Insights/ Training
Modern learning and cybersecurity: why fully automated AI training falls short
AI training promises scalable, low-cost learning, but research shows knowledge alone doesn't change behaviour. Why training needs human context to work.

Organisations are increasingly turning to AI-driven cybersecurity training, attracted by promises of scalable, personalised learning at low cost. But empirical research and educational science both point the same way: knowledge acquisition alone does not reliably produce lasting behaviour change, especially in a risk-laden domain like cybersecurity. This article looks at where fully automated training falls short, why humans learn best through contextual, supported engagement, and how measurement and adaptive intervention strengthen real-world outcomes.
Data sourcesPrümmer et al., Leiden University: 2024 meta-analysis of 69 studies · University of Bari, AI-Generated Phishing Training study, arXiv (2025) · Verizon Data Breach Investigations Report (2025) · Lain, Kostiainen and Čapkun, ETH Zurich (IEEE S&P 2022)Figures reflect publicly reported research as of October 2026. Findings evolve. Verify current studies before citing externally.
01Where training quietly fails
Why learning is not just information transfer
Modern cognitive and educational science underscores a fundamental truth about adult learning: knowledge is not equivalent to behaviour change. Learners may be able to recall information, yet still fail to apply it in real-world contexts where automatic habits and social pressure dominate decision-making.
Knowing a rule and acting on it are different things.
A 2024 meta-analysis out of Leiden University, reviewing 69 studies on cybersecurity training, found that training reliably shifts the precursors to behaviour, such as knowledge, attitudes and intentions, but its measured effect on actual behaviour was small and not statistically significant.
Adults learn most effectively when training is active, contextualised, and reinforced, and when learners see direct relevance to their role and get consistent feedback over time.
02The AI ceiling
AI as a tool: helpful, but not sufficient
AI-generated content and adaptive learning systems have a genuine role. A December 2025 study out of the University of Bari, run across two controlled studies with 480 participants, found that AI-generated phishing training produced measurable pre-post learning gains regardless of which prompting strategy was used: evidence that generative AI is a viable, scalable way to author training content.
But the same evidence points to clear limits:
- Training alone often fails to produce sustained behavioural change. The Leiden meta-analysis found that while awareness and attitudes improved after training, measured secure behaviour did not change substantially.
- Generic or unmodified AI output may deliver short-term gains but doesn't address the underlying habits and decision processes that lead to unsafe choices.
AI is best understood as one component of a broader, human-centric strategy, not a stand-alone solution.
AI is a co-pilot. It is not yet an autopilot.
It can generate the training content faster and cheaper than a person can write it by hand. It cannot yet read a room, sense which employee is quietly struggling, or decide when a policy needs a human conversation instead of another module. Those are still, for now, a person's job.
03What actually moves behaviour
The importance of behavioural reinforcement and feedback
Training that only transmits knowledge leaves learners without the behavioural anchors they actually need. Programs that build in frequent simulations, feedback loops, and adaptive intervention show a far stronger link to improved behaviour in real organisational contexts.
studies in a 2024 meta-analysis: training moved knowledge and attitudes, but measured behaviour barely shifted.
Prümmer et al., Leiden Universityemployees in a 15-month study where training shown straight after a failed simulation left people slightly more likely to click later.
ETH Zurich, IEEE S&P 2022higher reporting rate of simulated phishing among staff with recent training: 21% against a 5% baseline.
Verizon DBIR, 2025The pattern is consistent: a module on its own changes little, while repeated practice, feedback and a culture that rewards reporting change how people act under pressure.
04Beyond completion rates
Why measurement matters
One of the biggest gaps in cybersecurity training is the absence of meaningful measurement. Automated systems can generate completion rates and quiz scores, but without deeper analysis, those figures can mislead as easily as inform.
Effective measurement goes beyond completion and looks at:
Reporting rates
Who is reporting suspicious items, and how fast?
Behavioural change over time
Is there sustained improvement across repeat exposures?
Realistic scenario outcomes
Are learners making the right call in context, not just picking an answer?
Running training without these metrics risks producing reports that look activity-rich while masking poor behavioural impact.
05The part no model can see
Human context and support networks
Cyber threats are social as much as technical. People are shaped by factors no automated platform can fully account for:
Training that accounts for these, through facilitated discussion, role-specific examples, and contextual reinforcement, fosters deeper internalisation than an isolated AI module ever can. Even the most advanced adaptive AI can't yet simulate these human and organisational dynamics without human oversight.
06Where we stand
The Mono perspective
At Mono, we treat AI and automation as useful tools, not complete solutions. Effective learning prioritises behaviour change anchored in organisational context, with ongoing measurement and adaptive response at the centre of real-world effectiveness.
We build training that combines thoughtful human facilitation, realistic simulation, and data-driven adjustment, so learners do more than know.
They act securely.
·Key takeaways
What this means for your organisation
Information alone rarely produces lasting behaviour change. Completion rates tell you almost nothing about real-world security.
AI can generate training content, but it needs human-centric support and meaningful measurement to produce real results.
Without real analytics, reporting rates, behavioural trends, scenario outcomes, training outcomes stay opaque at best.
Real learning is built on repeated practice, feedback, and relevance to daily work, not a single module completed once a year.


