ABSTRACT:
Recommender systems have become a significant component of the online economy and consequently of the Information Systems (IS) research. They underpin the operation of the platformized marketplaces, while markets—since the deep antiquity—are the successful form of economic organization. Mixed economies that include markets rely on them as the driving and productive force. The matching of sellers with buyers is at the heart of the marketplaces, and recommenders facilitate this matching. The availability of the massive data on the buyers’ preferences and the sellers’ offerings has, over time, increased the recommendations opportunities—and difficulties. Further new opportunities have emerged with the recent use of generative artificial intelligence (AI) models that are trained on the general data and the proprietary data of the platform. Recommenders lower the friction in the markets and increase consumer surplus by the ability to expand the breadth of the offerings and to increase the aptness of the match. As they generally operate in a black-box mode, they also offer an opportunity for market manipulation, notably by the platform owner gaining preference in the competition with other platform sellers.
The general category of recommenders includes the algorithms that operate on social media platforms. These media recommenders (MR), as I will call them, ought to pursue different objectives than those of the marketplace recommenders. Their aim is—or should be—to recommend to the participants of the social media platforms, such as YouTube or Instagram, the content the consumers would be interested in, based on their constructed profile and the nature of the content item. An unbiased recommendation of other content counteracts balkanization and serves the users well. Here, we come to realize that the social media platforms include at least one more key stakeholder, that is, the advertisers or similar underwriters of the platform’s existence. The currency circulating here is users’ engagement. This construct wraps in their time, attention, and possible participation (at least as the raters of the content). This changes the role of the non-transparent MR to the caterers to the users’—yes, interests—but as skewed by their emotional baggage, generally known to the platforms from the accumulated data. The high level of emotional arousal leads to higher engagement. This has consequences.
Two papers that open this issue of Journal of Management Information Systems (JMIS) investigate the dark side of the MRs that stems from the engagement-seeking by social media platforms. AI has become a means of recommendation. Many stories on social media are recommended by AI agents, without the disclosure of this fact. There exists an expectation (and regulatory recommendations in some countries) that labeling stories as AI-recommended would stem the propagation of fake news on the premise that the users would devote more time to comprehending such stories and assessing their credibility. The authors of the opening paper of the issue, Hanzhou Ma, Alexander S. Dennis, Alan R. Dennis, and Wei Huang, base themselves in the social influence theory to conduct a series of experiments to verify or falsify these assertions. The authors’ findings show that when reading a story known to be recommended by AI the users tend to rely on the fast cognition known as System 1 (approximately, a fast judgment) rather than on the ratiocination of the more deliberative System 2 of cognition. In other words, the labeling of stories as AI-recommended would lead to the increased spreading of fake news. We have an important result as we face the corrosion of the public sphere by the spreading and amplification of untruths on the globally used media.
Societal polarization has become a fundamental issue in our sociopolitical lives. The role of social media in the growth of this phenomenon has been widely debated at all levels of our society. A transparent mechanism design for MRs would alleviate the concerns. The authors of the next paper, Carlos Carrasco-Farré, Didier Grimaldi, Marc Torrens, and Enzo Longobuco, make a highly significant contribution to our understanding of the role of social media. Using social identity theory as their lens, the researchers show empirically, by garnering a large data corpus of YouTube videos and recommendation links, how this polarization occurs on the level of social media. The authors study the role of MRs in acrophily, the tendency of the media users to expose themselves online to the contributions of others who hold more extreme political views than themselves. The work studies empirically both the ingroup acrophily (the preference for the extreme content aligned with the view of the focal individual) and the outgroup acrophily (favoring the extreme content of the group holding the opposing view). The latter—seemingly paradoxical—construct is introduced by the present authors and will remain valuable in the study of social media. The authors go to the psychological root causes of both types of acrophily and show empirically how users’ biases are amplified by the MRs. The authors offer pragmatic suggestions for the MR design that would lead to the remediation of the acrophylic effects. Here we have an expansive research study that analyzes some the roots of our social and societal polarizations.
We are witnessing an ever-deeper infusion of AI in our organizational functioning—and in our thinking and research. Three subsequent papers focus on the widely ranging aspects of this infusion. Along with the amplification of polarization and fake news, hate speech is another deleterious phenomenon afflicting social media. Here, Kyuhan Lee and Sudha Ram present a theoretically grounded method of deploying large language models to detect hate speech. This detection is generally a thornier problem than the detection of fake news, owing to a more nuanced (if this is the word in the sordid context) differentiation of the hate elocutions from the more ordinary ones. The authors introduce and test a multi-agentic framework where the agents generate prompts, adding contextual details to the suspected hate-speech elocutions. Beyond that, the framework includes motivational details in order to sharpen the classification as hate speech. The authors demonstrate significant performance advantages of their framework.
The AI-human collaboration is clearly the nodal issue of the present and the future. Our discipline has already contributed extensive knowledge to the domain, with numerous highly contributing papers published in JMIS and the other top-tier journals. In the next paper of the issue, Yotam Liel and Lior Zalmanson empirically study the overreliance on AI from a new angle, that of normative pressure. Human collaborators may feel that the AI systems are so valorized in the common organizational environment as to merit following their advice regardless of the—perhaps—better judgment. We can clearly see here the connection between these findings and those of Ma et al. in a preceding work in this issue. However, the cognitive mechanism revealed here is different, namely a discomfort of disagreeing with the AI as a perceived authority. This work showcases another aspect of the need to design a human-in-the-loop system with extreme care. The human who is in the loop is subject to the human biases and, in the future, perhaps deliberate deception by AI-based algorithms.
The design of innovative systems for weighty classes of problems is (or should be) an important component of the development of our IS discipline. Here, Reza Ebrahimi, James Lee Hu, Ning Zhang, Jay F. Nunamaker, Jr., and Hsinchun Chen present a design-oriented methodology for malware detection as a part of our disciplinary research programs addressing IS security. Adversarial attacks, which are conducted by malware with ongoing perturbations, are difficult to detect and the current methods of doing so are deficient. Ebrahimi et al. deploy a variant of recurrent independent neural networks to devise a robust generative AI model without unrealistic knowledge requirements. The detectors generate proactively the variants of the potential future malware to defeat them and can serve in the defense against an AI-enabled adversary. The authors demonstrate the advantages of their method with a set of real data and a case study.
Cybersecurity is also the focus of the next paper, by Eunae Yoo, Christopher W. Craighead, and Sagar Samtani. The authors are researching the vulnerabilities of the software supply chains, where the software systems deployed encompass vulnerable software from multiple sources. This is the reality of the current software environments. The research questions here center on the influence of the structural dependencies of the software systems. The two dependencies crucial to the propagation of security vulnerabilities are the complexity of the dependency network of software components and the strength of the interdependence (coupling tightness) among these components. Based on the empirics, the authors demonstrate the role of both in the security vulnerabilities of the entire software system. Hardening software supply chains is the ultimate objective.
The blockchain has become in a relatively short time since its invention, based on a highly ingenious amalgamation and enhancement of numerous cryptographic ideas and data-management concepts accumulated by computer science, the fundamental infrastructure of token-based economy. It has further potential. Haozhao Zhang, Zhe (James) Zhang, Zhiqiang (Eric) Zheng, and Varghese S. Jacob present and exercise an innovative generative blockchain. It goes beyond validating and recording transactions to generating them prior to these operations. Notably, transaction generation may be in some case categories a computationally-intensive operation—think again about matching a buyer and a seller. The authors demonstrate that the amalgamated transact-and-record blockchain could lead to efficiencies of workflow and to decentralization, preventing the bottlenecks of a single-point of processing (and failure). The generative blockchain proposed here comes with its own proof-of-merit mechanism for prioritizing the transactions. This paper constitutes another demonstration that our field can generate innovative IS designs. It is to be hoped that this work will be built on.
Omni-channel commerce, encompassing both the online and physical selling channels, is today the face of activities of major retailers. The natural question emerges: What products should be sold through what channel and in what quantities? This is known as assortment planning and is the subject of the paper by Amar Sapra and Subodha Kumar. The authors offer a formal model that takes into account product popularity (commonly-purchased or niche), product prices, and demand-fulfillment level (probability of no-stockout). The model answers research questions most significant to omni-channel retailing. It has been said that all models are wrong, but some are useful. The model presented here is, as all economic models are, an approximation. However, it makes reasonable assumptions and is highly useful.
We have seen unintended consequences identified by Ma et al. in the opening paper of the issue. We can see them again in the work of Antino Kim and Che-Wei Liu. Some online selling platforms introduced markers to identify minority-owned businesses in order to help them compete with others on the platform. The authors investigate the outcomes of this apparently worthy initiative. They are only partly encouraging. The potential buyers who are motivated to patronize such businesses, did so. However, the more common effect was the backfiring of the effort. The work relies on and contributes to the cue-utilization theory. The authors’ analysis of the potential causes and remediations contributes to out striving for a society of inclusive well-being.
In the concluding paper of the issue, Mario Schaarschmidt, Sven Heidenreich, and Matthias Bertram investigate user behavior in the post-adoption stage of digital services. Such services are expected to bring in a continuous stream of revenue, a highly desirable effect. In a well-known to all of us attempt to innovate these services also keep changing their offerings. Some of us resist that to the point of abandoning the service. Innovation resistance theory has been tested to study the outcomes. Our present authors broaden this theory to inform their empirically-based research that goes beyond what is known to the discovery how passive innovation resistance can be related to the users’ loyalty posture. The work expands the theory of resistance and shows the ways to user retention. As all the papers in the issue, this work contributes to our theoretical knowledge while informing the practical side of Information Systems.