The Evidence Turns Against Deployed Interventions

Organizations spend millions on mandatory phishing training, assuming it reduces click-through rates. Compliance requirements mandate it, but nobody measures whether employees actually get better at spotting attacks.

Method: A 12,511-person field study at a US fintech firm used the NIST Phish Scale to standardize phishing difficulty across test emails. Even after training, employees showed no significant improvement in detecting phishing attempts. The study controlled for email difficulty—something prior research ignored—and found training effects vanished when you account for varying attack sophistication. The NIST scale rates phishing emails from 0-5 based on cue detectability; trained employees performed identically to untrained ones across all difficulty levels.

Caveats: Single organization study in fintech. Results may not generalize to industries with different threat models or employee technical literacy.

Reflections: Do continuous micro-training interventions (e.g., real-time warnings) outperform annual training modules? · Can personalized training based on individual click patterns improve outcomes? · What's the cost-benefit ratio of training versus technical controls across different organization sizes?

Blind and low-vision players are locked out of most video games. Developers build accessibility features game-by-game, requiring custom programming for each title—an approach that doesn't scale.

Method: GamerAstra uses a multi-agent AI framework that sits on top of existing games without modifying source code. One agent extracts visual information from the screen buffer, another translates it to audio descriptions, and a third handles player commands. The system works across multiple 2D games—tested on platformers, RPGs, and puzzle games—by treating the game as a black box. It provides spatial audio cues for object locations and converts visual menus into navigable audio interfaces. The framework reduced task completion time by 40% compared to screen readers in user testing.

Caveats: Only works for 2D non-twitch games. Fast-paced action games with split-second timing remain inaccessible due to audio latency.

Reflections: Can the framework extend to 3D games with depth perception challenges? · How do players customize audio density without overwhelming information? · What's the latency threshold where audio feedback becomes unusable for gameplay?

Mental health support communities assume universal language for distress. But if partisan culture shapes how people express suffering, platforms may be inadvertently filtering or mismatching users based on political identity.

Method: Analysis of mental health subreddits found Republicans use 23% more religious language ("pray," "faith," "blessed") and Democrats use 31% more clinical terminology ("therapy," "diagnosis," "medication") when describing identical symptoms. The study used LIWC dictionaries to categorize language patterns across 50,000 posts, controlling for symptom severity. Republicans frame distress as moral struggle; Democrats frame it as medical condition. This isn't just word choice—it predicts which support resources users find credible and which moderators they trust.

Caveats: US-only sample from Reddit. Partisan identity may correlate with other factors like religiosity or education that independently affect language.

Reflections: Do partisan language patterns affect treatment adherence or outcomes? · Can platforms bridge partisan divides by exposing users to alternative framings? · How do moderators from different partisan backgrounds handle identical distress posts differently?