Verification Layers Emerge as the Architecture for Unreliable AI

LLM-based mobile GUI agents automate tasks via natural language, but their probabilistic nature makes them unreliable—they execute actions that don't match user intent, breaking workflows.

Method: VeriSafe Agent translates natural language instructions into formal specifications, then uses rule-based verification to check each agent action before execution. Tested on 300 instructions across 18 mobile apps using GPT-4o, it achieved 94.33%-98.33% accuracy in verifying actions—outperforming existing LLM-based verification by 16.33%-30.00%—and increased task completion rates by 90%-130%.

Caveats: Tested only on mobile apps. Desktop GUI agents or web automation may require different formalization approaches.

Reflections: Can autoformalization scale to ambiguous instructions where user intent itself is underspecified? · How does verification latency affect user experience in real-time automation scenarios? · What happens when the formal specification itself misinterprets user intent?

LLMs summarize scientific research for public audiences, but they may strip caveats and scope limitations, overgeneralizing findings beyond what studies actually support.

Method: Tested 10 LLMs on 4900 summaries of scientific texts. Even when explicitly prompted for accuracy, DeepSeek, ChatGPT-4o, and LLaMA 3.3 70B overgeneralized in 26% to 73% of cases. LLM summaries were nearly five times more likely to contain broad generalizations than human-authored summaries (OR = 4.85, 95% CI [3.06, 7.70]). Newer models performed worse than earlier ones.

Caveats: Tested on scientific texts. Generalization bias in other domains (news, policy) remains unverified.

Reflections: Why do newer models overgeneralize more than earlier versions—is it training data or architectural changes? · Can fine-tuning on scope-preserving summaries reduce generalization bias? · Do users actually prefer overgeneralized summaries because they're simpler, creating a misaligned incentive?

Platform moderation fails to remove ableist hate, leaving disabled users exposed. Personalized filters could help, but it's unclear what configuration options users actually want.

Method: Interviewed 23 disabled social media users with design probes showing different filter configurations (intensity vs. type of ableism) and presentation options (AI rephrasing, content warnings). Participants preferred filtering by type of ableist speech and favored content warnings over AI rephrasing. They expressed deep distrust in AI-based moderation accuracy and showed varied tolerances for viewing ableist content.

Caveats: 23 participants, qualitative study. Quantitative validation of filter effectiveness and user adoption rates needed.

Reflections: How do you design type-based filters that scale across the evolving language of ableism? · Can users' distrust in AI moderation be mitigated with transparency about accuracy rates? · What happens when users' tolerance for ableist content changes over time—how should filters adapt?