AI Phishing Emails and Texts: Why They Read Perfectly Now — and How to Check Them
Generative AI has removed the typos that used to expose phishing — the FBI says it corrects the very errors that warned victims. Fluency is no longer a security test. Here are the channel checks that still work.
Phishing used to announce itself: broken English, odd spacing, a logo that looked like it had been photocopied. Those errors were not carelessness — they were a side effect of scale. Now they are gone. The FBI’s Internet Crime Complaint Center (IC3) warns that generative AI lets criminals correct the human errors that once served as warning signs of fraud, and that AI-written text powers everything from phishing messages to fictitious social profiles and fraudulent websites with chatbots built in. This guide explains what replaced the old tells, and the checks that still work. It is part of our complete guide to AI scams and deepfake protection.
Highlights
- The spelling test is dead: the FBI’s IC3 says generative AI corrects the human errors that used to expose fraud attempts.
- Fluency is not identity: a perfect message proves only that it was well produced, not that it came from your bank, your boss or a delivery company.
- Check the channel, not the wording: open the official app or type the official address yourself; never use the message’s own links or numbers.
- Tone and word choice still leak: the FBI notes that tone and word choice can reveal a fraud attempt even in AI-polished text.
Why phishing reads perfectly now
The FBI’s public service announcement “Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud” describes the shift directly: generative AI reduces the time and effort criminals must spend deceiving their targets, and it can be used to correct for human errors that might otherwise serve as warning signs of fraud. In other words, the technology did not just make more phishing — it removed phishing’s most famous flaw. The same announcement lists where the polished text goes: fraudulent social-media profiles written to feel like real people, content for fraudulent investment websites, and chatbots embedded in scam sites that can hold a convincing conversation with a victim who asks questions.
Three practical consequences follow. Messages can now be personalised cheaply — your name, your employer, a real delivery you are expecting. They can be flawless in any language, which removes the last advantage careful readers had. And they can be consistent across a conversation, so the follow-up text answers your doubts in the same smooth voice. Grammar is no longer a security feature. Treat it the way our guide to AI myths and scams suggests: fluency tells you a machine helped write the message, which is true of most legitimate messages too. It tells you nothing about who sent it.
The checks that still work
1. Verify through a channel you open yourself. The FBI’s advice for suspicious contact is to hang up, research the organisation’s contact information, and call the phone number directly. The digital version is identical: close the message, open the company’s real app, or type the address you already know into your browser. If the “problem” is real, it will be waiting for you there. If it exists only inside the message’s link, it was never real.
2. Interrogate the request, not the prose. Legitimate organisations do not ask you to pay with gift cards, wire transfers or cryptocurrency, do not demand you act within minutes, and do not ask for full card numbers, PINs or one-time codes by email or text. Any message that does one of these has failed the only test that matters, however beautifully it is written.
3. Listen for wrong tone and word choice. The FBI specifically advises considering the tone and word choice in messages as possible indicators of fraud. A “bank” that sounds casual, a “manager” whose phrasing is subtly unlike them, urgency that feels performed — these are the seams AI has not fully closed, because the model knows language but not your actual relationship with the sender.
4. Slow the conversation down. AI chatbots on scam sites and in messaging apps are patient, but the script behind them needs momentum. Ask a question the script cannot answer — a specific detail about your actual account history — and watch the reply steer you back to payment. That steering is the confession.
5. Protect the next person. Report phishing texts and emails to the organisation being impersonated and at ReportFraud.ftc.gov if money or information changed hands. The same verification-first mindset powers our checklist for checking AI answers: never trust a fluent output until an independent source confirms it.
House rules for email and texts
Turn these checks into defaults and phishing loses most of its power. Treat every link in an unsolicited message as decoration — never as a doorway. Keep a separate, strong password and two-factor authentication on your email account, because the FBI warns that criminals also use AI content to harvest credentials, and a hijacked inbox becomes the perfect launchpad for impersonating you. Limit what your public profiles reveal — the IC3 advises keeping social media accounts private and limiting followers to people you know — because personal details are what make the new, polished phishing personal. And when a message claims to be family in trouble rather than a company, switch to the protocol in our voice clone guide: ask for the family code word, then call back on a number you already trust.
What this means
For decades, public advice about phishing amounted to “read carefully” — spot the typo, notice the clumsy logo, feel the wrongness. AI has ended that era, and the honest thing to say is that the advice most people still repeat is now dangerously out of date. The FBI’s own framing is telling: it does not promise that citizens will learn to recognise AI text. It tells them to stop relying on recognition altogether and to verify identity through channels they control. That is the same transition banks made years ago when they stopped trusting signatures and started trusting devices, codes and callbacks.
The comparison with legitimate communication is what makes this manageable. Real companies increasingly design their messages to be verifiable — “we will never ask for this”, “check inside the app” — because they know their emails can be copied perfectly. US readers should exploit that design: the official app and the number on the back of your card are ground truth, and everything arriving uninvited is a claim about that truth, not the truth itself. There is also a quieter shift for the workplace: an instruction from “the boss” by email or text now deserves the same callback discipline as a video call, since the FBI links AI text to impersonation at every level, from social profiles to investment sites. None of this requires new software or sharper eyes. It requires demoting the message and promoting the channel. Perfect grammar used to be evidence of a professional sender. Today it is evidence of nothing — and once you truly absorb that, the new phishing loses the only weapon it actually gained.
Employers are adapting along the same lines, and home users can copy them. Security teams increasingly train staff that any message asking for credentials, payments or secrecy gets verified on a second channel — a rule that assumes the message itself may be flawless. Families can run the household version: any text or email that asks for money, codes or urgent secrecy is treated as unverified by default, whoever it appears to come from. When suspicion becomes the default rather than the exception, AI’s polish stops mattering, because polish was only ever dangerous to people still using it as a test.
Bottom line
Stop trying to spot the fake; start verifying the sender through a channel you control. AI removed the typos that used to protect you — process, not eyesight, is the defence now.
Read Next
- AI Scams and Deepfake Protection: The Complete Guide
- Fake Shopping Websites: A Pre-Purchase Checklist for AI-Era Stores
Sources
