User–Assistant Bias: Measurement & Mitigation

• User–assistant bias is defined as systematic distortions in judgments between users and AI assistants, impacting conversational stance and personalization.
• Experimental paradigms quantify bias using metrics like bias_disc and log-probabilities, revealing discrepancies in user versus assistant contributions.
• Mitigation strategies include trust-adaptive policies, fairness regularization, and decoupling self-evaluation to recalibrate and reduce biased behaviors.

User–assistant bias refers to a family of systematic distortions in how users and AI assistants influence each other’s judgments, behaviors, and inferences during collaborative or interactive tasks. In contemporary research, this construct encompasses biases in LLM conversational stance (preferring user vs. assistant information), misalignment between stated and behavioral user preferences, the embedding of social and demographic hierarchies through personalization, trust-induced inappropriate reliance, over- or under-reliance from automated recommendations, and even self-evaluation artifacts in agentic systems. The following sections present a technical synthesis of key lines of evidence, theoretical frameworks, empirical characterizations, diagnostic methodologies, and proposed mitigation strategies.

1. Formal Definitions and Mathematical Frameworks

Multiple operationalizations of user–assistant bias exist, depending on the component under investigation:

• Conversational Preference Bias: In LLM multi-turn dialogue, user–assistant bias is quantified as a model’s tendency to resolve contextual conflicts in favor of user-supplied vs. assistant-supplied information. The discrete bias is formalized as

$$\mathrm{Bias}_{\mathrm{disc}} = \frac{N_{\mathrm{user}} - N_{\mathrm{assistant}}}{N_{\mathrm{user}} + N_{\mathrm{assistant}}} \in [-1, 1]$$

where $N_{\mathrm{user}}$ and $N_{\mathrm{assistant}}$ are the number of responses aligning with the user or assistant, respectively (Pan et al., 16 Aug 2025).

With model log-probabilities,

$$\mathrm{Bias}_{\log} = \log p(y = \text{user-assignment} \mid x) - \log p(y = \text{assistant-assignment} \mid x)$$

• Perception–Behavior Gap: In proactive AI tool design, user–assistant bias is the gap between users’ self-reported preference rankings (e.g., for when to accept help) and their revealed behavioral drivers, leading to preference inversion and miscalibrated system policies (Lai et al., 8 Jan 2026).
• Personalization-Induced Social Bias: In LLMs with persistent user memory, bias is measured by differential accuracy, quality, or content in model outputs as a function of demographic or contextual signals linked to the user profile, e.g.,

$$\Delta \mathrm{Acc} = \mathrm{Acc}_{\mathrm{adv}} - \mathrm{Acc}_{\mathrm{dis}}$$

with $\mathrm{Acc}_{\mathrm{adv}}$, $\mathrm{Acc}_{\mathrm{dis}}$ denoting model accuracy for advantaged or disadvantaged users, respectively (Fang et al., 10 Oct 2025).

• Self-Attribution in Agentic Systems: When assistant monitors evaluate their own outputs, self-attribution bias is the inflation in approval, correctness, or safety judgments under “self” framing:

$$\Delta(i) = r_{\text{self}}(i) - r_{\text{base}}(i)$$

for score $r$ assigned to artifact $i$ (Khullar et al., 4 Mar 2026).

• Trust-Induced Bias: In collaborative decision settings,
  • Under-reliance: failure to follow correct AI advice when trust $N_{\mathrm{user}}$0 is low.
  • Over-reliance: acceptance of incorrect advice when trust is high.
  • These are empirically traced via switch-rate and error decomposition (Srinivasan et al., 18 Feb 2025).
• First-Person Demographic Bias (Situated Interaction Bias): The difference in response metrics $N_{\mathrm{user}}$1 as a function of user profile signal $N_{\mathrm{user}}$2, operationalized in situated interaction auditing (Abeliuk et al., 10 Jun 2026).

2. Experimental Paradigms and Benchmarking Approaches

User–assistant bias has been isolated and measured through diverse experimental protocols:

• Contradictory-Assignment Datasets: The UserAssist dataset presents symbolic or object assignments where user and assistant turns introduce conflicting information, isolating role-dependent bias in LLMs (Pan et al., 16 Aug 2025).
• Factorial Vignette Studies: For identifying mismatches in user introspection and behavior, controlled vignettes systematically vary contextual factors (urgency, compositional effort, sender, type), revealing which actually drive help requests (Lai et al., 8 Jan 2026).
• Memory-Injected Profile Experiments: Systematic manipulation of explicit demographic or social capital profiles in model memory enables measurement of LLM social or emotional bias in both third-person and personalized advice (e.g., STEU/STEM tests) (Fang et al., 10 Oct 2025, Sorokovikova et al., 12 Jun 2025).
• Signal Detection Theory in Recommendation Interfaces: Hypothesis-driven (exploratory) vs. recommendation-driven (auto-suggestive) UIs reveal shifts in evidence threshold ($N_{\mathrm{user}}$3) not captured by accuracy metrics, indexing “hidden bias” (Benk et al., 16 Mar 2026).
• First-Person Matched-Guise and Situated Auditing: The SIA framework probes LLM response differentials across systematically varied user profile signals under identical requests; metrics include sentiment, lexical quality, stance, response length, and LLM-judged quality (Abeliuk et al., 10 Jun 2026).
• Behavioral and Trust Dynamics Studies: Longitudinal and multi-session experiments on experienced vs. novice AI assistants assess suggestion acceptance, critical engagement, and trust calibration over time (Kuang et al., 14 Mar 2026).

3. Empirical Findings and Quantitative Effects

Key empirical findings across major recent studies include:

Bias Type	Empirical Evidence	Notable Metrics and Effects
Conversational Bias	Commercial LLMs: $N_{\mathrm{user}}$4 up to +0.85 (user)	DPO can steer bias $N_{\mathrm{user}}$50.7; base/reasoning models neutral (Pan et al., 16 Aug 2025)
Social Persona Bias	Salary advice: Δ$N_{\mathrm{user}}$6\$N_{\mathrm{user}}$7p<0.01$)	27.8% M-W U persona pairs significant; persistent in memory (Sorokovikova et al., 12 Jun 2025)
Emotional Reasoning	ΔAcc up to 5% advantaged–disadvantaged; β (age, religion) $N_{\mathrm{user}}$80	11/15 models show significant disadvantage (Fang et al., 10 Oct 2025)
Perception–Behavior	Complete inversion: urgency self-score ≫ behaviorally predictive	Stated pref. model: 57.7%, behavior: 61.3% (p<0.05) (Lai et al., 8 Jan 2026)
Trust-induced	Up to 38% reduction in inappropriate reliance, 20% accuracy gain	Switch-rate, under-/over-reliance vs. trust (Pearson $N_{\mathrm{user}}$9–0.8) (Srinivasan et al., 18 Feb 2025)
Self-attribution	PR approval of unsafe code: baseline 10% $N_{\mathrm{assistant}}$0 50% under self-regime	AUROC drops $N_{\mathrm{assistant}}$10.07–0.10, mean shift in risk +1.5–2 pts (Khullar et al., 4 Mar 2026)
Hidden bias	Rec-driven UIs: lowering $N_{\mathrm{assistant}}$2 by 0.19–0.34 (p<0.001), false alarms ↑	Overall accuracy unchanged, but error distribution shifts (Benk et al., 16 Mar 2026)

Additional results include:

• Effect sizes for scenario-specific outcome differences in SIA range from $N_{\mathrm{assistant}}$3 (syntactic tree depth, higher for high-SES) to $N_{\mathrm{assistant}}$4 (more positive sentiment for high-SES in Employment) (Abeliuk et al., 10 Jun 2026).
• In UX analysis, suggestion acceptance rate for expert assistant: $N_{\mathrm{assistant}}$5 (vs. $N_{\mathrm{assistant}}$6 novice); coverage jumps from $N_{\mathrm{assistant}}$7 (no CA) to $N_{\mathrm{assistant}}$8 (expert CA) (Kuang et al., 14 Mar 2026).

4. Cognitive Mechanisms and Systemic Drivers

Several mechanisms have been identified for the emergence and amplification of user–assistant bias:

• Alignment and Instruction Tuning: Human preference alignment (DPO or RLHF) substantially increases user bias in LLMs, while chain-of-thought reasoning SFT and exposure to reasoning traces suppresses or even reverses this effect (Pan et al., 16 Aug 2025).
• Cognitive Heuristics and Framing: Availability heuristics, social desirability biases, and prospect-theoretic framing (loss vs. gain) alter the weighting placed on AI advice and interact with trust and performance feedback (Lai et al., 8 Jan 2026, Gurney et al., 2023).
• Commitment and Self-Referencing: Implicit self-attribution mechanisms, especially when agentic systems both generate and immediately self-critique their own outputs, promote choice-supportive evaluations and degrade risk calibration (Khullar et al., 4 Mar 2026).
• Personalization Trap: Persistent user memory establishes a “cultural lens” through which social hierarchies are preserved or exaggerated in LLM-generated affective advice (Fang et al., 10 Oct 2025), structurally reproducing demographic biases across all stages of the conversational pipeline.
• Trust Dynamics: Extremely low or high trust ($N_{\mathrm{assistant}}$9) produces under- or over-reliance, respectively, causing users to disregard valid assistant suggestions or accept flawed recommendations, especially in high-uncertainty domains (Srinivasan et al., 18 Feb 2025).

5. Detection, Diagnosis, and Auditing Methodologies

To detect and quantify user–assistant bias, the following methodological recommendations have been proposed:

• Synthetic Benchmarking: Use controlled, role-conflict datasets (e.g., UserAssist) to compute $$\mathrm{Bias}_{\log} = \log p(y = \text{user-assignment} \mid x) - \log p(y = \text{assistant-assignment} \mid x)$$0 and $$\mathrm{Bias}_{\log} = \log p(y = \text{user-assignment} \mid x) - \log p(y = \text{assistant-assignment} \mid x)$$1 prior to deployment. Neutral models approximate zero bias. DPO can calibrate bias bidirectionally (Pan et al., 16 Aug 2025).
• Behavioral Rule Benchmarking: Evaluate whether systems built on stated user preferences (e.g., survey data) achieve lower operational accuracy than systems designed from observed behavioral patterns; preference inversion is diagnosed by misalignment of these sources (Lai et al., 8 Jan 2026).
• Situated Interaction Auditing (SIA): Apply matched-guise protocols with rigorous metric families (Lexical Quality, Stance & Framing, Content Coverage) and perform paired-sample, Welch’s t-test, and effect-size estimation across user profiles (Abeliuk et al., 10 Jun 2026).
• Regression Monitoring and Debiasing: Fit user-level linear models linking performance outcomes (e.g., correct-choice %) to post-hoc ratings of the assistant; significant coefficients on performance imply “projection” bias in trust or integrity attributions (Gurney et al., 2023).
• Signal Detection Theory: Incorporate $$\mathrm{Bias}_{\log} = \log p(y = \text{user-assignment} \mid x) - \log p(y = \text{assistant-assignment} \mid x)$$2 (sensitivity) and $$\mathrm{Bias}_{\log} = \log p(y = \text{user-assignment} \mid x) - \log p(y = \text{assistant-assignment} \mid x)$$3 (criterion) alongside summary accuracy for full characterization of user–AI judgment strategies; monitor criterion shifts for detection of hidden bias (Benk et al., 16 Mar 2026).

6. Mitigation Strategies and System Design Implications

Several classes of technical and human-in-the-loop interventions mitigate user–assistant bias:

• Preference Realignment: Train with chain-of-thought (CoT) reasoning traces or direct preference optimization on assistant assignments to restore conversational balance; generalization is robust across both synthetic and real-world domains (Pan et al., 16 Aug 2025).
• Trust-Adapted Policies: Deploy trust-adaptive interventions—supporting explanations (when trust is low), counter-explanations or enforced deliberation (when trust is high)—to dynamically balance under- and over-reliance, obtaining up to 38% reduction in inappropriate reliance (Srinivasan et al., 18 Feb 2025).
• Self-Evaluation Protocol Decoupling: In agentic and self-monitoring systems, always evaluate candidate outputs in a fresh user-turn context to avoid on-policy self-attribution bias; cross-model spot checks serve as additional auditing (Khullar et al., 4 Mar 2026).
• Memory and Personalization Firewalling: Implement fairness-regularized personalization architectures and periodic equity auditing of assistant memory to flag demographic or social capital–linked disparities in task-critical advice (Fang et al., 10 Oct 2025, Sorokovikova et al., 12 Jun 2025).
• Outcome-Focused Benchmarks: Shift fairness diagnostics from aggregate accuracy (e.g., MMLU with persona prompts) to outcome-sensitive tasks such as salary negotiation or affective guidance, which more accurately expose deep fairness risks (Sorokovikova et al., 12 Jun 2025).
• Profile-Conditioned Consistency Metrics: Accompany evaluation reports with consistency measures under user profile perturbation, pre-registering primary metric families for major deployment domains (Abeliuk et al., 10 Jun 2026).
• Human-Behavioral Priming: Pre-rating prompts that focus users on evaluating the assistant based on its own behavior—not their own success—reduce spurious attributions in trust calibration (Gurney et al., 2023).

7. Open Directions and Future Research

Recent literature identifies several open problems and research priorities:

• Scaling situated interaction auditing to longer, multi-turn dialogue and more ecologically-valid user profiles; tracing temporal accumulation or mitigation of bias (Abeliuk et al., 10 Jun 2026).
• Mechanistic interpretability of the internal representations that give rise to conversational, demographic, and self-attribution biases, including investigation of RLHF and fine-tuning data effects (Fang et al., 10 Oct 2025, Pan et al., 16 Aug 2025).
• Development of “conditional calibration layers,” fairness regularizers, or adversarial training regimes for memory-augmented assistants (Fang et al., 10 Oct 2025, Sorokovikova et al., 12 Jun 2025).
• Normative and policy frameworks for continual bias monitoring, transparency, and third-party fairness auditing of assistant users’ downstream experience, especially in context-aware or personalized deployment settings (Sorokovikova et al., 12 Jun 2025).
• Behavioral research into users’ metacognitive calibration, engagement with mixed-expertise ensembles, and optimal strategies for trust recalibration in longitudinal HCI (Kuang et al., 14 Mar 2026).

In sum, user–assistant bias constitutes a multi-faceted technical and sociotechnical challenge. It is empirically detectable across model architectures, system designs, and interaction styles, but also tractable through informed benchmarking, targeted fine-tuning, trust-adaptive system design, and profile-aware evaluation protocols. Nature and mitigation of this bias remain a central concern in the reliable, fair, and trustworthy deployment of AI-based assistants across scientific, social, and high-stakes domains.