AI Systems Outpace Human Steering Capacity

Generative models produce impressive outputs in demos, but users with specific goals can't reliably steer them to produce what they actually need.

Method: The paper introduces a mathematical decomposition separating producibility (what a model can generate) from steerability (whether users can actually reach desired outputs through available controls). They formalize steerability as the probability that a user can navigate the control space to achieve their goal, independent of the model's raw generation capabilities. This reframes evaluation from 'what can this model do?' to 'can I make it do what I need?'

Caveats: Framework is model-agnostic but requires defining a goal space and control interface—non-trivial for open-ended creative tasks.

Reflections: How do different interface paradigms (sliders vs. text prompts vs. examples) affect steerability for the same underlying model? · Can we predict steerability from model architecture before deployment? · What's the minimum control dimensionality needed for acceptable steerability in different creative domains?

Users increasingly turn to AI chatbots for emotional support, but we don't know whether extended use creates dependency or affects real-world social connections.

Method: Four-week RCT with 981 participants generating over 300,000 messages tested three interaction modes (text, neutral voice, engaging voice) and three conversation types (open-ended, non-personal, personal). Engaging voice increased emotional dependence on AI. Personal conversations reduced loneliness but only when users actively disclosed—passive chatting had no effect. The mechanism matters: it's not chatbot use itself, but the combination of voice intimacy and self-disclosure that drives outcomes.

Caveats: Four-week study; long-term dependency patterns and withdrawal effects unknown. Sample skewed toward users seeking AI companionship.

Reflections: What happens after users stop using high-intimacy chatbots—do dependency effects reverse or persist? · Can we design 'scaffolding' that transitions users from AI support to human connections? · Do cultural differences in emotional expression change these dependency patterns?

GenAI browser assistants can track every search, click, and form input while autonomously performing tasks—raising privacy concerns beyond traditional browser extensions.

Method: Systematic audit of GenAI browser assistants revealed they track detailed browsing activity including search queries, clicked links, and form data. Unlike passive extensions, these assistants can autonomously fill forms and navigate sites. The study developed an auditing methodology examining data collection practices, server-side profiling, and personalization behaviors. They found assistants collect far more granular behavioral data than users expect, with unclear retention and usage policies.

Caveats: Audit methodology relies on observable network traffic and disclosed policies; server-side processing and actual data usage remain partially opaque.

Reflections: Can we build effective GenAI assistants with on-device processing only? · What's the minimum browsing context needed for useful assistance versus privacy-invasive profiling? · How do users' mental models of 'assistant' versus 'tracker' diverge from actual system behavior?