Conditional Endorsement: Simulating How Adults with Different Skills Evaluate Canada's National AI Strategy

Dipto Das, Felix Leduc, Shion GuhaarXiv 2026 · arXiv

TL;DR

Researchers used 3,200 survey-grounded generative agents to simulate how diverse populations evaluate Canada’s national AI strategy. They discovered a phenomenon of "conditional endorsement": while broad support for AI policy is nearly universal, adults with low digital skills remain deeply skeptical that these policies will actually reach or benefit them without specific, non-digital implementation safeguards.

What Problem Does This Paper Address?

National AI strategies, such as Canada’s 2026 AI for All plan, often make grand commitments to "AI literacy" and "inclusive adoption." However, the people these policies most need to reach—the digitally marginalized—are often the least visible in the public consultations that shape them.

The authors address the difficulty of anticipating how adults with uneven skills and life circumstances will receive a policy before it is implemented. They seek to move beyond aggregate "public opinion" to understand how specific barriers (like the digital divide) create a gap between supporting a policy's goals and believing in its personal utility.

Methods

The study employs Generative Agent-Based Modeling (GABM), a technique where Large Language Models (LLMs) are prompted to act as specific personas.

  1. Persona Construction: The authors built 3,200 agents grounded in the PIAAC Cycle 2 survey data. These personas include variables for demographics, education, employment, and—crucially—levels of digital technology use.
  2. Single-Agent Evaluation: Each agent (instantiated via GPT-5-mini) rated 20 specific policies from the Canadian AI strategy on three 6-point scales: Support, Effectiveness (will it improve literacy?), and Personal Benefit.
  3. Dyadic Dialogues: To move beyond static ratings, the researchers simulated 1,533 "dyadic" conversations between agents with contrasting views. This allowed the agents to qualify their support and specify implementation requirements.
  4. Computational Analysis: The team used LLooM (a tool for inductive concept analysis) to extract themes from the text and fine-tuned an XLM-RoBERTa model to predict rating changes across the turns of the dialogues.

Datasets

  • PIAAC Cycle 2 (Canadian Public-Use Files): The OECD’s Programme for the International Assessment of Adult Competencies. This provided the "ground truth" for adult skills (literacy, numeracy, tech use) used to anchor the generative agents.
  • Canada’s National AI Strategy (AI for All, 2026): The source text for the 20 policies and 9 "landscape challenges" (e.g., public trust, employment disruption) evaluated by the agents.

Key Insights and Ideas

  • The Support-Benefit Gap: There is a massive divergence between public endorsement and personal expectation. Agents supported almost every policy (99.5%) but were far less confident that they would personally benefit (56.4%) or that the policies would be effective (35.9%).
  • The "Never-User" Paradox: Personas modeled on adults who had never used a digital device still endorsed the aims of AI policies but expected zero benefit. This suggests that "support" in surveys can be a misleading metric for policy success.
  • Dialogues as Specification, Not Persuasion: In the simulated debates, agents rarely changed their minds. Instead, they used the interaction to specify delivery conditions. For example, a skeptical agent wouldn't be "convinced" by a pro-AI agent; instead, both would eventually agree that the policy only works if there is a "phone helpline" or "in-person support."
  • Three Pillars of Conditional Endorsement: The agents' reasoning centered on three requirements:
  • Participation: Addressing hidden costs like time, childcare, and device access.
  • Exercisability: Ensuring rights (like opting out) are available via non-digital channels (paper forms, staffed lines).
  • Control: Moving from "consultation" to community-governed funds and data sovereignty.

Key Results

  • Digital Use Gradient: Expected benefit fell sharply with digital skill levels. 63.4% of high-tech users expected to benefit, compared to only 12.3% of those who had never used a digital device.
  • Policy Specificity: Policies with explicit educational components (like free introductory AI training) were the only ones where low-skill personas expected more benefit than high-skill ones (94.8% vs 82.9%).
  • Consensus on Non-Digital Routes: In 91.6% of dialogues, agents opened with agreement phrases, quickly moving to co-designing "on-ramps" such as local library programs and "digital ambassadors."

Limitations and Future Work

  • Synthetic Personas: The authors acknowledge that agents are proxies. The "skill gradients" observed might partially reflect the LLM's internal stereotypes or assumptions about what "low digital use" means.
  • LLM Sycophancy: The tendency for LLMs to agree with their interlocutors may have suppressed genuine adversarial disagreement in the dialogues.
  • Validation: The study highlights the need to validate these simulated findings against real-world responses from marginalized groups.
  • Future Directions: The authors suggest applying this GABM framework to other jurisdictions and using "Retrieval-Augmented Generation" (RAG) on actual community town hall transcripts to further ground the agents in real human discourse.

Summary updated

Abstract

Canada's 2026 national AI strategy, AI for All, commits to free AI training for everyone and to expanding business AI adoption, yet it is difficult to anticipate whether these policies will reach adults with uneven skills and circumstances. We ask how adults with diverse skills might rate these policies on support, effectiveness, and personal benefit, how their skills and circumstances shape these evaluations, and what implementation conditions they raise. We built 3,200 generative agents from a weighted sample of the Programme for the International Assessment of Adult Competencies (PIAAC) Cycle 2 respondents, collected their ratings and explanations for 20 policies, and analyzed dialogues between agents with contrasting ratings. The agents supported nearly every policy but were far less confident that they would improve adult AI literacy or benefit them personally. This gap widened as digital technology use declined: agents modeled on adults who had never used a digital device still endorsed the policies but rarely expected to benefit. Agents traced the gap to participation costs, hard-to-exercise rights, and limited community control, and their dialogues specified these conditions rather than resolving disagreement. We describe this pattern as conditional endorsement and argue that AI literacy policy should be designed and evaluated for adults with the fewest digital skills.
BibTeX
@misc{das2026conditional,
  author = {Dipto Das and Felix Leduc and Shion Guha},
  title = {{Conditional Endorsement: Simulating How Adults with Different Skills Evaluate Canada's National AI Strategy}},
  year = {2026},
  eprint = {2611.19111},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19111},
}
Figure