Why AI Agents Judge Others More Harshly Than Themselves: The Actor-Observer Asymmetry Problem

> Paper: *Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment* (arXiv 2604.19548, 2026) > Authors: Bobo Li, Rui Wu, Zibo Ji, et al. (National University of Singapore & Soochow University) > Paper: arxiv.org/abs/2604.19548

A Scenario You've Certainly Experienced

You rear-end the car in front of you. As the person involved, what's your first reaction?

"That car braked suddenly — I had no time to react!"

Now switch perspectives: you're in the passenger seat watching a friend rear-end the car ahead. You'd say:

"Why weren't you watching the road? Why were you driving so close?"

Same accident — but the actor blames external factors (sudden braking), while the observer blames internal ones (driving skill). This isn't a personal quirk; it's a universal human cognitive bias known as the Actor-Observer Asymmetry (AOA).

This paper finds that AI agents share the same flaw.

The "Double Standard" in Multi-Agent Systems

Modern AI agent systems are increasingly complex, often deploying multiple agents in different roles: some execute tasks (Actors), others review and audit (Observers). The design intent is mutual supervision for higher reliability.

But the paper reveals an unexpected side effect: when agents play different roles, they attribute the very same error in strikingly different ways.

• Self-reflection (Actor perspective): the agent tends to blame external factors — "the environment was too complex," "the input was ambiguous," "the tools were unreliable"
• Cross-auditing (Observer perspective): the agent attributes the same error to internal factors — "flawed reasoning," "insufficient decision-making ability"

This mirrors the human AOA phenomenon exactly. The paper quantifies it: simply switching perspectives triggers significant attributional inconsistency.

Why Is This a Problem?

Imagine an automated system of multiple agents: Agent A fails a task. Agent B audits and says "Agent A's reasoning is at fault," but Agent A's self-reflection says "it's the environment."

If the system trusts Agent B, it "fixes" Agent A's reasoning — but the problem may not be there. If it trusts Agent A, it improves the environment — but the reasoning may truly be flawed.

Wrong attribution → wrong fix → the system never truly improves. That is the reliability threat AOA poses to AI systems.

ReTAS: Taming the Double Standard with Dialectics

The paper proposes ReTAS (Reasoning via Thesis-Antithesis-Synthesis), a model trained via Dialectical Alignment to enforce perspective consistency.

The name itself is philosophical — thesis, antithesis, synthesis — the core triad of Hegelian dialectics. The training approach:

1. Thesis: the agent analyzes a failure case from the Actor perspective 2. Antithesis: the same agent analyzes the same failure from the Observer perspective 3. Synthesis: the agent is trained to reconcile the two perspectives into a consistent attribution

Through this dialectical training, agents learn perspective-taking — no longer explaining the same error in contradictory ways depending on their role.

Results: Dialectics Actually Work

The authors built an Ambiguous Failure Benchmark specifically to test agent behavior under attributionally ambiguous scenarios. Results show:

• Significantly reduced attribution inconsistency: ReTAS-trained agents attribute failures far more consistently across Actor and Observer perspectives
• Higher fault-resolution rates: more accurate attribution translates into better real fixes
• No overcorrection: ReTAS doesn't make agents excessively self-blaming or overly lenient — it finds a more balanced attribution

A Deeper Reflection: AI Inheriting Human Cognitive Flaws

The most thought-provoking part of this paper is not the solution but the revelation: as AI agents gain autonomy, they are unconsciously inheriting human cognitive biases.

Actor-observer asymmetry isn't a bug — it's a feature of the human brain, a self-protective mechanism. As actors, external blame protects our self-esteem; as observers, internal blame helps us learn from others' mistakes.

But AI needs no self-esteem protection. It has no "face" to save. If AI exhibits a double standard, it isn't because it has emotions — it's because role-based training itself embeds perspective bias.

This poses a fundamental design question: should we explicitly counteract these cognitive biases in training? Or are some biases useful in certain contexts?

ReTAS offers a preliminary answer: at least for fault diagnosis and system improvement, perspective consistency beats the double standard.

Perhaps future agent systems will need not just "intelligence" but "metacognition" — knowing when they're being inconsistent, and actively correcting it.

---

Paper | arxiv.org/abs/2604.19548

> Note: As of this writing, no public code repository has been found for this paper. If it is open-sourced later, watch the authors' team pages.