A configurable AI teammate can steer team behavior: in a six-session classroom trial, a cognitive-scaffolding persona raised contributions and linguistic alignment, whereas a socially-supportive persona warmed team climate and lowered over-reliance.

Main Finding

A configurable AI teammate—implemented as a reproducible design object combining personality (Big Five persona), selective participation, and dual memory—can causally shape team dynamics over time. In a six-session classroom deployment (≈51 students), switching the AI’s persona produced a clean double dissociation: a cognitive‑scaffolding persona increased perceived contribution and linguistic alignment with the AI, while a socially‑supportive persona improved team climate and reduced over‑reliance.

Key Points

• TRAIL (Team Research and AI Integration Lab) is a web platform that treats an AI teammate as a reproducible, configurable design object.
• Core design levers:
  • Big Five persona assignment (personality).
  • Selective‑participation message pipeline (controls when and how often the AI speaks; kept the AI a stable minority participant).
  • Dual memory system (presumably short‑term and long‑term memory to support continuity).
  • Chained longitudinal experiments (multiple sessions with persistent configuration).
  • Export‑ready analytics (enables downstream analyses, e.g., AI–human linguistic similarity).
• Deployment: real classroom setting, six sessions, ~51 student participants; sustained, real‑time collaboration across sessions.
• Experimental manipulation: single blind persona change (cognitive‑scaffolding vs socially‑supportive).
• Main empirical results (double dissociation):
  • Cognitive‑scaffolding agent: higher ratings of AI contribution and closer linguistic/textual alignment between students and the AI.
  • Socially‑supportive agent: warmer reported team climate and lower measures of participant over‑reliance on the AI.
• Measurement innovations: maintained AI presence as a controlled minority of discourse; used exportable logs to compute text‑similarity (linguistic alignment) and other analytics.

Data & Methods

• Platform: TRAIL—web-based, instrumented collaboration environment that makes AI teammate configurations reproducible and chainable across sessions.
• Design features implemented:
  • Persona module using Big Five traits to generate consistent AI behavior.
  • Message pipeline to selectively inject AI messages (timing/rate control).
  • Dual memory to sustain context across interactions and over time.
  • Logging/export of full conversation transcripts and metadata for offline analysis.
• Experimental setup:
  • Participants: ≈51 students in a classroom course.
  • Duration: six collaborative sessions per team (longitudinal chaining).
  • Blinding: single blind persona change so participants did not know they were being switched between persona conditions.
• Outcomes and measures:
  • Subjective ratings (perceived AI contribution, team climate, reliance).
  • Behavioral measures (AI share of messages held to a stable minority).
  • Linguistic/textual measures (AI–human text similarity / alignment computed from exported logs).
• Analysis: compared outcomes across persona conditions and interpreted pattern as a double dissociation between cognitive vs social persona effects.

Implications for AI Economics

• Productization and feature valuation:
  • AI teammate design is an economically meaningful product axis—personality and participation-control are features that affect worker behavior and team outputs. Firms should treat these as monetizable/product design choices rather than incidental attributes.
• Complementarity vs substitution:
  • Different persona designs produce different complementarity channels. Cognitive‑scaffolding increases alignment and perceived productivity (augmentation), while socially‑supportive designs change coordination and reliance patterns (affecting substitution risk and moral hazard).
• Labor productivity and skill formation:
  • Socially supportive agents reduce over‑reliance, which may preserve human skill formation; cognitive agents that increase alignment might raise short‑term productivity but could risk longer‑term deskilling—important for cost‑benefit models of adoption.
• Organizational design and incentives:
  • Employers should consider persona as part of task allocation, monitoring, and incentive systems. For tasks requiring autonomous judgment, socially supportive/limiting participation agents may be preferable; for tasks needing structured cognitive support, scaffolding personas may increase throughput.
• Measurement, causal inference, and policy:
  • TRAIL demonstrates an experimental infrastructure to estimate causal effects of AI design on human behavior. Policymakers and researchers can use similar reproducible setups to quantify externalities (e.g., over‑reliance, coordination failures) and to design evidence‑based regulation or workplace guidelines.
• Pricing and adoption models:
  • Willingness to pay and adoption decisions may depend on persona effects—firms and vendors should segment offerings (e.g., “productivity‑scaffold” vs “team‑wellness”) and measure downstream economic impacts (errors avoided, time saved, turnover, training costs).
• Externalities and market outcomes:
  • Widespread deployment of personality‑tuned AI teammates could shift equilibrium behaviors (coordination norms, trust networks) with aggregate effects on labor markets—both positive (efficiency gains) and negative (skill erosion, increased dependence).
• Research agenda:
  • Need for larger, longer, and field‑scale studies linking persona/configuration to hard economic outcomes (productivity, error rates, wages, employment), heterogeneous team contexts, and optimal contract or compensation responses.

Limitations to note (relevant for economic inference): small classroom sample (~51), single blind persona switch, specific task/context, and relatively short horizon (six sessions). These constrain external validity; scaling TRAIL‑style experiments in workplaces will be necessary to quantify macroeconomic effects.