AI that proposes alternatives and synthesizes (Dialectical Inquiry) yields larger informational gains but raises users' mental workload, while critique-only assistants deliver smaller gains with lower cognitive cost; targeted strategy training lessens the cognitive burden and improves net outcomes.

Main Finding

This paper proposes and tests a dual-pathway model of Human–AI strategic decision-making comparing (1) tool-shaping (embedding conflict techniques in AI bots) and (2) mind-shaping (training users to apply those techniques when using a general AI). It predicts a benefit–cost trade-off: conflict-inducing interventions (Devil’s Advocate, DA; Dialectical Inquiry, DI) increase information elaboration (a cognitive benefit) but also increase cognitive load (a cost). DI is expected to produce the largest gains in elaboration and the largest cognitive costs (DI > DA > Info-only). Users’ cognitive flexibility is expected to amplify benefits and attenuate costs. Two between-subject experiments are designed to test these mechanisms and moderated mediation.

Key Points

• Research gap: prior AI-supported constructive conflict work has focused mainly on AI Devil’s Advocate; DI (generating competing plans and synthesis) is understudied and may perform differently.
• Dual interventions:
  • Tool-shaping: specialized AI bots (Info-only, DA bot, DI bot, vs no-AI control).
  • Mind-shaping: user strategy training (DA training, DI training, vs control) used with a general-purpose chatbot.
• Theoretical mechanisms:
  • Benefit pathway: information elaboration (expected to improve decision quality and innovativeness).
  • Cost pathway: cognitive load (can undermine decision outcomes if excessive).
  • Moderator: cognitive flexibility (higher flexibility → stronger elaboration gains and reduced cognitive load).
• Hypotheses summary:
  • H1: More sophisticated conflict technique → greater information elaboration (No AI < Info-only < DA < DI).
  • H2: More sophisticated technique → greater cognitive load (No AI < Info-only < DA < DI), and cognitive load negatively affects outcomes.
  • H3: Cognitive flexibility moderates both paths (boosting benefits, reducing costs).
• Practical emphasis: comparison of shaping the tool vs shaping the mind from a sociotechnical perspective.

Data & Methods

• Design: Two independent between-subject experiments that test the same dual-pathway model while shifting the intervention locus.
  • Study 1 (tool-shaping): 4 conditions — Control (no AI), Info-only Bot, DA Bot, DI Bot.
  • Study 2 (mind-shaping): 3 conditions — Control (basic AI intro), DA training, DI training; all use same general-purpose chatbot.
• Participants: Professionals recruited via Prolific with ≥3 years management experience and involvement in strategic decisions.
  • Power targets: Study 1 N ≥ 200 (target >180), Study 2 N ≥ 160 (target >159).
• Task: 30-minute supply-chain diversification strategic decision task (moderate difficulty). Study 2 includes a 15-minute training module pre-task.
• Experimental materials / implementation: Bots implemented with Claude Sonnet 3.7 via Anthropic API; role-constrained system prompts enforce Info-only / DA / DI behaviors.
• Measures:
  • Mediators: Information elaboration (adapted Kearney et al., 2009); Cognitive load (NASA-TLX).
  • Moderator: Cognitive Flexibility Scale (Martin & Rubin, 1995), measured pre-task.
  • Outcomes: Decision quality and innovativeness — rated by two independent experts blind to condition.
  • Manipulation checks and baseline cognitive load measures included (baseline before task; in Study 2 baseline measured after training).
• Analysis: Mediation and moderated mediation tested via PROCESS (Hayes) Models 4 and 7; exploratory cross-study comparisons of effect-size patterns.
• Current progress: Three bots (Info-only, DA, DI) prototyped with Claude Sonnet 3.7 and Anthropic API; role prompts encode allowed/forbidden behaviors.

Implications for AI Economics

• Investment trade-offs — tool development vs. human capital:
  • Organizations must decide whether to allocate resources toward building specialized AI conflict agents (tool-shaping) or toward workforce training (mind-shaping). The paper frames this as a ROI trade-off where DI-style systems may raise decision quality and innovation but also impose higher cognitive costs and potentially higher training needs.
• Complementarities and heterogeneity:
  • Returns to AI augmentation will be heterogeneous across workers. Cognitive flexibility functions like a skill complement — firms with a more cognitively flexible workforce will capture more of the AI-induced benefit and suffer less cognitive-cost leakage. This has implications for hiring, screening, and skill-development investments.
• Marginal productivity and task redesign:
  • DI bots could raise marginal productivity in strategic tasks (better innovation and decision quality) but may require changes in task design (e.g., more time per decision, different workflows) to manage cognitive load. Firms should account for these operational costs when estimating productivity gains.
• Allocation of scarce cognitive resources and scaling:
  • Cognitive load is an economic cost (time, performance risk, potential need for remedial processes). Scaling DI-style interventions across many decision-makers could amplify aggregate cognitive costs; firms should evaluate per-decision cost–benefit and consider targeted deployment where payoff is highest.
• Organizational strategy and compensation:
  • If DI increases innovation, it may produce greater firm-level value in uncertain/high-stakes environments; compensation and incentive systems might be adjusted to encourage use where net benefits are largest.
• Policy, governance, and procurement:
  • Procurement and regulation should consider not just measurable decision outcomes but also cognitive burdens on workers. Audits of AI decision aids should track both quality improvements and user cognitive load. Standards could require reporting on both benefit and cognitive cost metrics.
• Labor market impacts:
  • Differential returns by cognitive flexibility may accelerate skill-biased technological change — increasing demand (and wages) for workers who can productively use DI-style AI. Training investments will influence labor supply responses and inequality in AI adoption gains.
• Measurement and evaluation:
  • Economic evaluations of AI tools should include mediating psychological costs (cognitive load) and heterogeneity (cognitive flexibility) rather than only aggregate performance metrics. Expert-rated decision quality and innovation are meaningful but should be supplemented with downstream performance/firm-value measures in future work.
• Limitations relevant for economic interpretation:
  • Results are based on experimental, task-level measures with managerial participants and expert ratings; general equilibrium and firm-level payoff implications require further field validation and longitudinal study.

Overall, this paper provides a framework and experimental strategy to quantify the benefit–cost trade-offs of different AI-supported conflict techniques and to inform organizational decisions about whether to invest in specialized AI tooling or in training human decision-makers — an applied microeconomic question central to AI-driven productivity and workforce policy.