The evidence, for and against

Training interventions had no significant effect on click rates or reporting rates, with negligible effect sizes. Phish Scale difficulty did predict behavior (7% clicks on easy emails vs 15% on hard), and in 36-55% of campaigns reports arrived before clicks, though training did not improve this.

Embedded feedback shown only to people who fail has limited reach and weaker field effects than lab studies suggested. Sending delayed feedback to all employees after a simulation (non-embedded training) emerged as a more promising way to reduce vulnerability over time.

Susceptibility and intervention effectiveness varied sharply by staff group, and risk from a few phishing emails lingered about three days. Technical measures (filtering, in-email warnings) worked best, generic [EXTERNAL] tags did little, and some staff reacted to the simulation with fear, shame, guilt, and hostility.

The main reason employees report suspicious emails is wanting to protect and help the organization and coworkers, followed by responsibility, awareness of consequences, and feelings of insecurity. Reporters were more likely to report convincing, well-impersonated emails, showing they judge potential impact.

Recent completion of annual awareness training had no significant link to failing phishing simulations, and embedded training produced only tiny differences in failure rates. Most users spent minimal time on training pages, and for some content types more training was associated with higher later failure rates.

Whatever benefit embedded training has comes from the nudge of being periodically reminded of the threat, not from the training content, which employees rarely read. Delaying training was as effective as immediate training, rewards did not improve behavior, and phishing looked like an attention problem rather than a knowledge problem.

Click-through was 19-28% for generic control phishing, 30-44% for GPT-4 generated emails, 69-79% for emails designed by hand using the V-Triad cognitive-bias rules, and 43-81% for GPT-4 combined with the V-Triad. Large language models were also fairly good at detecting phishing intent, sometimes beating humans, and cut attacker costs.

Fully AI-automated spear-phishing emails drew a 54% click-through rate, matching human experts (54%) and far above arbitrary control phishing (12%), a big jump from comparable AI results a year earlier. The AI's reconnaissance profiles were accurate and useful for 88% of targets, and AI can raise attacker profitability up to 50-fold at scale.

People who report suspected phishing typically describe evidence they noticed, possible impacts, what they did or did not do, and questions they have. Some build clear arguments for why the email is phishing and why the organization should act.

Government security awareness programs lean heavily on compliance metrics such as training completion rates and struggle to find other ways to judge whether behavior actually changed.

Embedded training shown after failing a simulation did not make employees more resilient and could even increase susceptibility, while email warnings helped. Employees reporting suspicious emails worked as a fast, sustainable crowd-sourced detection system, with reporters staying active over long periods.

Handling phishing reports is a distributed process across several teams, each with narrow system access and knowledge. Sudden large campaigns flooded the help desk with reports, disrupting work and slowing mitigations and reflection.

After an awareness program, employees identified phishing and legitimate emails significantly better right away and at four months, but the gain was gone by six months. Reminders based on videos and interactive examples worked best and lasted at least another six months.

Click rates should be expected to vary with how hard a phishing email is for a given audience, especially when its premise fits the recipient's work context. The authors propose the NIST Phish Scale so programs can rate exercise difficulty and interpret click rates.

Experts detect phishing in three stages: making sense of the email and noticing small discrepancies, becoming suspicious when something (usually a link asking for action) triggers the phishing explanation, then investigating and deleting or reporting. Training should build this sensemaking process, not just checklists.

Measuring Cialdini-style persuasion triggers in real reported phishing let the authors predict which attacks would draw the most clicks, enabling response teams to prioritize takedowns. Consistency and scarcity triggers were associated with more clicks, reciprocity appeared counterproductive, and authority, social proof and liking showed no clear trend.

A review of the wider fear-appeal literature finds real disagreement over whether fear appeals are helpful or advisable, and wide variation in how cyber security fear-appeal experiments are designed. The authors propose a standard protocol for such studies.

A problematization analysis finds government, industry and hacker discourse treats humans as the problem, with controls designed to constrain them. The authors propose 'Cybersecurity, Differently', which assumes people are well-intentioned and builds on what contributes to positive outcomes and resilience.

In a role-play study of 985 people, emails using consistency and reciprocity were most effective while scarcity and social proof were least effective. People who scored as susceptible to a given principle were usually more fooled by emails using it, and age, computer time, social-proof susceptibility and impulsivity predicted detection ability.

43% of participants clicked at least one simulated phishing email, with older women most susceptible. Young users' susceptibility dropped over the 21 days while older users' stayed flat, and the effectiveness of each persuasion technique and life-domain topic varied by age group. Older users also rated their own susceptibility lower than it was.

Across a simulation sent to about 62,000 employees, emails carrying authority cues raised the likelihood of clicking a suspicious link. Focus groups pointed to workplace factors, such as routine email habits and work pressures, that shape whether employees fall for spear phishing.

Because training effects fade as people slip back into email routines, the authors built a model (SCAM) combining conscious cognitive processing, preconscious suspicion and habitual, automatic email use, and tested it across two phishing experiments. Email habits emerged as a key predictor of susceptibility alongside cognitive processing.

Attention to repeated security warnings declined measurably in the brain across a workweek, partially recovering between days. In the field, adherence to permission warnings fell over three weeks, while warnings whose appearance varied (polymorphic designs) substantially reduced this habituation.

Facts-and-advice training beat no training only when presented by a security expert, while story-based training worked much better when told by a peer. Who delivers training can strongly change security outcomes.

Although the interviews never asked about fatigue, over half of the 40 participants described it: resignation, loss of control, fatalism, risk minimization and decision avoidance. This fatigue fed their sense that following security advice has little benefit.

Awareness campaigns fail when they only provide information: people must be able to understand and apply advice and be motivated to act, which requires attitude and intention change. Reviews persuasion techniques, including fear appeals, and lists factors behind campaign success or failure.

Differently framed embedded training messages had no significant effect on whether people clicked later spear-phishing emails, and many people either clicked every link or none regardless of training. Employees largely did not read the training materials.

Beyond comply/not-comply, security-conscious employees who cannot follow policy build their own workarounds ('shadow security') that balance getting work done with managing risk. The authors recommend learning from these practices rather than stamping them out.

Employees' rationalizations for rule-breaking ('neutralization' techniques from criminology) explained intentions to violate security policy better than deterrence theory's sanctions. The authors argue policies should address these rationalizations.

Embedded email-based training (PhishGuru) and a game (Anti-Phishing Phil) built on learning-science principles improved users' ability to recognize phishing. The authors frame user education as a complement to automated detection, noting that users are unmotivated and that training can raise false alarms on legitimate messages.

Threat severity, response efficacy, self-efficacy and response costs shaped attitudes to security policy. Organisational commitment and social influence significantly drove compliance intentions, and available resources boosted self-efficacy. Employees underestimated how likely breaches were.

Argues that users ignoring security advice is economically rational: advice imposes large, constant effort costs while its benefits are often speculative. For example, the time cost of everyone checking URLs would dwarf all phishing losses.

User awareness of security policies, SETA programs and computer monitoring each deterred intentions to misuse IS, working through perceived sanctions. Perceived severity of sanctions mattered more than certainty, and the effect varied with individuals' morality.

Employees decide whether to comply by weighing the personal costs and benefits of compliance against perceived benefit to the organization. Proposes the 'Compliance Budget': a finite store of goodwill that security demands draw down and that must be managed.

Drawing on marketing-persuasion theory, a field study tested whether the factors that make marketing campaigns work also explain who falls for phishing and pretext phone attacks, and found that they do: dispositional factors tied to persuasion predicted victimization.

Habits are learned links between context cues and responses that form through repetition and can then be triggered directly by those cues without the goal that originally drove them. Goals shape habits mainly by motivating the early repetition and by steering people toward cues, rather than by directly activating the habit.

Introduces team psychological safety: a shared belief that it is safe to take interpersonal risks. Psychological safety, not team efficacy, was associated with learning behavior such as seeking feedback and discussing errors, and learning behavior mediated the link to team performance.

Insecure password practices (e.g., writing passwords down, linking passwords) stemmed from memory overload and poorly designed policies, not user carelessness. Security departments that withheld information and treated users as a threat worsened motivation and compliance.

Healthy participants began choosing advantageously and showed anticipatory skin-conductance responses to risky options before they could explain which strategy was best. Patients with prefrontal damage never developed these signals and kept choosing badly even after knowing the right strategy.

People's expectations that they can successfully perform a behavior determine whether they try, how hard they work and how long they persist. These efficacy beliefs are built mainly through actual successful performance, and also through watching others, verbal persuasion and emotional arousal.