Search Results for « AI-assisted decision-making »

Artificial Intelligence (AI) has the potential to improve human decision-making by providing decision recommendations and problem-relevant information to assist the human decision-maker. However, a number of challenges remain in fully realizing the human-AI collaborative potential. First, we need to understand the conditions that support complementarity where the human performance assisted by an AI exceeds the performance of an unassisted human or the AI. This requires that humans are able to recognize the situations where the AI should be leveraged as well as new AI systems that learn to complement the human decision-maker. Second, we discuss the need to accurately assess the human mental models of the AI that contain expectations of the AI as well as reliance strategies that depend on self-confidence, AI confidence, past performance, and prior beliefs. Third, we discuss the need to understand the effects of different design choices for the human-AI interaction including the timing of AI assistance as well as the amount of model information to present to the human decision-maker to avoid cognitive overload and ineffective reliance strategies. For each of these three challenges, we present an interdisciplinary perspective of recent empirical and theoretical findings and discuss new research directions.

Abstract AI has become a common assistant for making choices, from minor to major ones. It can inform our beliefs relevant to a decision by both helping us to find existing information and generating new information. But in what ways and to what extent is AI useful when making a rational decision? The present paper provides answers to this question for three different types of choices: easy choices, hard choices, and transformative choices. In easy choices, where the rational action is, in principle, straightforward, AI can make the decision-making process more efficient and accurate, increasing derived value (at least in the long-term). In hard choices, where options are on a par, AI can help us when we commit to an option by assisting us in the creation process of new will-based reasons. In transformative choices, where we cannot, even in principle, know by ourselves which option maximizes expected value, AI cannot fill the epistemic or metaphysical gap characteristic of such choices, and therefore cannot enable rational decision-making. Overall, the analysis shows that if the values of our options do not already allow us to determine the rational choice without AI, its assistance does not change that.

In parallel with the quick uptake of artificial intelligence, there has been a heightened concern in organisations about issues of governance, oversight, and accountability. In much of the contemporary debate, it is presumed that as long as accountability is human, decision making will remain unaffected. This paper questions that presumption. It suggests that the influence exerted by artificial intelligence need not be that of formal authority or decision-making in order to have an effect on organisational outcomes. Influence may, rather, develop gradually via recommendations, prioritisation, and decision-making processes, which may change decision-making processes prior to any formal decision being made. Organizations may be concerned with assessing their accountability without considering that decision-making behavior might be changing all the while. The main idea is that whereas accountability and influence traditionally were considered related concepts, they could become separate when it comes to the decision-making processes enabled by AI. It is suggested that measuring the growing disconnect between accountability and decision influence could become an important task for the future.

Abstract Objective : This study evaluates the applicability of artificial intelligence (AI) in clinical decision-making for malignant pleural mesothelioma (MPM) by comparing treatment recommendations from large language models (LLMs) with expert decisions. Methods : A retrospective analysis was conducted on 12 MPM cases treated between 2021 and 2023 at a tertiary university hospital. AI-generated recommendations from ChatGPT, Gemini, and Copilot were compared with multidisciplinary tumor board decisions regarding initial treatment strategy, radiotherapy (RT) timing, target volume delineation, dosimetric assessment, and RT plan approval. AI responses were scored using a 5-point Likert scale by three radiation oncologists. Readability and quality assessments were performed using the DISCERN scale and established readability metrics. Statistical analyses included intraclass correlation coefficients, Friedman tests, and Wilcoxon signed-rank tests with Bonferroni correction. Results : The study analyzed 12 cases with 60 questions comparing three LLMs in mesothelioma treatment decision-making. ChatGPT demonstrated superior performance with the highest mean score (4.50 ± 0.57) and a median score of 5, significantly outperforming Gemini (mean 3.77 ± 0.43, median 4) and Copilot (mean 3.85 ± 0.52, median 4) (p < 0.001). Category-specific analysis showed that ChatGPT consistently excelled across all decision-making domains, particularly in RT timing and dosimetric data evaluation (median scores of 5). It significantly outperformed the other models in four of five categories: Initial Treatment Recommendation, Radiotherapy Timing, Radiotherapy Planning, and Dosimetric Data Evaluation (all p < 0.05). Gemini maintained moderate performance with median scores of 4 across all categories. Copilot showed variable performance with median scores ranging from 3 to 4. In RT Plan Approval, ChatGPT and Gemini performed similarly (p = 1.000), while Copilot scored significantly lower (p = 0.025). ChatGPT achieved the highest DISCERN score (70/75, excellent quality), while Copilot (62/75) and Gemini (61/75) were rated as good. Readability analyses classified all AI outputs as "difficult to read," with Copilot being the most readable (Flesch Reading Ease Score = 35.52). Conclusion : Among the evaluated AI models, ChatGPT provided the most accurate and clinically relevant recommendations for MPM management. While AI tools show promise in decision support, further validation is required before integration into clinical workflows. Future research should focus on enhancing readability and reliability for clinical applications.

Large language models are increasingly embedded in decision-making environments across education, healthcare, law, finance, and software engineering. While these systems improve efficiency and accessibility, they also introduce systematic changes in how humans evaluate information and allocate cognitive effort. This paper presents a structured framework for understanding these changes through three interacting mechanisms: Fluency-Driven Authority (FDA), Cognitive Offloading (CO), and Verification Decay (VD). These constructs describe how fluent AI-generated outputs influence perceived correctness, reduce independent verification, and increase reliance on machine-generated reasoning over time. We integrate established concepts from cognitive offloading, automation bias, trust calibration, and human-computer interaction into a unified model of human-AI cognitive dynamics. A Domain Sensitivity Coefficient (DSC) is introduced to capture how these effects vary across domains with different risk levels and verification costs. The framework provides a synthesis model for analyzing AI-assisted decision-making and is intended as a baseline reference for future empirical research in cognitive science, human-AI interaction, and AI safety.

Abstract Recent advances in generative AI models enabled the creation of digital health assistants for patients. However, it remains unclear how patients - especially in the presence of cognitive biases - would utilize them. Drawing on behavioral decision theory (BDT), we analyzed how bounded rational patients use AI health assistants to make healthcare choices. Our findings show that cognitive biases lead patients to underutilize these assistants, limiting their potential to prompt high-risk patients to seek necessary care and to reduce unnecessary clinical visits among low-risk patients. Moreover, we found that bounded rational patients become less sensitive to differences in risk, and their decision to seek clinical care is determined primarily by the cost of access to healthcare rather than by the underlying health risk. These findings highlight the need for developers to design bias-mitigating interfaces and general transparency in the model, and for policymakers to establish safeguards to support effective adoption of these technologies.

The use of Artificial Intelligence (AI) is on the rise in the Criminal Justice System (CJS), being applied to a variety of functions such as Predictive Policing, Criminal Investigation, Risk Assessment, Evidence Analysis, Sentencing, and Judicial Decision Support.AI possesses several advantages such as improved effectiveness, uniformity, and the capability to process considerable amounts of intricate information in an efficient manner, but it brings serious legal problems connected with Algorithmic Bias, Transparency, Privacy, Accountability, and Procedural Fairness. The challenges related to AI-based evaluations and suggestions are especially crucial when it comes to making decisions which can affect a person's rights and liberty.Thus, the aim of the present research is to investigate the legal consequences of using AI in the decision-making process in CJS considering the interrelation between support of technologies and the established principles of Due Process, Equality Before the Law, and Judicial Independence. Along with that, the research assessed the drawbacks of Algorithmic Risk Assessment such as the possibility of repeating previous injustices due to biased training data, the challenges of Non-Traditional Decision-Making Systems, and the need for Essential Human Supervision.Last but not least, the study evaluated how Artificial Intelligence can be integrated into the Criminal Justice System without sacrificing basic legal safeguards and discovered the measures that would ensure that technological progress goes hand in hand with transparency, justice, accountability, and rule of law.

Abstract Whenever an AI model is used to predict a relevant (binary) outcome in AI-assisted decision making, it is widely agreed that, together with each prediction, the model should provide an AI confidence value. However, it has been unclear why decision makers have often difficulties to develop a good sense on whento trust a prediction using AI confidence values. Very recently, Corvelo Benz and Gomez Rodriguez have argued that, for rational decision makers, the utility of AI-assisted decision making is inherently bounded by the degree of alignment between the AI confidence values and the decision maker’s confidence ontheir own predictions. In this work, we empirically investigate to what extent the degree of alignment actually influences the utility of AI-assisted decision making. To this end, we design and run a large-scale human subject study (n = 703) where participants solve a simple decision making task—an online card game—assisted by an AI model with a steerable degree of alignment. Our results show a positive association between the degree of alignment and the utility of AI-assisted decision making. In addition, our results also show that post-processing the AI confidence values to achieve multicalibration with respect to the participants’ confidence on their own predictions increases both the degree of alignment and the utility of AI-assisted decision making.

Abstract Artificial intelligence (AI) alignment is often framed as the problem of making a model reliably pursue human intentions or values. That work is indispensable, but deployed systems also raise a governance problem: who controls information, private context, correction, escalation, and action? This conceptual article proposes triarchic corrigibility as a role-separated governance design pattern for AI-assisted decision-making. The pattern combines (1) an embodied and answerable human, or an accountable human decision body; (2) a strictly isolated local AI layer for private contextual continuity; and (3) an externally connected AI layer for current public-world evidence. Final authorization of consequential action remains human, while epistemic and procedural competences are deliberately distributed so that no component can unilaterally control framing, memory, procedure, and execution. The article specifies a normative basis, trust-domain separation, decision rights, human-mediated data flows, reciprocal challenge, escalation, affected-party recourse, proportionate audit, and exit procedures. It distinguishes personal and organizational instantiations, compares the pattern with single-assistant, local-only, edge–cloud, multi-agent, human-in-the-loop, meaningful-human-control, and polycentric approaches, and provides criteria and rejection conditions for evaluating its value. It also advances an evolutionary-symbiotic hypothesis: repeated, bounded dependence on complementary roles may make correction and cooperative continuity preferable to unilateral control while preserving interruption, replacement, and exit. A worked case illustrates the protocol. The result is a testable governance hypothesis whose capacity to reduce correlated failure, preserve privacy and human judgment, and sustain correction under residual uncertainty requires empirical evaluation.