Compound Human-AI Bias: Interaction & Amplification

• Compound human-AI bias is the interactive amplification of human cognitive and algorithmic biases that degrades decision quality across domains.
• Key factors such as automation and confirmation bias, alongside algorithmic misestimations, create self-reinforcing error cascades in systems like hiring and healthcare.
• Effective mitigation requires both system-level debiasing strategies and user-focused interventions to disrupt feedback loops and enhance decision accuracy.

Compound human-AI bias is the phenomenon wherein distinct human cognitive biases and algorithmic biases interact or reinforce each other during human–AI collaboration, yielding systematic errors or distortions that would not arise from either source alone. This compounding may degrade decision quality, amplify inequity, or entrench error cascades across diverse domains such as annotation, hiring, creative evaluation, healthcare, and algorithmic governance (Beck et al., 10 Sep 2025, Lloyd, 2018, Peng et al., 2022, Felten et al., 18 Nov 2025).

1. Conceptual Foundations and Formal Definitions

The notion of compound human-AI bias centers on the recursive interaction between human decision-makers—who import a repertoire of cognitive tendencies (e.g., automation bias, confirmation bias, base-rate neglect, metacognitive disengagement)—and AI systems, which are subject to algorithmic biases such as class imbalance, data-driven misestimation, or unfair inductive generalizations (Gulati et al., 2022, Vakali et al., 30 Jul 2024, Felten et al., 18 Nov 2025). Formally, if $\mathrm{HCB} = \{h_1, \dots, h_n\}$ and $\mathrm{AIB} = \{a_1, \dots, a_m\}$ are the spaces of human and AI biases, the interaction can be mapped as

$$\mathrm{CHAB} : \mathrm{HCB} \times \mathrm{AIB} \to \mathcal{C},$$

where $\mathcal{C}$ is the set of compound bias phenomena manifested in collaborative performance (Felten, 26 Apr 2025).

A canonical model relates the combined effect to the sum or amplification of standalone biases:

$$B_{\mathrm{CH,AI}} = \mathrm{Outcome}(H+\mathrm{AI}) - \mathrm{Outcome}(\mathrm{optimal}),$$

which is typically greater than $B_H + B_{\mathrm{AI}}$ in the presence of positive feedback or over-reliance (Gulati et al., 2022). In structured interactions, compound bias can be quantified as the “alignment” between biases (e.g., gender bias alignment indices) or the degree of mutually reinforcing attitudinal and behavioral distortions (Zipperling et al., 15 May 2025).

2. Mechanisms of Compounding: Feedback Loops and Behavioral Pathways

Compound bias arises through several interdependent behavioral and algorithmic mechanisms:

• Automation bias and cognitive shortcuts: Overreliance on AI outputs (undercorrection), driven by favorable attitudes toward automation, leads to increased error acceptance, particularly under increased correction effort requirements; conversely, skepticism may yield overcorrection of correct AI suggestions (Beck et al., 10 Sep 2025).
• Confirmation bias and bias alignment: Humans tend to accept AI recommendations mirroring their own preexisting biases, reinforcing confirmation effects even when those biases are erroneous, especially if AI recommendations are algorithmically miscalibrated (Rosbach et al., 1 Nov 2024, Zipperling et al., 15 May 2025).
• Compounding via conformity: When both AI model and human share the same directional bias (e.g., favor certain demographic groups), hybrid performance displays amplified disparity (measured via group TPR difference, demographic parity, etc.), often exceeding the bias magnitude of either agent alone (Peng et al., 2022).
• Attitudinal mediation: Compound bias is strongly modulated by attitudinal variables; high “AI-liking” participants produce a $+13$ percentage-point increase in undercorrection rates and $-0.22$ log-odds decrease in overcorrection per attitude point (Beck et al., 10 Sep 2025).
• Self-reinforcing feedback loops: AI systems trained on outputs or decisions of biased human reviewers further entrench bias in subsequent data cycles and model retraining, enabling a self-perpetuating cycle of discriminatory or homogenizing outputs (Lloyd, 2018, Castro et al., 2023).

3. Quantitative Frameworks and Empirical Metrics

Empirical studies deploy a range of metrics to capture and dissect compound human–AI biases:

Metric	Formula/Definition	Context
Accuracy	$\frac{C_C + I_C}{C + I}$	Annotation audits (Beck et al., 10 Sep 2025)
Overcorrection	$\frac{C_O}{C}$	Annotation audits (Beck et al., 10 Sep 2025)
Undercorrection	$\frac{I_U}{I}$	Annotation audits (Beck et al., 10 Sep 2025)
Conformity rate	$C = \frac{1}{N}\sum_i \mathbb{I}\{h_i = m_i\}$	Hybrid hiring (Peng et al., 2022)
Gender bias $\Delta$TPR	$TPR_f - TPR_m$	Group outcome bias (Peng et al., 2022)
Bias alignment	$$\text{Alignment}_{DP} = \frac{1}{2}(2 - |DP(\mathrm{AI}) - DP(H)|)$$	Human/AI parity (Zipperling et al., 15 May 2025)
RAIR	$(\sum_i CAIR_i)/(\sum_i CA_i)$	Reliance calibration (Felten et al., 18 Nov 2025)
RSR	$(\sum_i CSR_i)/(\sum_i IA_i)$	Reliance calibration (Felten et al., 18 Nov 2025)
Bias amplification	$$BA = \frac{\text{Disparity}_{\text{model}}}{\text{Disparity}_{\text{data}}}$$	Group disparity (Lloyd, 2018)

Controlled studies (e.g., 2,784 annotators (Beck et al., 10 Sep 2025); 38,400 hiring trials (Peng et al., 2022); $N=46$ for disease classification with class imbalance (Felten et al., 18 Nov 2025)) document that compounding effects can be directly measured and statistically parsed, including via regression on attitudinal predictors, alignment indices, or within-subject crossover designs.

4. Domains and Case Studies: Manifestations Across Application Areas

Compound human–AI bias is observable and robust across a spectrum of practical domains:

• Annotation and data auditing: Crowdsourced annotation tasks reveal that correction requirements induce cognitive shortcutting, amplifying AI omission errors among automation-biased individuals (Beck et al., 10 Sep 2025).
• Hiring and high-stakes selection: Human-AI collaboration in candidate shortlisting demonstrates that interpretable but biased models (e.g., bag-of-words) can amplify group disparities through conformity, especially where human and model biases align (Peng et al., 2022).
• Healthcare and diagnosis: In time-constrained computational pathology, confirmation bias is strongly triggered when AI advice coincides with human error; under time stress, this shifts to automation bias, with indiscriminate trust compounding final error (Rosbach et al., 1 Nov 2024).
• Base-rate neglect and class imbalance: Users' base-rate neglect interacts with AI class imbalance, yielding a mutually reinforcing cycle—in unbalanced settings, users increasingly trust AI's rare-class predictions, distorting disease prevalence estimates (Felten et al., 18 Nov 2025).
• Creative evaluation and attribution: In literary evaluation, both humans and LLMs manifest pro-human attribution bias, but LLMs amplify this bias by $2.5\times$, systematically devaluing AI-generated content when labeled as such (Haverals et al., 9 Oct 2025).

5. Theoretical Developments and Interactionist Frameworks

A rigorous interactionist paradigm models compound human–AI bias as the emergent output of underlying bias pairs $(h_i, a_j)$ with coefficients of amplification or mitigation (Felten, 26 Apr 2025):

$c_{ij} = \alpha_{ij} h_i + \beta_{ij} a_j$

where $\alpha_{ij}$ and $\beta_{ij}$ encode sensitivity of the human–AI team to the respective bias. Mitigation strategies must thus consider both vectors: system-side (algorithmic debiasing, post hoc calibration) and user-side (counter-bias prompts, icon arrays, cognitive-forcing interface elements). Procedural frameworks recommend mapping the full bias cross-product $\mathrm{HCB}\times\mathrm{AIB}$ to an intervention matrix and empirically iterating to reduce compounding (Felten, 26 Apr 2025, Vakali et al., 30 Jul 2024).

A socio-technical mapping links core heuristics—representativeness, availability, anchoring, and affect—to canonical AI bias manifestations at each pipeline phase (pre-, in-, post-processing), making explicit how human biases are reflected and then amplified in data, model design, and deployment decisions (Vakali et al., 30 Jul 2024).

6. Mitigation and Design: Strategies for Breaking Compounding Cycles

Effective mitigation demands interventions tailored to the structure of compounding:

• Workflow design: Decoupling correction from verification, equalizing cognitive effort across instance types, and stratified sampling on psychological attitudes can minimize error propagation and under/over-correction (Beck et al., 10 Sep 2025).
• Active debiasing: Confidence-adaptive time allocation, explicit explanation prompts, and cognitive-forcing functions directly target anchoring and automation, as shown to improve correction accuracy (Rastogi et al., 2020).
• Attitude measurement and interface personalization: Measuring and stratifying samples on AI trust, domain familiarity, and bias alignment ensures that neither overreliance nor blanket skepticism dominates cumulative outcomes (Beck et al., 10 Sep 2025, Horowitz et al., 2023).
• Interactionist evaluation: Auditing for bias amplification via conformity and alignment indices is critical in hybrid deployments, with corrective interventions triggered by high amplification metrics (Peng et al., 2022, Zipperling et al., 15 May 2025).
• Societal-level interventions: Reducing user–AI communication friction, balancing training data, and monitoring population-wide variance in outputs can inhibit homogenization spirals and feedback-driven amplification (Castro et al., 2023).
• Ensemble and transparency mechanisms: Multiplicity in evaluators (human–AI hybrid panels) and prompt-based metadata obfuscation can counteract attribution feedback loops, especially in creative domains (Haverals et al., 9 Oct 2025).

7. Limitations, Open Challenges, and Research Directions

Despite advances, key challenges remain:

• Generalization of bias alignment metrics beyond binary or one-dimensional biases, and extension to intersectional or contextual compound effects (Zipperling et al., 15 May 2025).
• Capturing higher-order or non-linear feedback in longitudinal and multi-agent human–AI ecosystems (Felten, 26 Apr 2025, Castro et al., 2023).
• Robust measurement and isolation of compound effects versus additive or independent biases, especially under varying task regimes and user populations (Gulati et al., 2022).
• Ethical and governance considerations, requiring continuous audit and adaptation as biases evolve through deployment (Lloyd, 2018).
• Integration with XAI and transparency pipelines to surface and resolve latent compound effects invisible to static model validation (Gulati et al., 2022, Vakali et al., 30 Jul 2024).

Future work is expected to systematically catalog bias cross-effects, develop standardized compound bias metrics, formalize synergistic and antagonistic interactions, and operationalize adaptive debiasing pipelines throughout the AI lifecycle. Interactionist frameworks and socio-technical mappings are increasingly positioned as foundational to rigorous evaluation and responsible deployment of human-in-the-loop AI systems.