When We Trust Robots More Than People

Author: Denis Avetisyan

A new study suggests that declining faith in human judgment may be driving increased reliance on artificial intelligence for decision-making guidance.

Original Caption: “(a) Results on the MuJoCo benchmark suite. The plot shows the average return over 50 evaluation episodes, with error bands representing one standard deviation. The solid and dashed lines represent the performance of the proposed method and the baseline, respectively.” The proposed method consistently outperforms the baseline across the MuJoCo benchmark suite, demonstrating a statistically significant improvement in average return-as indicated by the consistently higher solid line and accompanying standard deviation bands-and establishing its robustness in complex robotic control tasks.

Research identifies ‘deferred trust’ as a key factor in human-AI interaction, where trust shifts from human agents to algorithms due to diminished confidence in human expertise.

Despite growing reliance on artificial intelligence, the foundations of AI trust remain surprisingly linked to pre-existing interpersonal dynamics. This is the central question addressed in ‘Trust in AI emerges from distrust in humans: A machine learning study on decision-making guidance’, which introduces the concept of ‘deferred trust’-the tendency to favor AI guidance when faith in human agents erodes. Through analysis of decision-making scenarios, the study demonstrates that diminished trust in individuals consistently predicts increased reliance on AI, particularly among those with higher socioeconomic status and lower technology engagement. As AI increasingly mediates crucial aspects of life, how can we calibrate vigilance and ensure responsible reliance on these systems without simply transferring human biases?

The Shifting Foundations of Trust in the Age of Artificial Intelligence

The foundations of trust, historically forged through reciprocal human interactions – observing competence, assessing integrity, and gauging benevolence – are undergoing a significant upheaval as individuals increasingly engage with Artificial Intelligence. These established models struggle to account for the unique dynamics inherent in interactions with AI, where cues are often synthetic and the basis for evaluation shifts from demonstrable character to algorithmic performance. Unlike assessing a human colleague, judging an AI’s trustworthiness requires evaluating its outputs, data provenance, and the opacity of its underlying processes, creating a fundamentally different cognitive and emotional landscape. This disconnect challenges long-held assumptions about how trust is established and maintained, raising crucial questions about accountability, transparency, and the potential for misplaced confidence in systems lacking genuine social understanding.

Established models of technology acceptance, such as the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT), were designed to assess user adoption of tools with defined functionalities. However, these frameworks struggle to account for the unique characteristics of Large Language Models (LLMs). Unlike traditional software, LLMs engage users in conversational interactions, blurring the lines between tool and social partner. The nuanced interplay of factors – including perceived empathy, stylistic consistency, and the model’s ability to handle ambiguity – significantly influences trust formation in a way that traditional predictors of usefulness and ease of use simply cannot capture. Consequently, current models offer an incomplete picture of how and why individuals choose to rely on, or distrust, these increasingly sophisticated AI systems, necessitating new approaches to understand this evolving dynamic.

As artificial intelligence permeates daily life – from medical diagnoses and financial advice to news consumption and personal recommendations – comprehending the foundations of trust in these systems becomes paramount. This isn’t simply a matter of user acceptance; it’s a fundamental requirement for the effective and ethical integration of AI into society. Individuals are increasingly delegating decisions and tasks to algorithms, and the willingness to do so hinges on a perceived reliability and trustworthiness. However, traditional notions of trust, built on demonstrable competence, benevolence, and integrity, are being tested by the opaque nature of many AI systems. Investigating why people choose to trust – or distrust – AI, and the factors that influence these perceptions, is therefore crucial for designing systems that are not only technically proficient but also align with human values and expectations, fostering a future where AI serves as a collaborative partner rather than a source of anxiety or manipulation.

Deconstructing Trust: The Pillars of Evaluation

Trust formation is not a singular process but rather a composite of distinct evaluative dimensions. Epistemic trust centers on objective assessments of an entity’s reliability and accuracy – whether information provided is truthful and consistently verifiable. Conversely, social trust is rooted in subjective perceptions of an entity’s competence – their ability to execute tasks effectively – and benevolence – their perceived goodwill and intention to act in one’s best interest. These two forms of trust often interact; an entity perceived as both competent and benevolent is more likely to be granted epistemic trust, while demonstrably reliable and accurate information strengthens perceptions of competence and goodwill. The weighting of these dimensions can vary based on context and individual factors, but both contribute to overall trust assessment.

Transparency and fluency significantly influence information acceptance. Increased fluency, referring to the ease with which information is processed, can reduce an individual’s level of epistemic vigilance – the tendency to question the reliability of information sources. Research indicates that easily processed information is often perceived as more truthful, even in the absence of supporting evidence. This effect lowers cognitive scrutiny and increases susceptibility to misinformation, as individuals allocate fewer cognitive resources to evaluating the veracity of fluent statements. Consequently, factors enhancing fluency, such as clear presentation or repetitive exposure, can paradoxically increase belief in false or unsubstantiated claims.

Algorithm appreciation, a tendency to view algorithms favorably, and automation bias, the propensity to favor suggestions from automated systems, contribute to inflated trust in AI outputs regardless of accuracy. Studies demonstrate that individuals often assign higher credibility to information presented by an algorithm, even when explicitly informed of potential errors or presented with contradictory evidence. This effect is exacerbated when the algorithm is perceived as complex or sophisticated, leading users to overestimate its reliability and underestimate the importance of independent verification. Consequently, flawed AI outputs can be accepted uncritically, impacting decision-making processes in areas such as healthcare, finance, and legal assessment.

Predicting Reliance: A Machine Learning Perspective

XGBoost, a gradient boosting algorithm, was selected as the primary predictive model due to its demonstrated performance in handling complex datasets and its ability to capture non-linear relationships between features. The model was trained to predict agent selection – whether a participant would choose a human or an AI agent – based on a comprehensive feature set encompassing individual characteristics such as age, gender, and cognitive traits, as well as interaction factors including task complexity, time pressure, and prior experience with AI. Feature importance was determined through analysis of the model’s learned weights, and regularization techniques were employed to prevent overfitting and enhance generalization to unseen data. The algorithm’s inherent capacity for feature selection and its robust handling of missing data contributed to its effectiveness in this prediction task.

SHAP (SHapley Additive exPlanations) analysis was applied to the XGBoost model to determine the contribution of each feature to the prediction of agent selection. This method calculates the Shapley value for each feature, representing its average marginal contribution to the prediction across all possible feature combinations. The resulting SHAP values were then used to rank features by their importance in influencing the model’s output; features with larger absolute SHAP values were considered more impactful. Analysis revealed that specific individual characteristics, such as participant age and prior experience with AI, alongside interaction factors including the complexity of the task and the perceived reliability of the agent, were the strongest predictors of whether a human or AI agent was selected. The magnitude and direction of these SHAP values provided insights into how each feature contributed to the prediction – positive values indicating a feature increased the likelihood of AI selection, and negative values suggesting the opposite.

Participant and scenario segmentation was performed using K-Means and K-Modes clustering techniques to identify groupings based on agent selection behavior. K-Modes clustering, applied to the scenarios, yielded a Davies-Bouldin Index of 0.809 and a Dunn-like Index of 1.052. These index values indicate effective cluster separation, suggesting that the identified scenario groupings exhibit distinct agent selection patterns and represent meaningful distinctions in trust propensity. The K-Means clustering of participants similarly revealed groups with differing tendencies toward selecting either human or AI agents, providing insight into individual variations in trust.

The predictive models demonstrated a high degree of accuracy in identifying the factors influencing participant selection of AI agents. Specifically, average precision values reached 0.8813 ± 0.0872 for Situation 24, 0.8638 ± 0.0995 for Situation 26, and 0.8705 ± 0.0801 for Situation 9. These precision scores, reported with their respective standard deviations, indicate a strong capability of the models to correctly identify the key predictors of AI agent selection across the evaluated scenarios. The consistent values above 0.8 suggest robust performance in differentiating factors leading to the choice of an AI versus a human agent.

The Rise of Deferred Trust: When Algorithms Become Preferred Sources

Recent research indicates a growing trend termed ‘Deferred Trust,’ wherein individuals increasingly place faith in artificial intelligence systems over traditionally trusted human sources. This phenomenon isn’t necessarily about an inherent belief in AI’s infallibility, but rather a response to perceived biases or incompetence in human agents – a shift fueled by the perception that AI offers a more neutral and objective assessment of information. The study suggests that when confidence in established authorities wanes, individuals may actively seek out AI as an alternative, viewing its data-driven outputs as less susceptible to manipulation or personal agendas. This isn’t simply a technological preference, but a social recalibration, indicating a potential reshaping of how individuals evaluate credibility and form opinions in an increasingly complex information landscape.

The erosion of faith in established institutions – encompassing media, government, and expert communities – creates a notable vacuum, increasingly filled by perceptions of artificial intelligence as a more reliable source of information. This shift isn’t simply about technological adoption; it reflects a deeper societal realignment where AI is viewed as comparatively neutral and objective, particularly by individuals disillusioned with traditional authorities. Consequently, AI doesn’t just offer answers; it provides a sense of stability and trustworthiness when conventional sources are perceived as biased or untrustworthy, effectively addressing a growing social need for dependable information and potentially reshaping how individuals form beliefs and make decisions.

The study reveals a compelling link between socioeconomic status and patterns of trust in both human and artificial intelligence. Individuals with limited access to reliable information, often correlated with lower socioeconomic standing, demonstrate a heightened susceptibility to deferring trust to AI systems. This isn’t necessarily a reflection of inherent technological optimism, but rather a consequence of diminished faith in traditional sources – institutions, experts, or even local networks – coupled with the perceived neutrality and accessibility of AI. Consequently, AI can become a disproportionately relied-upon source of information for those marginalized from conventional knowledge channels, potentially exacerbating existing societal inequalities if algorithmic biases are present. The research suggests that equitable access to verified information is crucial to prevent a scenario where trust becomes stratified along socioeconomic lines, with AI inadvertently reinforcing existing disparities.

A predictive model successfully identified individuals predisposed to prioritize AI as a trusted source with an average precision of 0.7961, accompanied by a standard deviation of 0.1127. This demonstrates a quantifiable ability to discern those more likely to exhibit ‘Deferred Trust’ – shifting reliance from human agents toward artificial intelligence. The model’s performance suggests that certain characteristics and patterns of information consumption can reliably indicate a preference for AI-driven information, opening avenues for understanding and potentially addressing the underlying factors driving this phenomenon. This predictive capability offers a valuable tool for researchers seeking to explore the complex interplay between trust, information sources, and the growing role of AI in shaping public opinion and individual decision-making.

Implications and Future Directions for Trustworthy AI

The pursuit of trustworthy artificial intelligence necessitates a fundamental shift in design principles, prioritizing transparency, explainability, and fairness as core components rather than afterthoughts. Current research highlights that simply achieving high accuracy is insufficient; users require an understanding of how an AI system arrives at a particular conclusion to build genuine confidence. This demands moving beyond “black box” models towards systems that can articulate their reasoning processes in a human-understandable manner. Furthermore, addressing inherent biases within training data is crucial to prevent the perpetuation – or even amplification – of societal inequalities. A commitment to these principles isn’t merely an ethical imperative, but also a practical one; without trust, the widespread adoption and beneficial integration of AI into critical aspects of life will remain significantly hampered, limiting its potential to address complex global challenges.

The increasing reliance on artificial intelligence necessitates a thorough investigation into the long-term societal impacts of what’s termed “Deferred Trust” – the provisional acceptance of AI recommendations without full understanding or scrutiny. Research indicates that widespread deferral could subtly erode social cohesion, as individuals increasingly outsource judgment and decision-making to algorithms, potentially diminishing critical discussion and shared understanding. Furthermore, this trend carries the risk of amplifying existing societal inequalities; biased algorithms, coupled with uncritical acceptance, could systematically disadvantage marginalized groups, solidifying and even exacerbating disparities in areas like loan applications, hiring processes, or even access to essential services. Future studies must therefore prioritize understanding how Deferred Trust interacts with pre-existing social structures and biases, and explore mitigation strategies to ensure equitable and inclusive AI adoption, preventing a future where algorithmic decisions deepen societal divides.

The pervasive integration of artificial intelligence into daily life necessitates a proactive cultivation of Epistemic Vigilance – a sustained practice of questioning information, evaluating sources, and recognizing potential biases. This isn’t merely about distrusting AI, but rather about fostering a critical mindset that extends to all information streams, demanding evidence and logical reasoning. Research indicates that individuals readily accept outputs from AI systems, often attributing an unwarranted level of authority, which can hinder independent thought and perpetuate misinformation. Therefore, strategies to promote critical thinking – encompassing media literacy, cognitive bias awareness, and the ability to discern correlation from causation – are not supplementary skills, but foundational requirements for responsible AI adoption and the preservation of informed decision-making in an increasingly automated world. Without such vigilance, the potential benefits of AI risk being overshadowed by the erosion of truth and the amplification of societal harms.

The study illuminates a fascinating shift in epistemic vigilance, wherein diminished faith in fellow humans paradoxically bolsters reliance on algorithmic systems. It suggests a trade-off: accepting potential opacity in machine logic for perceived objectivity. As Alan Turing observed, “We can only see a short distance ahead, but we can see plenty there, and we can see plenty further than we ever could before.” This foresight resonates with the findings; the system, while not perfect, offers a glimpse beyond the limitations of human judgment, even if that vision is constructed from data and inference. If the system looks clever, it’s probably fragile – a caution applicable to both human and artificial reasoning, yet increasingly, the latter is being granted the benefit of the doubt.

The Road Ahead

The observation that reliance on artificial intelligence may correlate with diminished trust in human agents is less a revelation than a re-framing of an ancient problem. Throughout history, individuals have sought reliable sources of guidance, often turning to alternatives when established authorities falter. This study’s emphasis on ‘deferred trust’ prompts a crucial question: what, precisely, are individuals optimizing for when they delegate decision-making to an algorithm? Is it accuracy, consistency, or simply the absence of perceived bias – even if that bias is merely human fallibility? The allure of algorithmic objectivity should be examined with care; structure dictates behavior, and the structures encoding these systems are, inevitably, human constructs.

Future research must move beyond documenting this transfer of trust and delve into its consequences. The calibration of trust – determining when and where to defer to AI – appears paramount, yet the mechanisms governing this calibration remain largely unexplored. Simplicity is not minimalism, but the discipline of distinguishing the essential from the accidental. A deeper understanding of the cognitive processes underlying trust transfer is needed, especially concerning the potential for over-reliance, automation bias, and the erosion of critical thinking skills.

Ultimately, the focus should shift from ‘algorithmic trust’ to ‘trust in a system’ – recognizing that AI is but one component of a broader socio-technical landscape. The real challenge lies not in building ‘trustworthy AI’, but in designing systems that foster appropriate trust – a trust grounded in understanding, transparency, and a clear delineation of responsibility.

Original article: https://arxiv.org/pdf/2511.16769.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/