Can AI write a convincing phishing email or scam message?
YesYes. Controlled studies show fully automated AI phishing now works as well as emails written by human experts, and the FBI and AI companies report criminals using chatbots to write and translate scam messages.
First shown: Aug 23, 2023. Harvard-affiliated researchers found GPT-4-written phishing emails drew 30-44% click-through from 112 participants, against 19-28% for generic phishing, though human-designed emails still did better. (Heiding et al. (arXiv))
What it means
Phishing means a message, usually an email or text, that tricks you into clicking a link, handing over a password or sending money. The old giveaway was clumsy writing. AI removes that: the FBI said in December 2024 that criminals use AI to write believable messages faster, reach more people, and translate them to avoid grammar and spelling mistakes.
Researchers have measured how well it works by sending test emails to consenting volunteers. In a 2023 study GPT-4 emails beat generic phishing but lost to emails crafted by an expert. By late 2024 the same research group found fully AI-automated emails, built from public information about each person, got 54% of 101 participants to click, the same as human experts and against 12% for ordinary spam-style phishing. The study was peer reviewed and published in 2026.
The bigger change is cost. IBM's security team found in 2023 that an AI built a phishing email in about five minutes, against about 16 hours for its human experts, though the human email still won narrowly. The 2024 study estimated fully automated emails cost about four cents each and could make large campaigns up to 50 times more profitable.
AI companies see this in the wild. OpenAI reported in October 2025 that scam networks in Cambodia, Myanmar and Nigeria used ChatGPT to translate messages and build fake investment personas, and in August 2026 that it had disrupted a Cambodia-based network using its models to write and translate messages to targets. OpenAI also says ChatGPT is used about three times as often to spot scams as to make them.
The evidence ladder
Every capability on this site is graded on the same five steps, from a single demonstration to real-world consequences.
| Step | Status | When | Evidence |
|---|---|---|---|
| Demonstrated shown at least once | yes | Aug 23, 2023 | GPT-4 phishing emails drew 30-44% click-through among 112 participants, against 19-28% for generic phishing; expert human emails still did better (69-79%). (Heiding et al. (arXiv)) |
| Reproduced independently repeated | yes | Oct 24, 2023 | IBM X-Force sent AI- and human-written phishing tests to more than 800 staff at a healthcare firm; humans won by the narrowest of margins, and the AI needed about five minutes against about 16 hours. (IBM X-Force) |
| Accessible available to ordinary people, or criminals | yes | Dec 3, 2024 | The FBI said criminals use generative AI to write believable messages faster for wider audiences and to translate them so they contain fewer grammar and spelling errors. (FBI PSA I-120324-PSA) |
| Reliable works most of the time for a defined task | yes | Nov 30, 2024 | Fully AI-automated spear-phishing got 54% click-through from 101 participants, matching human experts (54%) against 12% for a control; peer reviewed in Expert Systems with Applications (2026). (Heiding et al. (arXiv)) |
| Consequences real-world use or harm documented | yes | Aug 4, 2026 | OpenAI disrupted a Cambodia-based scam network that used its models to create fake personas and generate and translate messages to targets; it may have reached hundreds of people. (The Record) |
From the record
- IBM's security team pitted an AI-written phishing test against one by its own experts at a healthcare firm; the humans won narrowly, but the AI took about five minutes instead of about 16 hours. (IBM X-Force)
- The FBI warned that criminals use AI-generated text for spear phishing and fraud, and to translate messages for US victims with fewer errors. (FBI)
- OpenAI reported scam networks in Cambodia, Myanmar and Nigeria using ChatGPT to translate messages and create fake investment personas. (SiliconANGLE)
- OpenAI said it disrupted a Cambodia-based scam network that used its models to write and translate messages to targets and to run its daily operations. (The Record)
- During the 2026 midterms the FBI and CISA warned election officials about AI-enhanced phishing. The case file → (Resemble AI watchlist)
What it means for you
Good spelling and a personal touch are no longer signs that a message is genuine. A scam can now mention your job, your colleagues or your interests and read like a real person wrote it, in any language.
What to do
- Judge the request, not the writingAny message asking you to click, log in, pay or share a code needs checking, however well written.
- Check through another channelContact the person or company using a number or app you already have, not the link or number in the message.
- Slow down on urgencyDeadlines, threats and secrecy are what make phishing work; take time to check.
- Use passkeys or an authenticator appThey stop a stolen password alone from opening your accounts.
Questions people ask
Are AI-written phishing emails better than human ones?
In a 2024 study they were as good: fully AI-automated emails got 54% of participants to click, the same as emails by human experts.
Can I spot AI phishing by bad grammar?
No. The FBI says criminals use AI to translate messages and limit grammar and spelling errors, so clean writing proves nothing.
Do chatbots refuse to write scam messages?
They are meant to, but OpenAI has repeatedly reported and banned scam networks that used ChatGPT to write and translate messages.
Can AI help me spot a scam?
OpenAI says ChatGPT is used about three times as often to identify scams as to create them; still, verify through a trusted channel.
Related
- AI scam texts: how to spot them and what to do (2026) →
- Is my boss really asking for gift cards or a transfer? →
- Is this call from my bank's fraud team a scam? →
- How to set up an AI chatbot for an older parent, safely →
Sources
- FBI PSA I-120324-PSA (Dec 2024)
- Heiding et al., Devising and Detecting Phishing (arXiv, 2023)
- Heiding et al., fully automated spear phishing (arXiv, Nov 2024)
- Schneier: peer-reviewed version in Expert Systems with Applications (2026)
- IBM X-Force: AI vs human phishing (Oct 2023)
- The Register on Microsoft Digital Defense Report 2025
- SiliconANGLE on OpenAI's October 2025 threat report
- The Record: OpenAI disrupts Cambodia scam network (Aug 2026)
Spot an error or something out of date? Tell us; corrections are made and logged. Written with AI assistance from public sources; every fact checked and approved by Colin Balfe. General information, not legal, medical or financial advice.