New analysis finds New York City rideshare users under-compare prices—leaving roughly $300m a year on the table—while LLM mediators can produce context-aware interventions in heated online discussions but deliver only modest attitude shifts, and workers’ beliefs about AI are varied and malleable yet show little behavioral response in initial pilots.

Main Finding

Jeffrey Fossett’s dissertation (Harvard PhD, 2026) studies how people, firms, and communities make choices inside contemporary sociotechnical systems. It contains three empirical projects with common themes (frictions, information, and behavior in digital environments):

• Chapter 1 (Ridesharing): Consumers compare prices between Uber and Lyft far less than canonical search models predict (observed comparison rate ≈ 16.1% conditional on opening one app; model predicts >90%), generating large aggregate inefficiencies — a back‑of‑the‑envelope estimate implies NYC riders collectively leave ≈ $300M/year (≈ 6% of platforms’ gross booking volume) on the table by not comparing.
• Chapter 2 (LLM mediation): An LLM-powered mediation tool can generate context-sensitive, theoretically grounded interventions (paraphrasing, finding agreement, encouraging perspective-taking) in contentious online conversations, but the randomized trial shows limited effects on participants’ attitudes toward disagreement partners.
• Chapter 3 (Worker beliefs about AI): Workers’ beliefs about the labor-market impact of generative AI/LLMs vary substantially; beliefs are somewhat correlated with self-reported adaptation behaviors (training, degree enrollment, workplace LLM use) but evidence is noisy; exposure to contrasting expert narratives causally shifts beliefs, though pilot data show no clear behavioral change yet.

Key Points

• Price search friction matters even when technological comparison is easy. Measured price dispersion between Uber and Lyft is substantial for identical trips, yet most users do not check the rival app. Small individual frictions aggregate to large surplus transfers away from consumers.
• A simple sequential search model calibrated to observed price gaps and benchmark value-of-time estimates greatly overstates observed search/comparison; the gap suggests behavioral or institutional barriers beyond classical search-cost explanations (e.g., attention, habit, app friction, platform design or terms restricting comparison).
• LLMs can implement mediation strategies at scale (automatic paraphrase, identify common ground, encourage perspective-taking). The tool produced context‑sensitive messages and successfully identified many intervention opportunities, but did not meaningfully change measured attitudinal outcomes in the trial (partial sample; full-sample results pending).
• Workers’ beliefs about AI impacts are heterogeneous and malleable. Short, contrasting expert narratives shift beliefs; this highlights the role of information channels and narratives (not only technical progress) in shaping adaptation decisions.
• Across chapters, the dissertation emphasizes that sociotechnical structures and information environments — and the small frictions that persist within them — materially shape market and social outcomes.

Data & Methods

• Chapter 1 (Ridesharing)

  • Data: Synchronized field price audit of Uber and Lyft for identical NYC trips; device-level app-usage data (to observe whether users opened both apps and how often); benchmark evidence on consumer search costs and value-of-time from literature.
  • Methods: Compute price dispersion and conditional comparison rates; calibrate a simple sequential search (optimal stopping) model to the empirical price-gap distribution and benchmark search-costs; compare model-predicted comparison frequency to observed behavior; back-of-envelope aggregation to estimate consumer foregone surplus in NYC.
  • Key empirical facts: Conditional on opening one app, only ~16.1% of the time the other app is opened the same day; calibrated model predicts >90% comparison.
  • Institutional observation: API/terms restrictions (e.g., Uber API terms) can inhibit automated price-comparison tools.

• Chapter 2 (LLM mediation experiment)

  • Setting: Online dyadic conversations on a contentious political topic (study recruitment via Prolific and partnerships with organizations). Used Perspective API to detect candidate conversations for intervention.
  • Intervention: An LLM system, guided by conflict-resolution principles, decides whether to intervene and generates short mediation messages (paraphrase, identify agreements, ask perspective-taking prompts). Intervention decision and content were automated via LLM prompts (full prompts in appendix).
  • Design: Randomized controlled trial where conversations flagged as candidate interventions were randomized to receive (or not receive) LLM mediation; analysis includes ITT, IV/LATE (instrumenting receipt by assignment among flagged conversations), heterogeneity checks, intervention-validation (did the tool identify correct moments? did it produce valid paraphrases?).
  • Outcomes: Measures of perceived productive disagreement, democratic reciprocity, affective attitudes (feeling thermometers), plus conversational metrics (length, toxicity scores). The system generated plausible interventions but showed limited downstream attitude change; appendices document intervention examples, validation checks, attrition, and measurement details.

• Chapter 3 (Worker beliefs about AI)

  • Data: Survey and randomized experiment targeting workers in two professional fields (law and management consulting); pilot sample (results preliminary).
  • Design: Belief elicitation about AI impacts (employment, task mix, efficiency, quality) using Likert and graphical trend interfaces; a randomized exposure to short expert-narrative videos presenting contrasting views (high-impact vs low-impact; e.g., Brynjolfsson vs Acemoglu transcripts included).
  • Outcomes: Self-reported adaptation behaviors (AI training, degree enrollment, workplace LLM adoption), belief measures, and causal effects of narrative treatments on beliefs and behavior intentions.

Implications for AI Economics

• Small frictions in digital markets can create large aggregate welfare effects. Even when comparison is technologically easy (two dominant apps on the same device), behavioral/institutional frictions can sustain price dispersion and transfer surplus to platforms. For AI economics this underscores that platform design, user attention, onboarding/habits, and information architecture — not only algorithmic pricing — determine distributional outcomes.
• Policy levers and platform governance matter. The findings support arguments for policy interventions (price‑transparency requirements, APIs that facilitate comparison, consumer nudges) to restore competitive pressure where comparison frictions persist. Restrictions in platform terms-of-use or proprietary APIs can lock in frictions even when technology could reduce them.
• LLMs have operational promise but limited immediate social‑attitudinal impact when used as automated mediators. The dissertation shows LLMs can reliably generate mediation content and detect candidate moments at scale, but changing entrenched attitudes or affective responses is harder. For AI economists assessing value from LLM-based social tools, expect meaningful engineering/implementation and behavioral design challenges; cost-effectiveness depends on whether downstream outcomes (e.g., reduced polarization, better deliberation) can be produced at scale and sustained over time.
• Beliefs and narratives are economically consequential. Workers’ heterogeneous and malleable beliefs about AI impacts influence their adaptation choices (training, tool adoption). This implies that public narratives, expert communication, and policy signaling can shift investment in human capital and thereby alter labor-market equilibria. Models of AI-driven labor-market transitions should incorporate belief dynamics and information interventions as mechanisms influencing adoption and retraining uptake.
• Research design takeaways for AI economics:
  • Combining synchronized audits, behavioral measurement (device usage), and calibrated structural models is a powerful approach for quantifying frictions in digital markets.
  • Randomized trials that use LLMs both as intervention designers and executors allow testing scalable social‑technology interventions, but rigorous validation (first-stage checks, content validation, heterogeneity) is critical.
  • Belief‑manipulation experiments (short expert narratives) can identify causal shifts in expectations that may precede longer-run behavioral change; longer follow-up and field-level outcome measurement will be necessary to assess economic impacts fully.

Limitations & next steps (noted by author) - Chapter 2 and Chapter 3 report partial/pilot samples; full-sample results may differ. Effects on real-world behavior (vs survey outcomes) require longer horizons and larger samples.
- The rideshare calibration abstracts from some real-world complications (loyalty, multi‑attribute choice, time costs beyond simple VOT); further work could incorporate richer behavioral models and field interventions (nudges, transparency tools) to causally change search behavior.

Overall contribution - The dissertation offers empirical evidence that modest frictions in digital systems (search costs, attention, narrative frames) have outsized economic and social consequences, and it demonstrates practical multimethod approaches (field audit + behavioral data + structural calibration; LLM-driven RCTs; belief‑manipulation experiments) that are directly relevant to AI economics and policy design.