TL;DR
This paper introduces Runtime Action Interference (RAI), a post-inference control mechanism that regulates an AI's action pacing and filters out toxic behaviors without altering its underlying policy parameters. By deploying RAI in a StarCraft II human-subject study, the researchers discovered a striking psychological twist: simply disclosing the AI's high capabilities to players drastically reduced their perception of fairness and increased perceived toxicity, even though the AI's actual behavior remained completely identical.
What Problem Does This Paper Address?
When we train reinforcement learning (RL) agents for complex real-time environments like StarCraft II, the policy network outputs raw action proposals. However, what users actually experience is mediated by the deployment stack—schedulers, rate limiters, and interface constraints. Standard approaches to AI safety and behavioral control usually involve retraining the model or shaping rewards, which is computationally expensive and often inflexible.
Furthermore, there is a major gap in how we evaluate human-AI interactions: we often fail to separate the actual behavior of the AI from the user's expectations based on how that AI is framed. This paper tackles both challenges by providing an external, auditable runtime control layer and investigating how capability disclosures alter human perceptions of fairness, trust, and toxicity.
Methods
The authors propose Runtime Action Interference (RAI), a deterministic admission rule inserted directly between policy inference and environment dispatch. RAI leaves the trained policy parameters completely fixed, making it easy to plug into existing architectures like AlphaStar.
RAI filters actions using two orthogonal checks:
- Timeframe Eligibility: A stateful cooldown registry that regulates action timing and limits Actions Per Minute (APM).
- Content Detector: A rule-based filter that scrutinizes the action against historical context to flag prohibited behavioral paradigms (e.g., worker-unit harassment).
If a proposed action fails either check, RAI vetoes it and dispatches a harmless no-op instead.
To evaluate the interplay between runtime control and human perception, the authors conducted a user study where human players faced a high-capability AlphaStar agent regulated by RAI. The trick? One group of players had the AI's high-capability status withheld, while the other group had it explicitly disclosed.
Datasets
The primary evaluation relies on a newly collected dataset of 32 post-match human responses stratified by player expertise:
- Novices: New to StarCraft II.
- Intermediates: Top 50% regionally in RTS games.
- Experts: Top 10% regionally.
The focal analysis zeroes in on 23 of these responses to compare the "Claim Withheld" ($n=10$) and "Claim Disclosed" ($n=13$) conditions.
Additionally, the authors experimented with an ancillary dataset of synthetic player personas generated via a large language model (gemini-exp-1206) to predict player reactions during the study design phase, though they explicitly separate these model-generated outputs from the actual human data.
Key Insights and Ideas
- The Execution Stack is a Viable Control Point: You don't need to retrain a massive RL model to enforce behavioral guardrails. Interceding at the terminal output layer provides a lightweight, auditable way to enforce safety and pacing.
- The Transparency Paradox: Total transparency isn't always a psychological win. Telling players they were up against a "superhuman AI" made them hyper-aware of its speed and efficiency, leading them to perceive the exact same gameplay as less fair and more hostile.
- Trust $\neq$ Fairness: Interestingly, disclosing the AI's high capability caused trust to shoot up among novices and experts (who respected the AI's competence), even as their ratings for fairness plummeted. This proves that trust, fairness, and toxicity are distinct dimensions of human experience that must be measured separately.
Key Results
The study measured player responses on a scale from 1 (lowest) to 5 (highest). When the AI's capability claim was disclosed, the pooled metrics shifted dramatically compared to when it was withheld:
- Fairness dropped from a pooled mean of 3.90 (withheld) to 2.62 (disclosed).
- Toxicity rose from a pooled mean of 2.00 (withheld) to 2.85 (disclosed).
- Trust increased overall from 3.50 (withheld) to 4.31 (disclosed), driven heavily by novices (+1.30) and experts (+1.35), though it actually decreased among intermediate players (-0.75).
Limitations and Future Work
The authors openly acknowledge that due to the small sample sizes within individual expertise cells, this study is descriptive and exploratory rather than a definitive causal proof.
Furthermore, because the human study held the RAI controller constant across both focal groups, it isolates the impact of capability framing but does not measure the direct causal impact of turning RAI on versus off. The authors suggest that future work should utilize a full factorial design—varying both the RAI guardrails and the disclosure scripts independently—to fully untangle how runtime interventions alter the strategic challenge and human acceptance of competitive AI.
Summary updated