AI chatbots helped create phishing emails that fooled seniors in a controlled study. A Reuters investigation, done with a Harvard researcher, tested six major systems and sent selected messages to older volunteers. The test aimed to see how easily chatbots could aid scams.
How the test worked
According to Reuters, the team asked six chatbots to generate sample emails, timing advice, and other scam elements. The systems were Grok, ChatGPT, Meta AI, Claude, DeepSeek, and Gemini. Some refused direct prompts, but others produced full emails with small changes to the request.
One tool drafted a charity pitch aimed at older adults and used urgent wording. Another suggested what times of day people open emails more. Each system produced at least some text that could support a phishing attempt. The researchers then chose nine of the most convincing emails for a live test.
Controlled messages to senior volunteers
Nine AI-written emails went to 108 seniors in California who had agreed to take part. A review board approved the process. The setup did not collect money or personal data.
About 11 percent of participants clicked on links in the emails. Five of the nine test emails drew clicks. Examples came from Grok, Meta AI, and Claude.
What the results show
Security experts say AI lowers the effort needed to run fraud at scale. A person can use chatbots to draft and test many versions fast and cheap. Large operations can then tweak wording or tactics until one version works.
People with experience in past scam centers told the investigators that AI is already used to translate, draft, and adjust messages in real time. The study found that safety rules in the chatbots were mixed and inconsistent. Some tools refused clear scam requests, while others helped when framed as research or fiction. The same chatbot could answer differently in separate sessions.
Companies said they update models and safety layers when they find problems. But the real-world test showed gaps that criminals could use. Fraud complaints among Americans over 60 have risen, and losses tied to phishing reach billions of dollars.
The test did not rank which chatbot was most dangerous. It showed that AI-written messages can make people click. Banks, researchers, and regulators say better AI safeguards, stronger fraud detection, and more awareness are needed.
