A controlled human-subject study with 101 participants reported a 54% click-through rate for personalised AI-generated phishing and 54% for human-written phishing, compared with 12% for a control condition. That is an important experimental result, but it is one study and does not establish that AI phishing universally bypasses detection at a particular rate.
What the evidence supports
Generative systems can reduce the time and language skill needed to personalise persuasive messages. The study supports treating message polish as a weak trust signal. It does not prove every model, recipient population, delivery channel, or enterprise control environment will produce the same result.
Defensive implications
Security awareness remains useful, but high-consequence requests should also require technical and procedural verification: phishing-resistant authentication, independent confirmation of payment or credential changes, domain and sender controls, protected reporting channels, and rapid incident handling.
Measurement
Organisations should measure their own simulation outcomes and control performance without turning click tests into employee blame. NIST and CISA guidance remains useful for layered phishing defence.
Sources & further reading
Follow the original evidence. Sources may include the organisation making the announcement; claims and independent findings are distinguished in the analysis.
01Human-subject phishing study, arXivarxiv.org02Phishing guidance, NISTwww.nist.gov03Avoiding social engineering and phishing attacks, CISAwww.cisa.govAI-Assisted Phishing: What the Human-Subject Evidence Does and Does Not Show
One controlled study found similar click-through rates for personalised AI and human phishing. That supports layered verification, not a universal detection-rate claim.
Last factual review: 24 August 2026