Phishing without typos: the new tells to look for

For years, the advice was simple: look for the spelling mistakes. Clumsy grammar, odd phrasing and a strange greeting gave plenty of scams away. That advice has aged badly.

In January 2024, the UK National Cyber Security Centre assessed that generative AI already lets attackers write convincing messages and lure documents without the translation, spelling and grammar errors that often revealed phishing. A polished email is no longer a sign of a real one.

The useful news is that the tells did not disappear. They moved. Every phishing message still has to arrive somewhere, ask you for something and push you to act before you think. Those parts are much harder to fake, and they are where a few seconds of attention pay off.

How much phishing is AI-written? The numbers disagree

You will see confident figures quoted for the share of phishing written by AI, and secondary sources disagree widely. The primary reports we could open measure different things and land in very different places:

• Verizon's 2025 Data Breach Investigations Report said that, according to data from one partner, synthetically generated text in malicious emails doubled over the past two years. The same report noted that as of early 2025, generative AI had "still not taken over the world."

• Hoxhunt's phishing trends report, based on samples of emails reported by its users, found that under 5% looked AI-generated for most of 2025. That share jumped to 56% in December 2025 and was 40% in January 2026. Hoxhunt notes its method only catches surface-level signs of AI use.

• ENISA, the EU cybersecurity agency, said in its ENISA Threat Landscape 2025 report that AI-supported phishing campaigns reportedly made up more than 80% of observed social engineering activity worldwide by early 2025.

What the evidence does agree on is effect. Microsoft's Digital Defense Report 2025 says AI-automated phishing emails achieved a 54% click-through rate, compared with 12% for standard attempts. The figures match an academic study with 101 participants that Microsoft cites, published in November 2024, in which AI tools researched each person and wrote a personalized email. Emails written by human experts also reached 54%. The study is small, but its point is clear: personalization works, and AI makes it cheap.

So don't build your defense on whether a message "sounds like AI." Assume any message can be well written, and check the things a writing tool can't fix.

The four tells that still work

Context: does this fit what you know?

Ask whether you expected this message. A supplier who has never mentioned a bank change, a manager who is supposed to be on vacation, a delivery you never ordered. Attackers can research you, but they rarely know what you were actually doing this week.

Channel: did it arrive the usual way?

Real requests tend to come through the same routes every time. Be wary when a conversation moves somewhere new. In May 2025, the FBI warned that criminals impersonating senior U.S. officials were sending texts and AI-generated voice messages, often with a link that claimed to move the chat to another platform. Microsoft also describes attackers flooding an inbox with sign-up emails, then contacting the target on Teams or by phone while posing as IT support to get remote access.

Pressure: are you being rushed?

Urgency, secrecy and authority are the oldest tools in the kit, and AI has not replaced them. In early 2024, a finance worker at engineering firm Arup in Hong Kong first suspected a message because it asked for a secret transaction. A video call with deepfake versions of the CFO and colleagues then talked him out of his doubts, and HK$200 million (about US$25.6 million) was sent across 15 transactions, according to CNN. His first instinct about the secrecy was right.

The request: what are you actually being asked to do?

Strip the message down to its verb. Pay, change bank details, log in, share a code, install a tool, buy gift cards. Those requests deserve a check no matter how friendly or senior the sender seems. The FBI's advice is blunt: never give a two-factor code to anyone over email, text or a messaging app.

A checklist you can copy

Before you click, pay or share anything, run through this. It takes less than a minute.

1. Did I expect this message, and does it fit what I know is going on?

2. Did it come through the channel this kind of request normally uses?

3. Is someone pushing me to hurry, keep it quiet or skip a normal step?

4. What exactly is the request: money, payment details, a login, a code, software, gift cards?

5. Do the address, number and link match what I already have on file, rather than what the message gives me?

6. Can I confirm it on a second channel I already trust, such as a call to a number from my own contacts?

7. If I'm still unsure, have I reported it, even if I already clicked?

That last step matters most. If you clicked, say so quickly. The NCSC warns that people who fear reprisals won't report mistakes promptly, if at all, and a fast report is what lets a team contain the damage.

Make it a team habit

Checklists work best when everyone uses the same one. A few habits make that easy:

• Agree on a pause-and-check rule. Any request to move money or change payment details gets a call-back on a known number, whoever it seems to come from. Leaders should say out loud that they want to be checked.

• Make reporting one click and say thank you. The NCSC recommends a reporting process that is clear, simple and quick, with feedback so people see their reports make a difference.

• Bring it home. The FBI suggests families agree on a secret word or phrase to confirm who is really calling.

• Use AI on your side. Paste a suspicious message (with personal details removed, and only if your organization's policy allows it) into an AI assistant and ask what it wants you to do and what looks off. In the same 2024 study, a language model spotted the intent of phishing emails with high accuracy and few false alarms. Treat it as a second opinion, not the final word.

Typos were always a weak signal. Context, channel, pressure and the request are stronger ones, and they work just as well on a perfect email.

Sources

• UK National Cyber Security Centre, "The near-term impact of AI on the cyber threat," Jan. 24, 2024: ncsc.gov.uk

• Verizon, "2025 Data Breach Investigations Report: Executive Summary," 2025: verizon.com

• Hoxhunt, "Phishing Trends Report (Updated for 2026)," 2026: hoxhunt.com

• ENISA, "EU consistently targeted by diverse yet convergent threat groups" (ENISA Threat Landscape 2025), Oct. 1, 2025: enisa.europa.eu

• Microsoft, "Microsoft Digital Defense Report 2025," Oct. 16, 2025: microsoft.com

• Heiding, Lermen, Kao, Schneier and Vishwanath (arXiv), "Evaluating Large Language Models' Capability to Launch Fully Automated Spear Phishing Campaigns: Validated on Human Subjects," Nov. 30, 2024: arxiv.org

• FBI Internet Crime Complaint Center, "Senior US Officials Impersonated in Malicious Messaging Campaign," May 15, 2025: ic3.gov

• FBI Internet Crime Complaint Center, "Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud," Dec. 3, 2024: ic3.gov

• CNN, "British engineering giant Arup revealed as $25 million deepfake scam victim," May 16, 2024: cnn.com

• UK National Cyber Security Centre, "Phishing attacks: defending your organisation," reviewed Feb. 13, 2024: ncsc.gov.uk

Or let NimBot handle it

You can do all of this yourself, and it works. If you'd rather not run it alone, NimBot handles the training, the drills and the reminders for your whole team. Try it at app.nim.bot.