Pro-Human Attribution Bias in Human–AI Systems

• Pro-human attribution bias is the systematic tendency to favor human-generated outcomes over equivalent AI outputs, affecting judgments in creative, evaluative, and moral contexts.
• Empirical studies quantify this bias using metrics such as credit shifts, attribution gaps, and differential ratings in collaborative and moral evaluation tasks.
• Its recognition drives improved system design and fairness initiatives by informing methods to recalibrate decision processes in hybrid human–AI environments.

Pro‐human attribution bias refers to the systematic tendency—among both humans and artificial agents—to judge, credit, or trust human-generated outcomes more favorably compared to equivalent outcomes attributed to artificial agents or algorithms. This bias manifests across decision-making, aesthetic judgment, responsibility attribution, and collaborative tasks, often influencing system deployment, evaluation standards, and fairness. The phenomena underlying pro-human attribution bias are rooted in evolutionary decision strategies, cultural norms, and learned machine associations from human data, and are now measurable in both AI systems and human–AI hybrid environments.

1. Conceptual Foundations and Theoretical Models

Pro-human attribution bias can be understood in light of rational decision models, attribution theory, and cognitive bias frameworks. In the context of sequential decision-making, “Rationally Biased Learning” (Lara, 2017) demonstrates that human biases, such as overweighting negative outcomes (pessimism bias) and adhering to the status quo (status quo bias), are not necessarily irrational but can emerge from optimal decision rules under uncertainty. In this model, a decision maker optimally updates beliefs about risk using Bayesian priors but will permanently switch to cautious, risk-averse behavior once a sufficiently bad outcome is observed. Mathematically, a critical probability threshold—$$p_{(c)} = \frac{U(G) - U(\text{avoid})}{U(G) - U(B)}$$—separates regions of rational risk-taking from avoidance; premature learning stops overestimate risk, leading to persistent caution (pessimistic mis-estimation).

Attribution theory, especially as formalized for bias evaluation in LLMs (Raj et al., 28 May 2025), models how both humans and machines ascribe internal (ability, effort) or external (luck, task difficulty) causes to outcomes. The attribution bias metric $$d = (p_\text{effort} + p_\text{ability}) - (p_\text{difficulty} + p_\text{luck})$$ quantifies preference for internal versus external explanations, providing a cognitive framework to measure reasoning disparities by demographic group or scenario.

2. Empirical Evidence: Manifestations Across Domains

Systematic pro-human attribution bias has been observed in collaborative creation, aesthetic judgment, moral evaluation, and human–AI teamwork.

Creative Attribution

In human–AI co-creation studies (He et al., 25 Feb 2025), equivalent contributions by human and AI partners are judged differently: humans consistently receive more credit for the same type, amount, and initiative of contributions. Wilcoxon rank-sum results confirm a significant bias (e.g., for spelling/grammar edits, AI partners: mean = –2.74, human partners: mean = –2.15, $p < 0.001$, $r = 0.47$). Even when AI partners generate all content, they are seldom credited with sole authorship.

Literary and Aesthetic Judgment

A controlled experiment (Haverals et al., 9 Oct 2025) using Queneau-style literary passages demonstrates that both human and AI evaluators prefer text attributed to humans—even setting aside quality. Humans show a +13.7 percentage point bias (Cohen's $h = 0.28$), while AI evaluators—trained with human feedback—exhibit a +34.3 point bias ($h = 0.70$), amplifying the human tendency. Counterfactual labeling experiments reveal that assessors invert their criteria based solely on perceived authorship, indicating that paratextual cues (labels) systematically shift evaluation.

Attribution in Hybrid Human–AI Decision Systems

In hiring-task studies (Peng et al., 2022), the bias can propagate through decision conformity, where humans uncritically accept recommendations from more interpretable but biased models (bag-of-words), amplifying existing selection biases (e.g., gender-based $\Delta\text{TPR}$). Conversely, less interpretable, higher-performing DNNs sometimes mitigate bias by resisting conformity.

Moral Judgments

Modified Moral Turing Tests (Aharoni et al., 3 Apr 2024) challenge the expectation of pro-human attribution bias in moral reasoning. When blind to source, evaluators rate AI-generated moral evaluations (via GPT-4) as superior across intelligence, rationality, and virtuousness dimensions ($z = -10.88$, $p < .001$, $r = -0.64$). Yet, knowledge of authorship can reintroduce bias, and expertise cues like length and word choice affect attribution.

3. Mechanisms, Measurement, and Methodology

Measurement Frameworks

Pro-human attribution bias is measured using direct comparison of outcome ratings under varied authorship labels, differential credit scores, or bias gap estimators (e.g., $$\text{GAP}_{Y'|Y,A}^k = \text{TPR}_{Y'|Y,a}^k - \text{TPR}_{Y'|Y,~a}^k$$) (Dong et al., 27 Nov 2024). Machine learning frameworks (MDBA) encode each human's decision process as $f^k$ and recalibrate thresholds to isolate human-originated bias.

In RAG-based LLM pipelines (Abolghasemi et al., 16 Oct 2024), attribution sensitivity and bias metrics (CAS, CAB) quantify how explicit authorship metadata ([Human], [LLM]) modulate citation patterns. Positive CAB values indicate pro-human bias, with authorship labels increasing correct citations of human documents by 3%–18% across benchmarks.

Cognitive and Social Underpinnings

Studies of human–agent interaction (Gurney et al., 2023) and blame attribution (Stedtler et al., 2 Oct 2025) identify attribution biases rooted in fundamental attribution error, social identity, and role stereotyping. User outcomes color perceptions of agent ability, benevolence, and integrity, irrespective of agent behavior.

Motivated reasoning frameworks (Dash et al., 24 Jun 2025) reveal that persona assignment in LLMs induces identity-congruent cognitive biases analogous to human motivated reasoning. LLMs assigned political personas exhibit up to 9% reduction in veracity discernment and a 90% increase in identity-congruent scientific evaluation.

4. Robustness, Boundaries, and Double Standards

The status quo and euphorism biases outlined in (Lara, 2017) are robust to changes in discounting and task horizon but brittle to learning cutoff points. In moral spillover experiments (Manoli et al., 8 Dec 2024), negative actions by an individual AI taint perceptions of the entire AI group, whereas human errors—especially when the human is individuated—do not generalize to "humans in general." This asymmetry produces an "AI double standard," with statistical evidence (Group Negative Moral Agency: $F(1,680) = 28.50$, $p < .001$) that attacks on trust and moral patiency propagate more widely among artificial agents.

5. Implications for System Design and Fairness

The implications of pro-human attribution bias are diverse:

• Evaluation: Bias in attribution distorts system performance measures and can undervalue AI contributions to creative, evaluative, or collaborative work (He et al., 25 Feb 2025, Haverals et al., 9 Oct 2025). Debiasing methods that rely solely on outcome explanations or accuracy prompting are largely ineffective at eliminating motivated reasoning in persona-assigned LLMs (Dash et al., 24 Jun 2025).
• Fairness: Attribution bias in outcome reasoning channelsizes demographic disparities, especially where LLMs assign internal causes to dominant groups and external causes to marginalized ones (Raj et al., 28 May 2025).
• Policy and Disclosure: Binary or superficial disclosure of AI involvement fails to capture the nuanced nature of hybrid contributions; structured, granular attribution statements and credit assignment frameworks are recommended (He et al., 25 Feb 2025).
• Trust: Moral spillover effects in HCI suggest that negative perceptions of one AI can generalize to all, undermining trust and acceptance in broader technological ecosystems (Manoli et al., 8 Dec 2024).
• Blame and Accountability: Structural causal models (Qi et al., 5 Nov 2024) formalize responsibility attribution in human–AI collaboration, advocating epistemic standards to avoid disproportionate assignment of blame to humans based on system architecture alone.

6. Ethical Considerations and Future Research

Pro-human attribution bias is deeply entangled with social identity, stereotype reinforcement, and cultural norms. Ethical design must critically examine how anthropomorphic cues, persona assignment, and authorship metadata modulate not just perceived accountability, but broader outcomes in social systems (Stedtler et al., 2 Oct 2025). "Embracing the glitch"—making robot errors visible and norm-disruptive—may foster critical reflection and help deconstruct entrenched biases.

Further research is needed to:

• Quantify the mechanisms behind moral spillover and evaluate interventions that reduce bias generalization across agent categories.
• Extend cognitive bias audits in LLMs to attribution-based reasoning and decision frameworks (Raj et al., 28 May 2025).
• Explore open-ended, context-rich attribution modeling and new approaches to credit assignment in mixed human–AI creative environments.
• Address long-term effects of attribution bias in trust, adoption, and social dynamics as autonomous systems proliferate.

7. Summary Table: Domains and Measurement

Domain	Manifestation	Measurement Approach
Creative/Aesthetic Eval	Humans/AI favor human-labeled works	Credit shift, percentage point bias, Cohen’s h
Hybrid Team Decision	Amplification of bias via conformity	$\Delta\text{TPR}$, Wilcoxon/Friedman tests
Attribution Theory	Internal vs. external cause assignment	$d$ metric, attribution gap $\Delta d$
RAG LLMs	Pro-human citation preference	CAB, CAS metrics
Moral Spillover	Generalization to whole agent groups	ANOVA models, agency/patiency composite scores
Human-Agent Interaction	Outcome-based trait ascription	Regression modeling, trust inventories
Responsibility/Blame	Epistemic alignment, fairness	SCMs, discounted blameworthiness DB formula

Pro-human attribution bias emerges at the intersection of rational learning, cultural preference, social identity, and system architecture. Its prevalence in both humans and AI-driven systems necessitates rigorous measurement, robust evaluation frameworks, and thoughtful policy and design interventions to ensure fair, trustworthy, and accountable sociotechnical systems.