AI Assistance Optimizes for Now, Erases Capability for Later

Regulators mandate AI disclosure, but nobody knows if labeling a chatbot as 'AI' actually changes its persuasive power. The assumption is that transparency matters—but which kind?

Method: A preregistered experiment with 1,500 UK adults tested two disclosure types against a persuasive chatbot. Disclosing AI identity alone had no effect: the chatbot shifted attitudes by 13.1 points versus 12.6 points in the control. But disclosing persuasive intent and instructions cut persuasion in half to 6.3 points. Participants also judged the campaign's methods as less acceptable and supported stronger penalties.

Caveats: Tested on policy attitudes in short conversations. Longer interactions or different domains may differ.

Reflections: Does intent disclosure remain effective after repeated exposure, or do users habituate? · Would disclosing training objectives (e.g., 'optimized for engagement') produce similar effects? · How does intent disclosure interact with user trust in the deploying organization?

Visualization design principles assume users are focused. But real users monitor dashboards while answering Slack messages. Does divided attention change which design choices matter?

Method: Two experiments tested expectation-aligned versus expectation-violating color and spatial mappings under single-task and dual-task conditions. Divided attention amplified the performance gap between aligned and violating designs, affecting response time, accuracy, and the ability to produce judgments under time constraints. A Linear Ballistic Accumulator model showed that expectation violations slow evidence accumulation, and concurrent tasks compound this cost.

Caveats: Tested specific color and spatial mappings. Other design dimensions may behave differently under divided attention.

Reflections: Do training or repeated exposure reduce the divided-attention penalty for unconventional mappings? · Which other visualization design choices (e.g., chart type, annotation style) show amplified effects under multitasking? · Can adaptive interfaces detect divided attention and adjust design conventions in real time?

TikTok research is booming, but researchers use different data collection tools without knowing if they're measuring the same platform. Do the tools produce comparable samples?

Method: Systematic comparison of TikTok Research API, Pyktok, and Apify across five endpoints found substantial cross-tool differences, especially for hashtag and keyword searches. The Research API uses back-end calls; Apify and Pyktok scrape the front-end, producing systematic differences in time periods and popularity levels. Only the user endpoint yielded consistent results across tools. None can guarantee random or randomized samples.

Caveats: Comparison conducted at a single time point. Tool behavior and platform APIs evolve rapidly.

Reflections: Can hybrid approaches combining multiple tools reduce sampling bias? · How do algorithmic recommendation differences across tools affect longitudinal studies? · What ground-truth benchmarks could validate TikTok data collection methods?