The As-If Machine: Personalized counterfactuals make distant policy risks feel close

Joshua Ashkinaze, Yubo Zhou, Pengyu Chen, Avi Bhattacharya, Hyebin Park, Yanman Zhang, Stephanie Preston, Eric Gilbert, Ceren BudakarXiv 2026 · arXiv

TL;DR

The "As-If Machine" (AIM) is a multi-agent AI system that helps citizens co-write personalized stories about how abstract policy risks—like satellite cyberattacks or science funding cuts—would directly impact their own lives. By simulating these "counterfactual" futures, the system significantly increases "psychological closeness," making distant societal risks feel immediate and personally relevant, which in turn can shift policy attitudes.

What Problem Does This Paper Address?

We only live one life, which limits our ability to personally experience the consequences of complex policy decisions. This creates a "psychological distance" problem: if you haven't personally felt the sting of a failing infrastructure or a funding shortage, you are less likely to prioritize policies that prevent those risks.

While narratives (stories) are known to bridge this gap, they are traditionally difficult to personalize at scale. A generic story about a power outage might not move you, but a story about your specific commute being disrupted might. The authors explore whether generative AI can solve this by acting as a "personalized simulation engine," helping users construct concrete mental models of unfamiliar futures.

Methods

The researchers developed the As-If Machine (AIM), a multi-agent system built on Large Language Models (LLMs). The system follows a four-stage pipeline to ensure "narrative transportation" (immersion):

  1. Initializer: Takes user intake data (job, location, routines) to write a plausible "onset" for the scenario.
  2. Predictor: Forecasts specific consequences for that individual.
  3. Storyteller: Connects the onset and consequences into a coherent narrative arc.
  4. Autocompleter: A real-time writing assistant that offers short suggestions as the user writes their story, maintaining the causal structure while letting the user lead.
The AIM system architecture and user interface.

The authors tested AIM in two pre-registered randomized experiments (N=1,007 and N=894). They compared AIM against two controls: reading an expert-written vignette on the same topic and a "placebo" version where users co-wrote with AIM about an unrelated topic (like meeting a celebrity).

Datasets

The study did not use a static external dataset for training but rather generated personalized content based on:

  • Participant Intake Data: Real-time responses from users about their daily lives, occupations, and locations.
  • Policy Scenarios: Two specific domains were used for evaluation: Science Funding (focused on NIH training grants and specialist wait times) and Satellite Cybersecurity (focused on disruptions to GPS, banking, and communications).

Key Insights and Ideas

  • Division of Narrative Labor: Linguistic analysis of over 23,000 interactions showed a clear split: the AI supplied the "scaffold" (concrete physical details and causal structure), while humans supplied the "soul" (first-person perspective, emotions, and interiority).
  • Closeness vs. Immersion: The primary driver for changing attitudes wasn't just being "lost in a story" (transportation), but specifically psychological closeness—the feeling that the issue is relevant to "me."
  • The Crystallization Effect: AIM was significantly more effective at changing attitudes in the satellite cybersecurity domain than in science funding. The authors suggest this is because science funding attitudes are more "crystallized" (fixed by ideology and media coverage), whereas satellite risk is a relatively "new" and less polarized issue for most people.
  • AI Skepticism Doesn't Block Efficacy: Surprisingly, the intervention worked even for participants who reported unfavorable views of AI. The "As-If" experience was powerful enough to overcome initial algorithm aversion.

Key Results

  • Psychological Closeness: AIM increased closeness by 0.36 to 0.38 standard deviations (SDs) compared to expert vignettes, and by 0.80 to 0.90 SDs compared to unrelated writing.
  • Policy Attitudes: In the satellite cybersecurity study, AIM shifted policy support by 0.54 SDs over the vignette. In the science funding study, the effect was smaller (0.16 SDs), likely due to the crystallization mentioned above.
Treatment effects on psychological closeness and policy attitudes across both studies.
  • Distributional Impact: Quantile regressions showed that AIM was most effective at "raising the floor"—it primarily helped people who initially felt the issue had nothing to do with them.

Limitations and Future Work

  • The Specificity-Credibility Tradeoff: While concrete details make stories vivid, they also make them "contestable." If the AI suggests a 14-month wait for a doctor and the user finds that unrealistic, they may reject the entire narrative.
  • Episodic vs. Thematic Framing: Because AIM focuses on personal "episodes," it can sometimes leave the broader political or institutional causes feeling abstract. Some participants reported being so focused on their "personal predicament" that they lost sight of the policy.
  • Negative Bias: Both tested scenarios were negative futures. Future work could explore whether "positive counterfactuals" (simulating the benefits of a policy) are equally persuasive.
  • Persistence: The study measured immediate post-task attitudes; it remains to be seen how long these "personalized simulations" stick in a user's mind.

Summary updated

Abstract

Life experiences can shape policy attitudes, but we only live one life. Citizens may therefore fail to support policies addressing risks they have never personally experienced. Narratives can make such risks immediate and personal, but personalized stories remain difficult to produce at scale. Here we ask whether AI can make distant scenarios feel concrete by simulating a participant's life as if that scenario occurred. We present the As-If Machine (AIM), a multi-agent system through which people co-write personalized counterfactual narratives---stories about their own lives under different policy futures. In two pre-registered and quota-matched experiments on science funding (N=1,007) and satellite cybersecurity (N=894), we compared AIM with two controls: reading an expert-written vignette on the same topic and co-writing with AIM about an unrelated topic. AIM increased psychological closeness---how close a risk feels to the self---by nearly identical margins in the two domains (0.36 and 0.38 standard deviations (SDs) against the vignette, 0.80 and 0.90 against unrelated writing). AIM increased psychological closeness even among participants with unfavorable views of AI. For policy attitudes, AIM's effects were 0.16 SDs against the vignette and 0.10 SDs (n.s.) against unrelated writing for science funding; the corresponding effects for satellite cybersecurity were 0.54 and 0.67 SDs. Across 23,496 human-AI autocomplete interactions, participants and the system played distinct roles: AIM supplied concrete detail that grounded the scenario, and participants supplied first-person perspective and emotion. Co-writing with the As-If Machine makes distant policy risks feel close.
BibTeX
@misc{ashkinaze2026as,
  author = {Joshua Ashkinaze and Yubo Zhou and Pengyu Chen and Avi Bhattacharya and Hyebin Park and Yanman Zhang and Stephanie Preston and Eric Gilbert and Ceren Budak},
  title = {{The As-If Machine: Personalized counterfactuals make distant policy risks feel close}},
  year = {2026},
  eprint = {2610.33301},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.33301},
}
Figure