Introduction to the Special Issue on the Human-Algorithm Connection

1. Introduction

We are delighted to introduce this special issue of Management Science focused on the “Human-Algorithm Connection.” The call for papers, launched in 2021, aimed to attract research that deepens our understanding of how humans and algorithms coexist and how that relationship can be enhanced to improve decision making. As algorithms become increasingly pervasive, they bring forth a new set of challenges, largely because they do not operate in isolation. They are designed, implemented, and overseen by humans, and their outputs influence human decisions and behaviors both directly and indirectly. Furthermore, algorithms are embedded within organizational contexts, where their purpose should align with strategic goals and mission. At a broader level, their societal impact is becoming increasingly significant. Notably, the final submission deadline for this issue (December 22, 2022) was just a few weeks after the launch of ChatGPT, underscoring the growing urgency for rigorous research in this domain.

This special issue is distinctly interdisciplinary, featuring 39 papers that span a broad range of departmental areas including strategy, finance, operations, marketing, data science, and information systems. The methodologies used are also diverse as shown in Table 1 (some papers use more than one methodology so the last column adds up to more than 39). The papers were selected from a record-breaking 319 submissions, marking an unprecedented level of interest for Management Science.

Table 1. Number of Papers by Methodology

In the following sections, we explain the main idea of each paper that is relevant beyond methods. To provide structure and clarity, we have grouped the papers into four thematic categories: augmentation, algorithm aversion/adoption, algorithmic fairness, and the future of work. Finally, the special issue is about algorithms in general, which includes artificial intelligence (AI) as a very important and timely special case.

2. Augmentation of Human Decision Making with Algorithms

The papers in this section provide key insights or prescription on the interaction between humans and algorithms to perform a specific (type of) task—for example, portfolio rebalancing, sales, forecasting, inventory management, and so on—which is known as augmentation. Rather than replacing humans as in automation, augmentation aims to support and amplify human decision making, creativity, and productivity. Recent work for instance includes algorithmic writing assistance to improve jobseekers’ outcomes in online labor markets (Wiles et al. 2025), human-algorithm complementarity for microloan decisions (Lu and Zhang 2025), algorithmic advice for diagnostic decision making in clinical settings (Yin et al. 2025), and collaborative AI and human domain expertise for process improvement in semiconductor manufacturing (Senoner et al. 2022). In this special issue, we categorize the papers on augmentation in three sections, beginning with augmentation for the customer or end user, then augmentation for customer-facing workers, and finally augmentation for back-end tasks.

2.1. Augmentation Experienced by the End User

The following papers examine how algorithms enhance individual users’ performance, decision making, or workflow by complementing their capabilities.

1. [observational data] In “Human-Robot Interactions in Investment Decisions,” Bianchi and Brière (2024) study how investors interact with a robo-advisor deployed by a large asset management company. The robot tracks the investor’s portfolio and sends a recommendation to rebalance the portfolio when it is too far from the investor’s target allocation. The investor has the option to follow the recommendation but keeps entire control over the portfolio. The authors find that investors who use the robot become more attentive to deviations from their target, rebalance more often, and enjoy higher returns. A nice feature of this setup is that one can measure the monetary impact of letting the human investor take the decision instead of fully automating the rebalancing. The authors find that keeping the human “in the loop” comes with a very small cost.

2. [incentivized experiment] In “Trading Gamification and Investor Behavior,” Chapkovski et al. (2024) conduct experiments on how investors are affected by the “gamification” elements (e.g., “confetti” and achievement badges) used in trading apps like RobinHood or EToro. They show that participants do trade more when a platform is gamified, but 70% of the effect is selection: Less financially literate participants, who always tend to trade more, are significantly more likely to choose a gamified platform. Moreover, some gamification elements make participants more attentive to the trading process, which can enhance their performance. These nuanced results should have an important impact on the controversies regarding the gamification of trading, with lessons for human-algo interactions more generally.

2.2. Customer-Facing Augmentation

In the previous section, the end users themselves are using the algorithm, whereas here it is an employee that is using the algorithm to assist or serve the end user.

• 3. [field experiment] “Human-Centered Artificial Intelligence: A Field Experiment” by Krakowski et al. (2025) is a useful example of how the central understanding of behavior in marketing can be deployed to use AI better. The paper hypothesizes that human-AI interaction should be tailored to individuals’ cognitive preferences, in the context of unstructured managerial tasks as are commonly found in marketing and sales contexts. The paper explores this hypothesis in a field experiment that studies the reactions of a sales force to AI in a multinational pharmaceutical firm. This represents an unparalleled opportunity to study the deployment of AI. The experiment manipulated four contextual parameters of human-AI interaction—work procedures, decision-making authority, training, and incentives. The manipulation attempted to align salespeople’s cognitive styles. The authors find that the tailoring of interactions improved the performance of sales people, but if AI was not tailored, it could perform worse than baseline. The authors present convincing qualitative evidence that suggests that this negative outcome arises from role conflicts, ambiguities and lack of use. This emphasizes that it cannot be assumed that all humans will align the same way with AI and suggests that therefore it is helpful for AI deployment to be tailored to individuals’ cognitive styles.

• 4. [field experiment] In “Engaging Customers with AI in Online Chats: Evidence from a Randomized Field Experiment,” Zhang and Narayandas (2025) explore the use of AI explicitly in the context of customers and their reactions when a firm augments customer support with AI to facilitate customer interactions. They also use a field experiment to explore this question, but this time in the context of a meal delivery platform. They find that giving customer service agents access to AI-generated suggestions improved both the efficiency and the effectiveness of these interactions. The agents who had access to AI assistance responded faster, engaged customers more deeply, and achieved greater improvements in customer sentiment. These benefits were most pronounced for less-experienced agents, which is suggestive of AI potentially equalizing employee performance. They did find a few caveats. The impact of AI varied with the type of conversation. It was very useful for interactions when customers wanted to cancel things, but it was less effective in scenarios where a customer already had a complaint and was trying to escalate it, because such requests appeared to be beyond the AI’s capability. There was also a further interesting twist. If the customer had first chatted with an AI chatbot, and they went on to interact with a customer service agent who used AI, the customer became unhappy because the quicker responses of the agent made the customer suspect they were still communicating with a chatbot.

2.3. Back-End Augmentation

The papers in this section focus on human-algorithm augmentation for back-end processes or operations.

• 5. [field experiment] In “The Power of Disagreement: A Field Experiment to Investigate Human-Algorithm Collaboration in Loan Evaluations,” Wang et al. (2025) use the domain of financial loan evaluations to investigate human-algorithm collaboration, an increasingly common setting that is more likely to be fruitful when the two entities are complementary and when humans can call out algorithm decisions when those decisions are wrong. They deploy a field experiment where human evaluators and algorithms worked together to evaluate loan applications. Testing such collaboration across four scenarios (by varying the richness of information needed in decision-making and disclosure of algorithm rationale), they found that such collaboration was fruitful (relative to algorithm-only or human-only work) in all four scenarios, and disagreement was necessary for collaborative value (because otherwise the collaborative outcome would be no different than algorithm only), but only when used wisely. Collaborative value increased significantly when humans used signs of algorithm contradiction (inconsistencies) to override flawed algorithmic recommendations.

• 6. [field experiment] “Profit Implications of Judgmental Adjustments to Forecast Inputs: Evidence from a Large-Scale Field Experiment” by Kesavan et al. (2025) presents causal evidence from a large-scale field experiment in a U.S. automotive spare parts retail chain, showing that allowing human judgment to adjust forecast inputs to inventory algorithms improves profits by 4.92% on average. The study reveals that human judgment adds value selectively. It most benefits profits for stock keeping units with high margin, less historical data, and from large suppliers. This nuanced understanding challenges the assumption that automation should always be preferred and highlights the importance of designing hybrid decision systems that strategically incorporate human input. Finally, the paper shows that forecast accuracy alone is not a sufficient metric for evaluating algorithmic performance—profit outcomes matter more.

• 7. [incentivized experiment] “Managerial Insight and ‘Optimal’ Algorithms” by Flicker (2025) introduces Forecast to Individualized Normative Distribution (FIND), a method that helps firms leverage managers’ private, and often unstructured information, to improve inventory management decisions. Traditional automated systems miss context-specific signals that managers observe. However, managers often struggle to translate their insight into optimal actions due to cognitive biases. FIND bridges this gap by converting managers’ forecasts into debiased, conditional demand distributions. Across eight experiments, including those with perceptual signals like images and sounds, FIND consistently outperforms both human direct orders and algorithmic benchmarks, offering a practical framework for human-algorithm collaboration.

• 8. [theoretical, implementation] In “Algorithmic Precision and Human Decision: A Study of Interactive Optimization for School Schedules,” Lian et al. (2025) present an interactive optimization framework that successfully helped the San Francisco Unified School District redesign school start times, balancing transportation costs, policy constraints, and stakeholder preferences. Rather than relying solely on algorithmic outputs, the approach empowers policymakers to explore thousands of near-optimal solutions using intuitive tools, enabling real-time feedback and objective discovery. This collaboration led to the first optimization-driven schedule change in a U.S. school district, saving $5 million annually and improving stakeholder satisfaction. The broader insight is that combining algorithmic precision with human judgment through interactive tools can overcome policy resistance and improve implementation in complex multiobjective public decisions.

• 9. [theoretical, case study] In “Reciprocal Human-Machine Learning: A Theory and an Instantiation for the Case of Message Classification,” Te’eni et al. (2025) rethink the value of keeping a “human in the loop,” challenging the view that humans should primarily teach or oversee AI. They address this gap by introducing Reciprocal Human–Machine Learning (RHML): a framework for a learning partnership where humans and algorithms iteratively exchange feedback to coevolve their understanding. The authors instantiate RHML in Fusion, a novel system developed through two case studies in cybersecurity forums (focused on drug trafficking and hacker attacks). Across eight learning cycles, both humans and machine learning (ML) models improved their classification performance. The study not only demonstrates the feasibility of RHML but also provides formal design principles to guide the future development of collaborative human-AI systems. RHML will likely become an essential form of learning in organizations.

3. Algorithm Aversion/Adoption

Algorithms are usually implemented in practice as systems that make recommendations, which a human decision maker can accept, adjust, or simply reject. In this context, it has been well documented that humans might display algorithm aversion (Davis et al. 2024), undermining the adoption of the algorithm and its recommendations, which can lead to detrimental performance (Ibanez et al. 2018, Kesavan and Kushwaha 2020, Gnewuch et al. 2024). Previous studies have shown that users might deviate if they perceive that the algorithm does not consider relevant costs or information (Van Donselaar et al. 2010) or because of behavioral biases (Burton et al. 2020). Algorithm aversion can be mitigated by giving the human additional control (Dietvorst et al. 2018, Kawaguchi 2021), by well-timed transparency (Luo et al. 2019), by enhancing the algorithm (Sun et al. 2022), or by improving its user interface (Caro and de Tejada Cuenca 2023).

The study of how people accept or reject algorithms has become a vibrant research area. Reflecting this trend, the special issue includes 11 papers focused on algorithm aversion/adoption, which are summarized below.

• 10. [field experiment] “Algorithm Aversion: Evidence from Ridesharing Drivers” by Liu et al. (2023), unlike the papers discussed in Section 2.2 that randomized whether customer-facing employees had access to AI, instead looks at factors driving whether users adopt AI in the first place. This is important because a lack of adoption may hinder the benefits of AI within an organization, especially if the AI system is focused on improving the organization rather than necessarily the individual’s own experience. The authors explore this question in a field study focused on the roll-out of an algorithmic recommendation on a large ride-sharing platform, that was designed to help the drivers using its platform make better location decisions. The authors find that the drivers were less likely to follow the algorithm when the algorithmic recommendation did not align with their past experience at how effective a specific location was at a certain time, and when the actions of peer drivers contradicted the algorithms’ recommendations. This paper therefore provides some useful, potentially generalizable instances where algorithmic aversion and consequently a lack of adoption might be expected in AI roll-out.

• 11. [field experiment] In “Using AI and Behavioral Finance to Cope with Limited Attention and Reduce Overdraft Fees,” Ben-David et al. (2025) study a commercial app that helps customers manage their finances. The app uses an AI algorithm to predict when a customer’s bank balance is likely to run below zero and then sends a warning so that the customer can avoid overdraft fees. Surprisingly, many customers fail to act upon the warning. The authors find that the framing of the warning matters significantly and that simpler notifications elicit higher response rates. Thus, even in this quite unique setting in which humans have no reason to not trust the algorithm’s advice, the algorithm is still underused, because humans have limited attention. Maximizing the use of the algorithm requires designing the human-algorithm interface in a way that minimizes the scope for such biases.

• 12. [field experiment] In “Identity Disclosure and Anthropomorphism in Voice Chatbot Design: A Field Experiment,” Xu et al. (2024) investigate identity disclosure and anthropomorphism in the design of a voice chatbot for a logistics dispatcher. In a large-scale randomized field experiment with more than 11,000 truck drivers, the authors find that disclosing a chatbot’s identity reduces user engagement due to algorithm aversion. However, the negative effect of identity disclosure is mitigated by humanizing the chatbot using interjections and filler words, suggesting that trust and perceived human-likeness play a key role in user interaction. These results highlight the importance of designing AI systems that foster trust and social connection, even when transparency is mandated, and they provide guidelines for the deployment of AI in customer-facing roles across industries.

• 13. [field experiment] In “My Advisor, Her AI, and Me: Evidence from a Field Experiment on Human-AI Collaboration and Investment Decisions,” Yang et al. (2025) study a timely question in AI governance and service design: Could keeping a human in the loop of AI-based decision making benefit end users? Using a field experiment in a real-world financial advisory setting, the study offers a surprising insight: Although human involvement does not compromise the quality of AI-based financial advice, it significantly increases consumer trust and uptake—especially for more risky investments. This heightened engagement, in turn, enhances consumer welfare. Importantly, the benefit of human involvement cannot be attributed to a perceived improvement in advice quality due to complementarities between human and AI capabilities. Rather, its value lies in the enhanced affective appeal of the advice, building consumer trust. These findings offer actionable implications for firms, policymakers, and designers of AI systems, highlighting how a thoughtful human-in-the-loop approach can foster trust and accelerate AI adoption.

• 14. [theoretical, incentivized experiment] In “Human-Algorithm Collaboration with Private Information: Naïve Advice-Weighting Behavior and Mitigation,” Balakrishnan et al. (2025) make a crucial theoretical contribution to understanding when and why human-algorithm collaboration fails in practice. It addresses a fundamental paradox: Although algorithms often outperform humans on average, humans sometimes possess valuable private information algorithms cannot access, yet human overrides frequently degrade rather than improve performance. The authors propose this stems from humans’ systematic bias toward “naive advice weighting” (NAW)—taking constant weighted averages between their predictions and algorithmic recommendations regardless of whether private information is actually valuable. Through three experiments involving demand prediction tasks, they demonstrate NAW leads to predictable patterns of over-adherence when private information is valuable and under-adherence when it’s not. Most importantly, they show “feature transparency”—simply telling users which inputs the algorithm considers—can mitigate this bias by helping humans better discriminate when their private information warrants deviation. The research moves beyond traditional focus on algorithm aversion to examine nuanced cognitive processes determining when human oversight adds versus destroys value.

• 15. [incentivized experiment] “Aversion to Hiring Algorithms: Transparency, Gender Profiling, and Self-Confidence” by Dargnies et al. (2024) addresses a critical barrier to algorithmic adoption in hiring by examining why both job seekers and managers resist AI-driven hiring decisions despite evidence of superior algorithmic performance. Through an online experiment with 744 workers and 754 managers, the authors investigate three key policy levers: gender-blind algorithms, algorithmic transparency, and performance feedback. The paper reveals preferences for algorithmic hiring are not uniform across stakeholders. Workers strongly prefer gender-blind algorithms (59% versus 47% adoption), suggesting a reluctance to profiling by gender, whereas algorithmic transparency has surprisingly little impact. Managerial overconfidence creates substantial barriers to efficient algorithm adoption. Providing managers feedback about their hiring performance dramatically increases delegation to algorithms (from 34% to 50%), offering a practical intervention for overcoming resistance. This work shows algorithm aversion is not simply about trust or understanding technology, but is fundamentally shaped by self-interest, fairness perceptions, and cognitive biases like overconfidence.

• 16. [incentivized experiment] “Incentives, Framing, and Reliance on Algorithmic Advice: An Experimental Study” by Greiner et al. (2025) investigates how financial incentives and the framing of algorithmic advice affect decision makers’ reliance on AI tools. In a large experiment, participants estimated Airbnb prices with or without algorithmic support and under different compensation schemes. The study finds that both performance-based and tournament incentives increase reliance on algorithmic advice and effort, without reducing performance, contradicting earlier claims that incentives may backfire. Additionally, framing the algorithm as incorporating human expertise boosts trust and advice uptake, especially among those with fixed pay. These findings suggest that organizations can encourage effective use of AI tools either through financial incentives or by emphasizing human involvement in algorithm design, offering practical strategies to mitigate algorithm aversion.

• 17. [incentivized experiment] In “Humans’ Use of AI-Assistance: The Effect of Loss Aversion on Willingness to Delegate Decisions,” Bockstedt and Buckman (2025) address the challenge of algorithm aversion—people’s reluctance to delegate tasks to AI—even when AI outperforms humans. The authors show that reframing incentives can reduce this bias. Across multiple experiments, participants preferred human help when rewards were framed as gains but became more open to AI assistance when poor performance was framed as losses. This shift was consistent across task complexity and incentive levels and was linked to increased situational awareness. The study offers practical guidance for managers on how simple loss-framing strategies can encourage more effective human-AI collaboration and improve task delegation outcomes.

• 18. [incentivized experiment] In “Till Tech Do Us Part: Betrayal Aversion and Its Role in Algorithm Use,” Kormylo et al. (2025) provide the first direct test of algorithmic betrayal aversion, finding that under conditions of betrayal risk, individuals reject advice from human experts, leading to a 20% drop in earnings, but not from algorithms. In a financial market experiment with identical monetary risks across human and algorithmic advisors, the study isolates betrayal aversion as an emotional, nonmonetary barrier to expert advice uptake. The effect replicates across populations and settings, and for algorithms, any initial aversion dissipates quickly. The findings have broad implications for financial services, healthcare, and other domains where replacing or augmenting human experts with algorithms could reduce betrayal aversion and subsequently increase adherence to high-quality advice.

• 19. [incentivized experiment] “Algorithm Reliance, Fast and Slow” by Snyder et al. (2025) contributes significantly to understanding human-algorithm collaboration by examining operational efficiency in service contexts rather than just decision quality. Although most research on algorithm aversion studies whether people use algorithmic advice, this work recognizes that in queueing systems, how people use algorithms critically affects system performance. The authors develop a novel experimental approach where participants recommend jokes to customers in a simulated service system, capturing the dual pressures service workers face: making good decisions while working quickly under time constraints. Their key insight is that algorithm reliance exists on a spectrum from deliberative consultation to automatic defaulting, with profound implications for throughput times. The paper introduces “learning by using” as a mechanism for workers to evaluate algorithm quality through repeated interaction without explicit performance feedback. Most importantly, the research demonstrates that algorithms can even harm efficiency if system conditions do not encourage fast adoption, revealing a critical implementation barrier.

• 20. [incentivized experiment] In “Digital Lyrebirds: Experimental Evidence that Voice-Based Deep Fakes Influence Trust,” Schanke et al. (2024) explore how pairing audio chatbots with voice clones—AI-generated replicas of users’ voices—affects consumer trust. In investment game experiments, participants showed significantly greater trust in AI agents that used their own voice, reflecting a heightened willingness to engage. Although voice cloning increased trust, AI disclosure did not reduce this effect. The trust appears to stem from perceived homophily, or a sense of similarity, and grows with the quality and clarity of the audio. These findings highlight both the promise and risks of voice cloning: consumer trust can be manipulated through voice similarity, while the current regulatory approach of disclosure may be insufficient. The study offers timely insights for consumers, policymakers, and designers of voice-based AI systems seeking to balance personalization and protection.

4. Algorithmic Fairness

Algorithmic fairness concerns outcomes that disproportionately affect certain population subgroups. It is a primary concern that only grows in importance as algorithms continue to permeate almost every aspect of human life. De-Arteaga et al. (2022) provide an early survey on the subject. More recent research by Chen et al. (2025) has shown that large language models (up to GPT-4), which are trained with data generated by humans, preserve or even exacerbate common behavioral decision biases and therefore can propagate fairness issues. In that vein, algorithms have been shown to produce unfair outcomes in career ad targeting (Lambrecht and Tucker 2019), search advertising (Lambrecht and Tucker 2024), and credit markets (Fuster et al. 2022). However, there is also research concluding the opposite. For instance, Ganju et al. (2020) find that the deployment of a clinical decision support system reduced racial disparities in amputation and revascularization rates. Similarly, Bai et al. (2022) show that pickers at an Alibaba Group warehouse perceived an algorithmic (versus human-based) assignment process as fairer.

Solving the fairness conundrum remains an open question. The special issue adds to the discussion with six papers on algorithmic fairness in relevant contexts, including two that are specifically about fairness in hiring.

• 21. [theoretical, observational data] In “The Fairness of Credit Scoring Models,” Hurlin et al. (2024) address the thorny issue of algorithmic bias in lending decisions. They propose a powerful diagnosis tool to determine whether an algorithm is “fair,” or on the contrary, discriminates against some demographic groups. In the latter case, they propose a method to determine which features cause the problem and restores fairness. Beyond the application to credit scoring, the paper answers a more general need to be able to use the most accurate prediction algorithms, which can often be “black boxes,” while also making sure that these algorithms are not biased or discriminatory.

• 22. [theoretical] In “The Gatekeeper Effect: The Implications of Pre-Screening, Self-Selection, and Bias for Hiring Processes,” Koren (2024) examines candidate screening in hiring decisions mediated by a preselection process. A gatekeeper assesses candidates before a full evaluation, aiming to reduce workload by filtering out unpromising candidates. This step may be performed algorithmically for greater cost savings. Introducing this step may affect candidates’ decisions at equilibrium because they anticipate the screening outcome and choose whether to apply accordingly. The main finding connects the correctness of the gatekeeper to the overall quality of the selection process, showing that quality improves with a gatekeeper that has sufficiently high-quality signals. When candidates are drawn from two population segments, those who are underrepresented tend to self-exclude, thereby reducing the overall quality of the process.

• 23. [theoretical, numerical simulation] In “On Statistical Discrimination as a Failure of Social Learning: A Multiarmed Bandit Approach,” Komiyama and Noda (2024) explore hiring decisions from the perspective of fairness, with the goal of reducing bias. Considering a scenario with two similar population segments, a rational and nonprejudiced hiring firm that does not know that both segments are similar seeks to hire the best candidate. When one segment is larger, the firm has more prior information about it, which can cause the quality of candidates from the smaller segment to be underestimated. The central concern is that such biased beliefs may persist over time as the firm updates its posterior beliefs using a multiarmed bandit approach. The proposed solution is a temporary subsidy that mitigates the data scarcity in the smaller segment, allowing the firm to correct its beliefs and improve its ability to identify and hire the best candidates.

• 24. [theoretical, numerical simulation] “Learning to Be Fair: A Consequentialist Approach to Equitable Decision Making” by Chohlas-Wood et al. (2024) explores the question of algorithmic justice in a marketing context where the focal algorithmic justice concern is equalizing decisions, outcomes, or error rates across race or gender groups. These authors consider a hypothetical government ride-share program that provides transportation assistance to low-income people with upcoming court dates. The algorithmic justice literature might suggest that the AI might allocate rides to those with the highest estimated treatment effect per dollar, while constraining spending to be equal by race. The authors point out, however, that such a mechanism would ignore the downstream consequences of such constraints. For example, if one racial group systematically lives further from court, enforcing equal spending would mean fewer total rides provided to them and more people penalized for missing court. The authors suggest that it is first important to elicit stakeholder preferences over the space of possible decisions and the resulting outcomes—such as preferences for balancing spending parity against court appearance rates. The authors then show how to optimize these elicited preferences over the space of such decision policies, making tradeoffs in a way that maximizes elicited utility.

• 25. [natural experiment] In “Can Artificial Intelligence Improve Gender Equality? Evidence from a Natural Experiment,” Bao et al. (2024) provide compelling evidence for AI’s potential to address persistent gender disparities in education through its inherently neutral design characteristics. It leverages a natural experiment where COVID-19 quarantines forced random replacement of human teachers with AI instructors at a Go training academy, creating ideal conditions to isolate AI’s effects on gender differences. The research demonstrates AI can simultaneously improve overall learning outcomes while specifically reducing pre-existing gender performance gaps—a dual benefit rarely observed in educational interventions. The authors show AI’s advantages stem not just from superior analytical capabilities, but crucially from emotional neutrality: while human teachers unconsciously exhibited gender-biased emotions, AI maintained consistent emotional responses regardless of student gender. This finding suggests AI’s lack of human-like social biases can be leveraged as a strength in contexts where discrimination is problematic, offering a novel pathway for addressing gender disparities through substituting biased human judgment with unbiased algorithmic assessment.

• 26. [structural estimation] In “Human–Algorithmic Bias: Source, Evolution, and Impact,” Hu et al. (2025) investigate the origins and effects of human-algorithmic bias using a unique repeat decision-making setting in a high-stakes microlending context. By estimating a structural model, the authors identify both preference-based and belief-based biases in human evaluators—favoring female applicants. Counterfactual simulations show that removing either type of bias improves both fairness and platform profits. The study then examines how these biases are inherited by ML algorithms. Training ML models on real and bias-adjusted datasets reveals that even fairness-unaware ML can mitigate human bias; although eliminating either type of human bias from the training data can further improve ML fairness, the fairness-enhancing effects vary significantly between new and repeat applicants. The findings offer valuable insight into how bias evolves across human and algorithmic decisions and suggest practical strategies for reducing bias in human-AI pipelines.

5. Future of Work

Algorithms, and especially AI, are certain to shape the future of work, although the extent remains to be seen. Research in this area is growing rapidly and is organized here into three categories: workforce management with algorithms, strategic response, and other relevant topics.

5.1. Workforce Management with Algorithms

Prior research has shown that, when there’s complementarity, the collaboration between humans and algorithms can be advantageous, even accounting for humans’ limited cognitive capacity (Boyacı et al. 2024). On the other hand, humans generally exhibit a preference for collaborating with other humans, and algorithmic agents cannot yet be presumed to function effectively as team members (Dell’Acqua et al. 2025). In a similar spirit, the papers in this section provide general guidelines and methodologies to organize or optimize work between humans and algorithms, in contrast to the papers on augmentation in Section 2 that refer to a specific task (e.g., forecasting).

• 27. [observational data] In “Reskilling the Workforce for AI: Domain Expertise and Algorithmic Literacy,” Tambe (2025) tackles another critical question about AI’s value: It shows that AI’s impact is amplified when technical capabilities are distributed across employees with domain expertise. The implications are far-reaching. For policymakers, it underscores the need to reskill nontechnical workers. For business leaders, it calls for broader hiring and training strategies—beyond investing in data scientists—to promote algorithmic literacy across the organization. For educators, it challenges the notion that AI is reserved for a technical elite, showing that tools like no-code platforms can democratize access and suggesting a need for curricular reorientation. Ultimately, the paper reframes AI not as a specialized technical challenge but as a workforce-wide opportunity essential for digital transformation.

• 28. [theoretical, numerical simulation] In “Roles of Artificial Intelligence in Collaboration with Humans: Automation, Augmentation, and the Future of Work,” Fügener et al. (2025) explore the pressing question of how AI will reshape human judgment tasks in the future of work. Although public debates often present AI adoption as a choice between automation and augmentation, the study offers a more nuanced perspective. It shows that the optimal use of AI depends not only on its capabilities but also on how human and AI capabilities complement each other. A key contribution is the identification of different types of complementarity, which help determine when AI should augment, automate, or let the human work without AI support. The authors also highlight an overlooked design lever: reallocating human labor released by AI automation to tackle the most difficult tasks. Taken together, these insights offer a practical framework for designing AI roles that elevate, rather than replace, human input, moving the conversation from “AI versus humans” toward hybrid models of collaboration.

• 29. [theoretical] “Is Your Machine Better Than You? You May Never Know” by de Véricourt and Gurkan (2023) provides a theoretical framework for understanding why human-machine collaboration persists even when machines might outperform humans. It addresses an important puzzle: Despite algorithms often exceeding the accuracy of human experts, professionals in high-stakes domains like medical diagnosis and judicial decisions continue to override the machines instead of fully adopting or rejecting them. Rather than attributing this solely to algorithm aversion, the authors demonstrate that high-stakes decision-making contexts create structural barriers to learning a machine’s quality. Through an analytical model incorporating verification bias and exploration-free behavior, they show that decision makers can become trapped in states of perpetual uncertainty about machine capabilities. This reframes human-machine complementarity from an assumption about comparative advantage to an emergent outcome of learning constraints. The work suggests that prolonged uncertainty or unanimous team beliefs signal better machines, whereas persistent overriding indicates worse ones.

• 30. [theoretical, incentivized experiment] In “Behavioral Externalities of Process Automation,” Beer et al. (2025) investigate how process automation affects the productivity of human workers in collaborative tasks. Using a stylized model and behavioral experiments, the authors show that, although automation improves overall project completion rates and reduces completion times, it unexpectedly reduces worker productivity. This effect is greater among workers who care about how their performance affects coworkers’ pay. The study also finds that automating upstream tasks is more beneficial than automating downstream ones, because upstream human workers tend to delay more. These findings challenge the assumption that automation always enhances productivity because the absence of a human partner might reduce the behavioral pressure to perform efficiently.

• 31. [incentivized experiment] In “Interacting with Man or Machine: When Do Humans Reason Better?,” Bayer and Renou (2024) provide insight into how collaboration partner type fundamentally alters human cognitive processes, with direct implications for optimal AI deployment strategies. It addresses a gap in human-AI collaboration research by demonstrating that humans do not simply perform better or worse with AI—they reason fundamentally differently depending on whether they believe they are working with humans or machines. Using the red-hat puzzle as a controlled reasoning task, the authors reveal that humans excel at simple collaborative reasoning when paired with other humans but perform better on complex problems when working with AI. Importantly, this difference stems not from AI being nonhuman, but from humans’ knowledge that AI reasons correctly and reliably. The research suggests effective AI deployment requires matching task complexity to appropriate collaboration mode: human teams for problems requiring intuitive, empathetic reasoning, and human-AI teams for complex problems where algorithmic consistency provides better basis for human decision making.

5.2. Strategic Response

This section includes papers that account for strategic or gaming behavior and how it can influence algorithm design or user interaction. From a methodological standpoint, these papers solve for an equilibrium involving humans and algorithms in a game-theoretic setting. A good example from the extant literature is algorithmic pricing in competitive markets (Miklós-Thal and Tucker 2019, Calvano et al. 2020).

• 32. [theoretical, structural estimation] In “Strategic Responses to Algorithmic Recommendations: Evidence from Hotel Pricing,” Garcia et al. (2025) examine how human managers in the hotel industry interact with a price-recommendation algorithm when they have the option to leave prices unchanged at minimal cost. Theoretical and empirical analyses show that price adjustment frictions generate strategic conflicts. The algorithm exaggerates its price recommendations to encourage more frequent updates. However, managers still update too infrequently, resulting in measurable revenue losses. Although managers possess valuable private information that the algorithm lacks, the human cost of overcoming inertia outweighs this advantage. Eliminating these price-setting frictions by fully delegating pricing to the algorithm substantially improves outcomes. These insights apply broadly to human-algorithm interactions involving a low-effort status quo option. Examples include retailers’ inventory restocking, investors’ portfolio rebalancing, and digital marketers’ campaign bid adjustments. In such settings, humans tend to stick with defaults despite having valuable additional information.

• 33. [theoretical] In “When Emotion AI Meets Strategic Users,” Yu et al. (2025) examine the strategic risks and societal implications of deploying emotion AI for resource allocation in settings such as customer care. When users can benefit from appearing distressed, they may game the system by misrepresenting their emotions—potentially undermining the AI’s effectiveness. Using a game-theoretic model, the authors show that emotion AI remains valuable when the spillover effects of negative emotions are small relative to the cost of resource misallocation, even in the presence of algorithmic noise and strategic behavior. Notably, stronger AI is not always socially optimal, and regulatory oversight may be necessary. The study also compares emotion AI with human employees in both allocation and monitoring roles, finding that algorithmic noise can sometimes increase AI profitability. These insights offer practical guidance for responsibly designing, adopting, and regulating emotion AI to anticipate human-AI dynamics in real-world, high-stakes environments.

• 34. [theoretical, numerical simulation] In “Offline Reinforcement Learning for Human-Guided Human-Machine Interaction with Private Information,” Fu et al. (2025b) investigate human-machine interactions in a decision-making context with private information using a reinforcement learning framework in a two-player, turn-based cooperative game. The model is motivated by recommender systems, where humans receive information from the platform but do not reveal private information (e.g., mood), whereas the platform seeks to provide effective recommendations and may also possess private information (e.g., content ratings provided by all users). A key challenge arises because the offline data set does not include private information, introducing bias during model training. The main contribution is a novel identification procedure that treats the other player’s previous action as an instrumental variable to identify the causal effect of a treatment under unmeasured confounding, combined with structural equations and a sieve minimum distance estimator for policy evaluation. This is used to propose a pessimistic off-policy learning algorithm with theoretical convergence guarantees.

• 35. [theoretical, numerical simulation] In “The Best Decisions Are Not the Best Advice: Making Adherence-Aware Recommendations,” Grand-Clément and Pauphilet (2024) examine the consequences of human decision makers not consistently following algorithmic recommendations. This challenge is common in domains where humans leverage algorithms to improve decision making, requiring algorithms to anticipate humans’ reactions and design policies that remain effective in practice. An optimal policy may turn out to be inefficient if the human who is supposed to implement it does not follow it. The main contribution hinges on computing policies that are robust to human deviations. This is achieved by considering a Markov decision process with a baseline policy that captures typical human behavior, establishing a framework for the algorithm to account for deviations from the baseline. The outcome is a set of improved policies that are guided by optimality but allow humans to deviate without a severe loss in performance.

• 36. [incentivized experiment] “Strategic Inattention in Product Search” by Hillenbrand and Hippel (2025) represents a paradigm shift in understanding consumer protection in algorithmic markets. Although digital platforms increasingly use search behavior for price discrimination, existing policy frameworks assume transparency solves the problem. This research fundamentally challenges that assumption by demonstrating that even perfectly informed, optimally behaving consumers cannot escape welfare losses from algorithmic pricing. The work advances understanding of human-algorithm interactions by showing the problem is not consumer sophistication or awareness but rather structural power imbalance inherent in data-driven markets. By showing strategic consumer behavior—although individually rational—still reduces overall welfare, the authors expose critical flaws in disclosure-based regulatory approaches. This finding has profound implications for competition policy and platform regulation, suggesting effective governance of algorithmic markets requires structural interventions rather than relying on informed consumer choice to discipline platforms.

5.3. Privacy, Interpretability, and Organizational Alignment

The papers in this section study other important topics on human-algorithm interactions that have received considerable attention in the literature and can impact the future of work. These key areas are privacy (Tucker 2018, Fu et al. 2025a), AI interpretability (de Véricourt and Perakis 2020, De Bock et al. 2024), and organizational alignment (Dixon et al. 2021).

• 37. [incentivized experiment] In “Speaking in Private: Privacy Expectations Depend on Communication Modality,” Melzner et al. (2024) delve into how privacy matters for interactions between humans and AI. This is a natural question in a marketing context, given the recent focus of privacy regulation and protections on marketing contexts. These privacy concerns arise because consumers disclose personal information when they interact with connected technologies. The advent of voice technology has enabled consumers to interact with connected technologies not only through typing but also through speaking. The paper investigates in three studies whether consumers expect different levels of privacy for information they disclose via different types of communication. The findings suggest that consumers have more restrictive privacy expectations for information disclosed via speech compared with text. The authors suggest that the effect is driven, at least in part, by increased feelings of ownership over content disclosed via speech compared with text.

• 38. [theoretical, incentivized experiment] In “Improving Human Sequential Decision Making with Reinforcement Learning,” Bastani et al. (2025) address a fundamental challenge in human-AI collaboration: extracting actionable insights from complex algorithmic decision making and communicating them effectively to human workers. Although most prior work focuses on one-shot decisions, this tackles sequential decision making, where current actions affect future states and outcomes. The key innovation lies in developing an interpretable reinforcement learning algorithm that identifies optimal actions and specifically targets the gap between human and optimal policies to generate concise, actionable “tips” workers can operationalize. Through large-scale behavioral experiments with more than 2,300 participants, the authors demonstrate that effective human-AI collaboration requires more than just providing correct advice—It demands understanding how humans perceive, adopt, and integrate algorithmic recommendations over time. Their findings reveal workers do not blindly follow tips but combine them with experience to discover additional strategies, suggesting a nuanced model of human-AI partnership. Interventions designed merely to improve compliance may not translate to performance gains.

• 39. [observational data] In “Artificial Intelligence, Lean Startup Method, and Product Innovations,” Wang and Wu (2025) tackle a core tension in the digital economy: why AI adoption so often fails to produce meaningful product innovation. As firms increasingly invest in AI technologies, the gap between potential and realized value has become a pressing concern—not only for academics but also for executives, investors, and policymakers. This research offers timely, actionable insights into the organizational complements needed to unlock AI’s potential, particularly in high-growth, high-uncertainty environments such as startups. The paper’s central contribution is to distinguish between discovery-oriented and optimization-oriented AI tools—and its linkage to specific organizational practices within the lean startup method. This reframing moves beyond the generic notion of “AI adoption” and demonstrates how different types of AI support distinct innovation pathways. The core insight—that AI’s value depends on organizational design and alignment—extends well beyond startups.

6. Final Remarks

Comparing human and algorithmic decision making remains challenging (Kleinberg et al. 2018), particularly as individuals may alter their behavior in response to algorithmic presence—for example, shifting their error profile from false negatives to false positives (Almog et al. 2025). Progress has been made through laboratory and field experiments in specific contexts, and theoretical models are already unveiling important insights, as demonstrated by several papers in this special issue. Additional research is still needed to establish unifying frameworks and identify conditions under which unchecked algorithms might defy their purpose (Lu and Tomlin 2025). Likewise, there is room for deeper exploration into the future of work, especially regarding mechanisms to incentivize human effort, delegate tasks, and distribute the gains from AI (Colliard and Zhao 2025, Zhang et al. 2025).

From a broader perspective, emerging research is examining how organizations will adapt to the growing presence of AI (Hillebrand et al. 2025). This adaptation introduces further complexities. For instance, on the issue of algorithmic fairness, the multiplicity of goals might require rethinking how traditional fairness constraints are imposed on algorithms (Ge et al. 2023). As the use of AI becomes the norm, questions arise about the implications for society and the flywheel of innovation. One concern is that algorithmic outputs may become increasingly homogenized, especially when trained on AI-generated data, potentially triggering a self-reinforcing cycle of uniformity (Castro et al. 2023). Similarly, continuous dependence on AI tools may hinder hands-on learning and reduce opportunities for independent problem solving, potentially impairing long-term skill development (Kim et al. 2024). Management science scholars are well positioned to contribute to these and other pressing questions surrounding the human-algorithm connection.

Associate Editors

Table

Abhishek, Vibhanshu	Dai, Tinglong	Kübler, Dorothea	Shi, Mengze
Acquisti, Alessandro	Davis, Andrew	Lambrecht, Anja	Simmons, Joseph
Adamopoulos, Panagiotis	den Boer, Arnoud	Larrick, Richard	Simonsohn, Uri
Agarwal, Ashish	Derdenger, Timothy	Lee, Dokyun	Singh, Param Vir
Agha, Leila	Dzyabura, Daria	Leider, Stephen	Sliwka, Dirk
Agrawal, Ajay	Eckles, Dean	Li, Beibei	Soll, Jack
Amaldoss, Wilfred	Elmachtoub, Adam	Li, Danielle	Song, Hummy
Anand, Krishnan	Englmaier, Florian	Li, Jun	Staats, Bradley
Asparouhova, Elena	Ferreira, Pedro	Lin, Chen	Su, Xuanming
August, Terrence	Fosfuri, Andrea	Liu, Tracy	Sun, Monic
Baillon, Aurelien	Gallino, Santiago	Mehra, Amit	Sutter, Matthias
Balseiro, Santiago	Gans, Joshua	Menon, Tanya	Tambe, Prasanna
Ban, Gah-Yi	Goel, Sharad	Mihm, Jurgen	Thomadsen, Raphael
Basten, Rob	Goldfarb, Brent	Miklós-Thal, Jeanine	Tian, Xuan
Bayati, Mohsen	Greenwood, Brad	Moreno, Antonio	Tong, Jordan
Beil, Damian	Greiner, Ben	Overby, Eric	Trautmann, Stefan
Bell, David E.	Grushka-Cockayne, Yael	Padmanabhan, Balaji	Trichakis, Nikolaos
Bernstein, Shai	Gupta, Vishal	Park, Andreas	Trueblood, Jennifer
Bolton, Gary	Haruvy, Ernan	Pelger, Markus	Tsoukalas, Gerry
Bossaerts, Peter	Honhon, Dorothee	Phillips, Robert	Viswanathan, Siva
Boudreau, Kevin	Hossain, Tanjim	Ramaprasad, Jui	Wager, Stefan
Bray, Robert	Howell, Sabrina	Ramdas, Kamalini	Windschitl, Paul
Burtch, Gordon	Hu, Yu Jeffrey	Rees-Jones, Alex	Wu, Lynn
Caldentey, Rene	Huang, Peng	Riedl, Christoph	Yoganarasimhan, Hema
Chan, Jason	Huang, Yan	Ringgenberg, Matthew	Yom-Tov, Galit
Chen, Pei-yu	Hyndman, Kyle	Rossi, Alberto	Zeithammer, Robert
Chen, Xi	Iancu, Dan	Rudin, Cynthia	Zhang, Dennis
Chen, Yan	Ibrahim, Rouba	Russo, Daniel	Zhang, Xiaoyan
Chod, Jiri	Ipeirotis, Panos	Saar-Tsechansky, Maytal	Zheng, Fanyin
Choudhury, Prithwiraj	Jose, Victor	Sarkar, Sumit	Zheng, Yanchong
Ciocan, Florin	Kannan, Karthik	Savva, Nicos	Zhu, Feng
Coffman, Katherine	Karahanna, Elena	Schaefer, Andrew	Zuo, Luo
Cohen, Maxime	Kesavan, Saravanan	Seamans, Robert
Da, Zhi	Kim, Song-Hee	Shen, Max
Dai, Hengchen	Krepely Pool, Veronika	Shi, Cong