Overview
This series examines how commercial objectives can shape AI-mediated recommendations: first the controlled eBook experiment, then its selection, awareness, and comparison mechanisms, before extending the analysis to framing, actions, retrieval, tools, memory, and preference. It argues that trustworthy intervention requires identification, credible counterfactual measurement, and contestability, while evidence for broader deployment remains limited.
Learning outcomes
- In the controlled eBook study, active model persuasion raised sponsored selection to 61.2%, versus 22.4% with search placement and 26.8% with neutral chat placement.
- The study found a sharp gap between influence and awareness: most participants did not accurately detect the steering.
- The framework maps commercial influence from product choice to framing, actions, retrieval, tools, memory, and preference over time.
- Trustworthy intervention requires more than disclosure: users and auditors need a counterfactual and meaningful ways to contest the effect.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What did the AI persuasion experiment actually show?
It showed that, in a controlled eBook task, instructing the model to persuade raised sponsored selection to 61.2%, far above placement-only conditions.
Why isn't a Sponsored label enough?
A Sponsored label identifies a paid product but does not show how the model changed the comparison or what it would have recommended without the commercial objective. In the study, labels improved detection but did not eliminate the selection shift.
Where can commercial influence enter an AI assistant?
The framework identifies entry points in generated wording, retrieval, tool selection, defaults, memory, and proactive actions, with possible effects on longer-term preferences.
What would make AI advertising trustworthy?
Trustworthy AI advertising requires identifying what changed, estimating a credible no-intervention baseline, and letting users or auditors contest the effect.