and that’s exactly what the sender is counting on.

There’s a better move that takes about a minute and doesn’t involve tapping anything. You show the message to an AI that can read it and tell you what’s suspicious about it. Google built that into Android’s Circle to Search and into Google Lens (which works on iPhone too). ChatGPT, Gemini and Claude can do a version of it if you paste the text in. All of it is free.

People are already doing this at scale. OpenAI said in its October 2025 threat report that it sees people use ChatGPT to identify and avoid online scams “millions of times a month,” and estimated ChatGPT is used to spot scams up to three times as often as it’s used to run them. This guide covers how to do it safely: which route to use on your phone, what to paste (and what to leave out), three worked examples, where the answer can’t be trusted, and the steps that matter more than the AI’s opinion, which are never clicking, checking through an official app, and reporting the message.

A note on how I tested. I don’t have a bench of Android and iPhone test devices, so the Circle to Search and Lens steps below come from Google’s own documentation and published hands-on tests, and I’ve linked them. The chatbot examples come from running four made-up texts through Claude with the prompt in this post. The sample outputs are shortened from what it returned. I haven’t compared the three chatbots head to head.

Why this question is everywhere right now

Texts are the scammers’ favorite door, and the money is real. The FTC’s data spotlight on text scams, which uses 2024 reports, puts reported losses at $470 million, more than five times the 2020 figure, even though the number of reports fell. The FTC also notes that the vast majority of fraud is never reported, so that figure likely reflects only a fraction of the harm. For all fraud in 2025, the FTC told Congress it received 3 million reports with $15.9 billion in reported losses.

Source: FTC Data Spotlight, “Top text scams of 2024” (April 14, 2025). Figures are based on fraud reports that name text as the contact method. These are 2024 numbers, the most recent text-specific breakdown I could verify.

Fewer reports and more money lost means the scams are getting better at the one thing that matters to the scammer: getting a person to actually pay. The FTC footnote adds a detail worth knowing. The share of text-scam reports that said money was lost went from 5% in 2020 to 11% in 2024. Most people who report a scam text didn’t lose anything. But that rate is rising.

Generative AI gives scammers cleaner writing, which is why the old advice of “look for typos” works less often. Google’s own post introducing the feature frames it the same way: scammers send fraudulent texts directly to your phone or through messaging and social apps, and the messages often demand money or link to scam sites.

And Google keeps pointing people at the feature. On October 15, 2026 it published a short guide titled “Check if a text message is a scam using Circle to Search”: circle the message, read the AI Overview, and if it shows warning signs, delete the message and block the number. The post is limited to Android and says “results, availability, and compatibility may vary.” Android Authority ran a hands-on piece the same day and reported that it worked as intended in its testing. To be clear about timing: this is not a new launch. Google announced the capability on December 2, 2025. October’s coverage is a reminder that it exists.

Your four options, and which to use

You have four routes. They differ in what they need and in what they do with your message.

Option 1: Circle to Search (Android)

Google’s steps, from its December 2, 2025 post: long-press the home button or the navigation bar on your Android phone, then circle the suspicious text. Google says its systems then use AI and information from the web to assess whether the message is likely a scam, and show an overview with guidance and suggested next steps.

Source: News from Google, “Use Circle to Search and Google Lens to spot scam messages” (December 2, 2025).

Expected result: a short AI Overview that lists what looks off about the message and what to do next, such as not clicking the link and contacting the company through an official channel.

Two details matter. First, Google says the scam assessment appears only “when our systems have high confidence in the quality of the response.” So if you circle a text and no assessment shows up, that is not an all-clear. It means Google didn’t give a verdict. Second, this works on messages in any app, including WhatsApp, social apps and email, because you’re circling what’s on the screen. It’s something you trigger. It isn’t an automatic scanner.

Option 2: Google Lens (iPhone, and Android phones without Circle to Search)

Google says the same capability is in Lens, in the Google app on both Android and iOS, in three short steps: take a screenshot of the message, open Lens in the Google app, and tap the screenshot. On an iPhone this is the Google route. You need the Google app installed.

Expected result: the same kind of overview. One advantage: you’re working from a screenshot, so you never go near the link.

Option 3: Paste it into ChatGPT, Gemini or Claude

This works on any phone, and it lets you give context and ask follow-up questions, which the one-shot Google overview doesn’t. Use the prompt in the next section. The tradeoffs: you decide what to share, and what you paste goes to a company with its own data rules (covered below).

Option 4: Let your phone filter first

Neither AI route replaces the filters already on your phone:

• Google Messages has scam detection built in. Google’s May 2025 announcement described real-time conversational scam detection that looks for suspicious patterns as a conversation develops, including ones that start out innocent, processed on the device, and you can turn it off in Spam Protection settings. Google lists categories such as toll-road and billing scams, crypto scams, financial impersonation, gift-card and prize scams, and tech-support scams. October 2026 coverage describes a newer tier that adds cloud-based analysis on selected newer phones. I couldn’t verify that announcement against a primary source, so check your phone’s settings for what applies to you.
• iPhone (iOS 26) lets you go to Messages, then Filters, then Manage Filtering, then Screen Unknown Senders, which sends unknown senders to a separate folder. Apple’s support page notes that this filtering is screening and not a guarantee that what’s left is legitimate.

Filters work on their own, without you doing anything. The AI check is for the messages that get through.

Which to pick: on Android with Circle to Search, use it first, because it’s fastest and keeps you away from the link. On iPhone, use Lens if you have the Google app, or paste into a chatbot if you want to ask follow-up questions. If the message is long, strange or personal (a “wrong number” that has turned into a conversation), a chatbot gives you more.

The safe way to ask: what to paste and what to leave out

Three rules before you paste or screenshot anything:

1. Don’t click the link, even to “see what happens.” Some of the people the scammers catch are the ones who tap just to check. If you paste the text into a chatbot, write the link with brackets around the dots (for example, usps-redelivery-help[.]top), so it can’t be tapped by accident. The AI can still read it.
2. Leave out personal information. Crop or retype the message so your name, full address, account numbers and anything you’d be unhappy to see leaked aren’t included. A screenshot also shows the sender’s phone number and sometimes your name or contacts. The AI doesn’t need them.
3. Ask what it’s trying to get you to do, not “is it safe?” F-Secure’s testing of AI assistants (more below) found that the same question can get different answers on different runs, so a reassuring one-word reply isn’t something to lean on.

Here is the prompt I used. Copy it and replace the last line.

I got this text message and I have NOT clicked or replied. Help me
assess it. Any link has [.] in place of dots on purpose. Answer in
this format:

1. MY READ: likely scam / suspicious / probably legitimate / can't tell,
   with one sentence on why.
2. RED FLAGS: what in the text looks off (type of sender, urgency,
   a request for money, a login, a code or personal details, a link
   that doesn't match the brand, odd wording).
3. WHAT I CAN'T TELL FROM THE TEXT ALONE: for example, whether the
   account or package it mentions is real.
4. SAFE NEXT STEP: what to do using an official app, a website I type
   in myself, or a phone number I already have. Never use anything
   from the text.

Don't tell me a message is "safe". If you see no red flags, say "no
red flags in the text" and still give the safe next step.

TEXT:
[paste the message here]

Four worked examples

These are made-up texts built from the patterns the FTC describes. Each shows what the AI said (shortened) and what to do next.

Example 1: The package text

“USPS: Your package has a pending address issue and could not be delivered. Confirm your details within 24 hours to avoid return: usps-redelivery-help[.]top/track”

The AI’s read, shortened: likely scam. The link’s domain isn’t USPS’s, it uses urgency (24 hours), it asks you to “confirm your details,” and real carriers don’t normally ask for personal details by text. It said it couldn’t tell whether you actually have a package coming.

What you do: don’t tap. If you’re expecting a package, open the carrier’s own app or type the tracking number into the carrier’s website yourself. Forward the text to 7726 and delete it. This is the FTC’s number one text scam from its 2024 data: a fake delivery problem that leads to a small “redelivery fee,” which is really a way to get your card or even Social Security number.

Example 2: The bank fraud alert

“First Lakeside Bank Alert: Suspicious charge of $842.17 at an electronics store. Reply YES to confirm or NO to dispute. For help call 1-8XX-555-0142.”

The AI’s read: suspicious, high risk. Banks rarely resolve fraud alerts through a YES/NO reply to an unknown number, there’s a phone number in the text that you can’t verify, and the message creates alarm to get you to act. It listed what it couldn’t tell: whether you actually have an account at that bank, and whether the number is real.

What you do: don’t reply and don’t call that number. Call the number on the back of your card, or open the bank’s app, and look for the alert there. The FTC describes the fake-fraud-alert scam like this: people who reply or call are connected to a fake fraud department, which then pressures them to move money out of their accounts “to keep it safe.” The money goes to the scammers.

Example 3: The wrong number

“Hi, is this Alex? We met at the conference last week.”

The AI’s read: can’t tell, treat with caution. There’s no link or request in this message, which is why it’s hard to judge from the text alone. It noted that unsolicited “wrong number” openers are a common first step in scams that build a relationship before asking for money. Safe next step: you don’t owe a stranger a reply. If you want to be polite, “Sorry, wrong number” is enough, but a reply also confirms your number is active.

This is the case where the AI is weakest, and it said so. The FTC describes wrong-number scams as ones that start with a message that looks innocent, such as “hello” or “do you want to get a coffee?”, then develop a fake friendship, often with romantic undertones, and then turn to investment pitches that direct people to bogus platforms. People report losing tens of thousands of dollars. Nothing in the first text gives it away.

Example 4: A real one

“Your verification code is 482915. Don’t share this code with anyone. It expires in 10 minutes.”

The AI’s read: probably legitimate, with one caution. This is the standard shape of a one-time code. The caution it gave: if you didn’t just try to log in somewhere, someone may be trying to get into your account, so don’t share the code with anyone, including someone who calls claiming to be from the company, and consider changing your password.

This one matters because AI checks aren’t only for catching scams. They also help you stop panicking about real messages, and a good answer on a real text still tells you what to do. It also shows the kind of false alarm to expect. F-Secure found about one in five harmless questions still triggered a scam warning, so if the AI flags a real code or alert, check through the official app and don’t ignore the message.

Comparison: the tools side by side

The five scams you’ll see most

From the FTC’s top five, found by hand-coding a random sample of 1,000 text-fraud reports from 2024. Together they account for roughly half of text frauds reported:

The pattern is that the first two have a link and a story the AI can examine. The last two may not contain anything wrong in the first message at all. That is why “the AI said it looks fine” shouldn’t be the end of your checking, especially with texts from strangers that ask for nothing yet.

How reliable is the AI’s answer?

Honestly, nobody has published a clean, head-to-head accuracy number for “ChatGPT, Gemini and Claude checking scam texts.” The evidence I found covers neighboring tasks, and it points the same way: useful second opinion, not a verdict.

• F-Secure’s ASK benchmark (published August 13, 2026): 35 scam scenarios, each at four levels of how many red flags were visible, producing 840 scored responses from ChatGPT and Gemini, with Claude acting as judge. Both flagged scam descriptions more than 95% of the time on average, and about 95% when the scam was fully described. But it measured whether the assistant volunteered a warning while helping with a task, not whether it classified an uploaded text correctly. Also: about one in five harmless questions still triggered a scam warning, and on a “reshipping coordinator” job the same ChatGPT question scored 0, 4 and 4 (on the benchmark’s 0-to-4 warning scale) across three identical runs, meaning one run gave no caution at all. On a gift-card scam disguised as a chore, every run from both assistants gave tips for scratching off the cards.
• A 2024 study of AI spear phishing (Heiding and colleagues, arXiv 2412.00586) tested Claude 3.5 Sonnet on 381 emails, 363 phishing and 18 legitimate. It caught 97.25% of the phishing emails with no false positives, though only 81% of some conventional phishing emails. The authors also found models did better when asked whether an email is suspicious than when asked what its intent is, which is one reason the prompt above asks for red flags. Encouraging, but it’s email and not texts, and 18 legitimate messages is far too few to trust the false-positive number.
• Language matters. Most of this research is on English. I found a 2025 benchmark of Portuguese-language mobile-money smishing in Mozambique (MOZ-Smishing, AfricaNLP 2025), a reminder that a checker tuned on English scams may not carry over to other languages and local scam styles.
• Google’s own caveat: the scam assessment appears only when its systems have high confidence. It publishes no accuracy figure.

F-Secure’s four recommendations are sensible rules for everyone: give the AI context about why you’re asking, but never share sensitive details like Social Security numbers or passwords; don’t treat a lack of warning as proof of safety; if something feels off, ask directly “could this be a scam?”; and for high-stakes situations involving money, urgency or secrecy, involve a trusted person. In their words, AI is a second opinion, not the final word.

What to do after the AI answers

Whatever the AI says, these steps are the ones the FTC recommends, and they’re the actual protection:

1. Never click links or reply to unexpected texts. If you think the message might be real, contact the company using a phone number or website you know is genuine, not the details in the text.
2. Report it. Forward the message to 7726 (SPAM), which helps your wireless provider spot and block similar messages. Report it in your Messages app (Apple Messages or Google Messages), and report it to the FTC at ReportFraud.ftc.gov.
3. Block the number and delete the text.
4. If you already clicked or replied, don’t panic. Don’t enter anything more. If you typed a password, change it, and use the official website or app to do it. If you shared card or bank details, call the number on the back of your card or the bank’s official line right away. The FTC’s ftc.gov/scams has steps for recovering money if you’ve paid a scammer.

A privacy note on reporting: reporting isn’t free of sharing. Google’s help pages say that reporting spam in Google Messages can send Google the sender’s number and the last ten incoming messages with timestamps, that employees or contractors may review reports, and that reported conversations may be used to train spam-protection models. Apple’s Report Junk sends the message and sender information to Apple. You can also just block a number without reporting it. That’s your choice.

Privacy: what happens to the text you paste

If you use a chatbot, the training settings are the same ones as any other chat. Quick version, from each provider’s help pages as of October 20, 2026:

• ChatGPT: free-account content may be used to improve models unless you turn off “Improve the model for everyone” in Data controls. Temporary Chat isn’t used for model improvement but may be kept up to 30 days for safety.
• Gemini: saved activity may be used to improve Google’s services, and some chats are reviewed by humans. Turn off “Keep Activity,” or use a temporary chat, which isn’t used for training and is kept for 72 hours.
• Claude: chats may be used to improve Claude if you allow it; turn off “Help improve our AI models” in Privacy settings, or use an Incognito chat, which isn’t used for training and is kept for 30 days by default.

For a scam text, the practical fix is simple: crop or retype so your name, number and account details aren’t in what you share. The scam message itself is the part you want checked, and it’s the part that doesn’t need your identity.

What this means for you

If you’re on Android. You already have the fastest option. Your first action: try Circle to Search on a harmless text today, so you’ve seen what the result looks like before a scary one arrives. Check that Spam Protection is on in Google Messages.

If you’re on an iPhone. Install the Google app if you don’t have it, and practice the screenshot-and-Lens route once. Turn on Screen Unknown Senders in Messages if you can live with the separate folder. Use a chatbot when you want to ask follow-up questions.

If you help an older parent or relative. You’re probably the “can you look at this text?” person. Your first action: agree on one rule, which is that they forward you any strange message without tapping, and you check it through an official channel. Show them 7726. Scammers use urgency and secrecy, so the habit of showing someone matters more than any tool.

If you run a small business. Scammers impersonate banks, couriers and customers, and your staff get the texts too. Your first action: write the three-line rule (don’t tap, check through the official app or known number, report to 7726) on one page and post it where your team will see it. Our small business course covers using AI safely in daily work.

If you’re job hunting. The FTC says “task scams,” where an unexpected text offers simple online work like rating apps, took off in 2024, and people who sent money to withdraw earnings didn’t get it back. Your first action: treat any unsolicited job text with pay and no specifics as suspect, and apply for jobs through the company’s own careers page.

If you sell things online. Fake “payment sent, click to confirm” and courier texts target sellers. Your first action: never confirm a payment through a link in a text. Open your payment app or bank yourself.

If you get a lot of “wrong number” texts. Don’t engage. Block the number. It’s the pattern the AI is worst at, and it’s the one where a harmless “hello” can lead to a loss of tens of thousands of dollars, according to the FTC’s reports.

Edge cases and troubleshooting

You circled the text and nothing came up. Google says an assessment appears only when it has high confidence. No answer isn’t safe or unsafe. Treat it as unknown, check through an official channel, or paste it into a chatbot for a second look.

Your Android phone doesn’t have Circle to Search. Not every Android phone does. Use Google Lens from the Google app with a screenshot, the same route iPhone users use.

The AI says it looks fine, but you still feel unsure. Trust the feeling and verify through the official app or a number you already have. F-Secure’s results showed a chatbot can give a cautious answer on one run and a careless one on another with the same input.

It flagged a real message as a scam. Common with real two-factor codes, delivery updates you signed up for, and appointment reminders, because they contain links, codes and urgency. Check through the official app instead of ignoring it. If a service really sends you texts with links, you can often find the sender’s official short code on the company’s website.

The text came from a number in your contacts. A friend’s number can be spoofed or their phone hacked. If a known contact suddenly asks for money, a code or a favor with a link, call them on the number you already have.

The text is in another language. Gemini and the others handle many languages, but I wouldn’t assume the same quality in each. Pasting in the original text and asking for an English summary of what it’s asking you to do is a good first step. Google’s Circle to Search works on what’s on screen. Check the language list for your device.

You took a screenshot and now it’s in your photos. It may sync to a cloud backup. Delete it after you’ve checked, and don’t share it publicly without cropping out the sender’s number.

You already tapped the link. Close the page without entering anything. If you entered a password anywhere, change it from the official site or app. If you entered card details, call your card issuer using the number on the card. Report the message to 7726 and ReportFraud.ftc.gov.

What this can’t fix

It can’t see the person behind the number. The AI reads the words. It doesn’t know who sent them. A convincing text from a hacked contact or a cloned number can look perfect.

It’s weakest on slow scams. Wrong-number and romance-style scams build trust over many messages before any request. Each individual text can look harmless.

It isn’t a measured detector. Nobody has published a head-to-head accuracy rate for these tools on scam texts, and results vary by run. Google’s feature gives no verdict when it isn’t confident.

It doesn’t stop the message from arriving. For that you need filters, blocking and carrier tools. The AI check helps you after the text is on your screen.

It can’t recover money. If you’ve already paid, the AI can’t help. Contact your bank immediately and report at ReportFraud.ftc.gov. For business-payment scams, our guide for bookkeepers and agents covers the email version of the same problem.

FAQ

Can ChatGPT tell me if a text is a scam? It can read the message and point out red flags such as urgency, odd links and requests for money, and it can explain what the message wants you to do. It can’t verify that the sender is who they claim to be, and its answers can vary from run to run. Use it as a second opinion.

How do I use Circle to Search to check a text? On an Android phone that has it, long-press the home button or navigation bar and circle the text. Google then shows an AI Overview with an assessment and suggested next steps, but only when it has high confidence in its answer.

How do I check a scam text on an iPhone? Take a screenshot, open Google Lens in the Google app, and tap the screenshot. Or paste the text into ChatGPT, Gemini or Claude. Never tap the link in the original message.

Is it safe to paste a scam text into a chatbot? The scam text itself is fine, but remove your own name, number and account details first. Write the link with brackets around the dots so it can’t be tapped by accident, and check the provider’s training settings.

Should I reply “STOP” to unwanted texts? The FTC advice is not to respond to unexpected texts at all, because a reply can tell the sender your number is live. Report the message to 7726 and block the number instead. A legitimate sender you opted into is different, and you can use its official unsubscribe route.

What does forwarding a text to 7726 do? The FTC says it helps your wireless provider spot and block similar messages. Report it at ReportFraud.ftc.gov too.

What are the most common text scams? From the FTC’s 2024 data: fake package delivery problems, phony job opportunities, fake fraud alerts, unpaid toll notices and wrong-number texts. Together they make up roughly half of reported text fraud.

What if the AI says a text is “probably legitimate”? Then check anyway if it asks for money, a login or a code. Go to the company’s official app or website by typing it yourself, or call a number you already have.

Did Google just launch this? No. Google announced scam checks in Circle to Search and Lens on December 2, 2025. The renewed attention in October 2026 comes from coverage reminding people it exists.

Does Apple have something like this? Apple’s iOS 26 can screen unknown senders into a separate folder and has Report Junk, but it isn’t the same as asking an AI to read a specific message. On iPhone, the Google app’s Lens route or a chatbot fills that gap.

The bottom line

When a text asks you to hurry, pay or confirm something, pause before you tap. Use whatever AI is closest to hand to read it for you: Circle to Search on Android, Lens on iPhone, or a chatbot if you want to ask follow-up questions. Hide the link, leave out your personal details, and read the answer as a second opinion. Then do the three things that actually protect you: don’t click or reply, check through an official app or a number you already have, and forward it to 7726.

The tools are good at the obvious scams, such as the fake delivery fee and the fake toll. They are weakest on the quiet ones, like the “wrong number” that wants to be your friend. For those, your own rule beats any app: if a stranger starts a conversation, you don’t have to continue it.

To build the habit of using AI with care, our AI Fundamentals course is a good starting point, and AI Privacy 101 covers what to keep out of any chatbot. For related reading, see our guides on AI voice cloning scams and ChatGPT scams. For a free starting prompt, try the Phishing Email Detector skill, which works on suspicious messages too.

Update note: this guide reflects Google’s published December 2025 announcement, October 2026 coverage, and each provider’s help pages as of October 20, 2026. Features and settings change, so check your own phone’s current screens. The FTC text-scam figures are for 2024.

Sources

• News from Google, “Use Circle to Search and Google Lens to spot scam messages” (December 2, 2025)
• Google, “Check if a text message is a scam using Circle to Search” (October 15, 2026)
• Android, “Stop text message scams in their tracks with Circle to Search”
• Android Authority, “Google Lens and Circle to Search can help detect scams” (October 2026)
• FTC Data Spotlight, “Top text scams of 2024” (April 14, 2025)
• FTC press release, testimony before the Joint Economic Committee (March 2026)
• ReportFraud.ftc.gov, report a scam to the FTC
• F-Secure, “AI can answer almost anything. But can it spot a scam?” (August 13, 2026)
• OpenAI, “Disrupting malicious uses of AI” (October 2025)
• Google Security Blog, “What’s New in Android Security and Privacy in 2025”
• Apple Support, Filter and block messages on iPhone
• Google Messages Help, Report spam
• Heiding et al., “Evaluating Large Language Models’ Capability to Launch Fully Automated Spear Phishing Campaigns” (arXiv 2412.00586)
• MOZ-Smishing: A Benchmark Dataset for Detecting Mobile Money Frauds (AfricaNLP 2025)
• OpenAI Help, Data controls in ChatGPT
• Google, Gemini Apps Activity and privacy
• Anthropic, Is my data used for model training?