AI Systems Succeed at Tasks While Eroding Human Agency

LLM-based mobile agents re-derive every task from scratch, burning inference costs and never improving. A task that worked yesterday gets the full reasoning treatment today—no memory, no speed gain, no reliability improvement.

Method: SkillDroid compiles successful LLM trajectories into parameterized skill templates—sequences of UI actions with weighted element locators and typed parameter slots—then replays them without LLM calls. Over 150 rounds with instruction variation and controlled perturbations, it hit 85.3% success (23 points above the stateless baseline) using 49% fewer LLM calls. Skill replay achieved 100% success across 79 rounds at 2.4x speed. Most critically: success rate converged upward from 87% to 91%, while the baseline degraded from 80% to 44%.

Caveats: Tested on controlled mobile GUI tasks. Real-world app updates and UI redesigns may break skill templates faster than the study captured.

Reflections: How frequently do real-world app updates invalidate compiled skills, and can update detection trigger preemptive recompilation? · Can skill templates transfer across similar apps (e.g., different email clients) or do UI variations require per-app compilation? · What's the minimum number of successful executions needed to compile a reliable skill template?

AI writing assistance improves quality but erodes ownership—the sense that the work is yours. Ownership matters for attribution, cognitive engagement, and rights. Researchers tested whether the stage of AI support changes this tradeoff.

Method: Between-subjects study (n=253) on short essay writing found all AI assistance decreased ownership, but planning support minimally so, while drafting support saw the largest drop. The variation mapped directly to AI-contributed text and ideas: more AI contribution meant less ownership. Notably, an AI-generated draft based on participants' own outline contributed more AI ideas than AI planning support. Meanwhile, more AI contributions improved essay quality.

Caveats: Tested on short essays. Long-form writing with multiple revision cycles may show different ownership dynamics.

Reflections: Does iterative revision of AI-generated drafts restore ownership over time, or does the initial authorship loss persist? · How does ownership vary when AI planning support is highly directive versus suggestive? · Do domain experts (professional writers) experience the same ownership loss as general participants?

During complex knowledge work, cognitive processes—interpreting information, connecting ideas—are hard to share with AI. They arise in situ, rarely get captured without interrupting flow, and when expressed, scatter or get reduced to system-generated summaries that miss the user's thinking.

Method: Contexty enables in-situ snippet memoing to capture cognitive traces, then lets users inspect and refine these contexts to reflect their understanding. Evaluation (N=12) showed it improved task awareness, thought structuring, and users' sense of authorship and control. Participants preferred snippet-grounded AI responses over non-grounded ones 78.1% of the time. The system preserves user agency by making AI context directly inspectable and revisable.

Caveats: Evidence grade B. Evaluation focused on task awareness and preference; long-term adoption and cognitive load effects remain unverified.

Reflections: Does the cognitive load of snippet memoing outweigh its benefits during high-intensity work phases? · How does snippet-grounded context compare to automatically extracted context in terms of AI response quality? · Can snippet organization patterns predict task complexity or user expertise level?