AI-written phishing fooled people 28% of the time in a BYU test, against 21% for human-written

Seven percentage points separate the two arms: 28% for the AI-written messages, 21% for the human-written ones ^[1][1]. In relative terms that is about a third more clicks per message ^[2]. It also leaves roughly seven in ten AI messages ignored ^[4]. Across the comparisons, the researchers report that AI matched or beat human authors at producing a click 80% of the time ^[2].

For anyone building awareness training, the identification result matters more than the click rate. Asked whether a person or a model had written a given message, participants were right 52% of the time ^[4]. On a two-way judgement, chance is 50%, so they were two points above guessing ^[3]. That was a question about authorship, not about whether the message was a scam. Still, a program that trains staff to judge a message by how it reads is asking for a discrimination these participants made at 52%.

What moved the click rate was the detail about the target. A message naming a co-worker was 2.3 times as likely to be clicked as one appearing to come from a generic organization ^[3], and messages carrying job-related details, a workplace or an occupation, were the most convincing the team tested ^[6]. The inputs are ordinary: company websites, LinkedIn profiles, social media accounts, staff directories ^[5].

A click in a study is a click. Whether it becomes a compromise depends on what sits behind the link, and the published account measures clicks, across text messages and emails ^[13]. It gives no participant counts and no confidence interval for the seven-point gap ^[15], so treat the size of that gap as provisional and the direction as the finding. The same account carries the figure that 90% of corporate hacks originate from spear phishing, unattributed ^[8].

The economic claim sits apart from the click numbers. "AI can reduce the time and effort required to create personalized spear-phishing messages, thus making them more effective and more common," said Derek Hansen, the BYU cybersecurity professor who led the work ^[9][7]. A per-message success comparison cannot test that; what could is time per message, or messages per hour.

Jerson Francia, a BYU cybersecurity doctoral student and co-author, points the defense at the channel instead of the text, recommending verification of unexpected messages through a separate, trusted route before anyone responds or clicks ^[7][12]. "We need to more thoroughly verify the authenticity of messages through other means, rather than relying solely on the content of the message itself," Francia said ^[11].