ABSTRACT:
The Special Section opening this issue of Journal of Management Information Systems (JMIS) is devoted to the works addressing various aspects of cognitive reapportionment occasioned by the recent dramatic progress in artificial intelligence (AI), supported by the advances in computing such as cloud computing and the GPU/TPU chips. The Special Section is guest-edited by Benn R. Konsynski, Abhishek Kathuria, and Prasanna P. Karhade. The first guest editor is a codeveloper of the concept of cognitive reapportionment over three decades ago [6], and his prescience is now in play as we study and aim to support the emerging developments in AI.
Cognitive reapportionment is an idea developed to study and apply the apportionment of tasks between the humans and the machines as the latter become capable of carrying out cognitive tasks, such as language understanding, generating the summaries of large corpora, and a degree of multimedia creativity, among other attainments previously ascribed exclusively to the humans. To paraphrase a question from the political realm: “Who/what should do what?”.
Cognitive reapportionment will be challenging, as the highly impressive AI systems have significant limitations at present, and the assessment of their capabilities will be necessary as they rapidly evolve to realize their fuller potential in the future. Cognitive misperception, gently anthropomorphized as hallucinations, may have serious negative effects. In testing Copilot and similar chatbots in the textual mode, I found them highly fluent at summarization and presenting alternatives (for example, prompt for “cognitive reapportionment” and you will see the summarization and illustration of the concept). I also found them not yet satisfactory in properly assigning weights to the evidence on more complex tasks—which is the basis of higher-level decision making. A variety of cognitive biases may ensue. Undoubtedly, there will be path-breaking progress here, while independent decision-making by AI agents is not on the cards at present for more responsible tasks.
While growing exponentially in their capabilities, the AI systems based on large language models (LLM) are already impressive as the partners of humans. They are particularly effective in handling well-structured tasks rooted in pattern recognition. For example, in a study comparing the performance of a neuroradiologist of more than 30-year experience with a specialized Glass AI, the radiologist’s results in recommending the imaging modality were scored at 2.20 and the LLM’s at 1.75, both out of 3 [7]. Considering that most patients do not have access to a neuroradiologist with that experience, there is room for human-AI collaboration in most radiological practices. The efficiency of the LLM use may be demonstrated by their impact on the software-code generation. Thus, Copilot’s use in the GitHub environment shows the acceptance rate by programmers of the code suggestions by the AI chatbot to be in the general range between 20-26% in the developers’ assessments, with the resulting productivity increase, as it also shows the need for collaboration with the human developers [8]. Already now, the design capabilities of the AI display a high degree of creativity in the multimedia environments [5]. Organizational deployment of AI has become a direct concern of the C-suites and in some cases of the chief executive officers [3].
The aforementioned result indicates that presently LLMs can act, if used properly, as partners of humans. Indeed, a study by Dennis, Lakhiwal, and Sachdeva has shown empirically that AI agents are largely acceptable to humans in organizations as partners in teams [4]. Notably, the emergence of such systems leads to the revaluation of the existing values of human endeavors, as shown by Brynjolfsson, Liu, and Westerman [2]. In the more philosophical notions of a longer look, human-AI collaboration is seen as one of several potential paths to “superintelligence” [1].
Our field is well positioned to lead the research on the human-AI collaboration, and specifically on cognitive apportionment, with our collective expertise both in information technology (IT) and in the behavioral disciplines. JMIS has published a number of research papers on this collaboration and they are discussed here in the guest editors’ introductory paper, along with the works they have included in their Special Section. It is important for the Information Systems discipline—and we hope for the world at large—that we pursue this research.
The first paper in the general section of the issue continues the theme of the Special Section. Recognizing the deficiencies of the LLM in their training through deep learning, the authors aim to build on the human learning processes in order to develop systems that bring into action the reservoirs of knowledge learned by humans over their lifetimes. The authors, Long Xia, Wenqi Shen, Weiguo Fan, and G. Alan Wang, present a novel design of such a knowledge-aware model. In the vein of design-science research, they deploy schema theory and employ the sequence of human cognition to produce a learning framework that is by orders of magnitude more economical than LLM in the use of computing resources, while matching the performance of those models on certain tasks. The authors demonstrate the effectiveness of their design framework in text analytics. The approach is promising as an add-on to deep learning in better matching the cognition of human collaborators and in the potential transparency of the task handling.
Crowdsourcing has become a potent source of ideas and ready solutions coming from a broad public or from a carefully selected set of ideators. The participants create value for the sponsoring platform or organization. What is their incentive to do so? It may be simply a cash reward; it also may be a non-cash reward such as winning the crowdsourcing contest or the satisfaction of offering a solution to a societal ill. The current state of research differentiating between the intrinsic and extrinsic motivations is replete with contradictions, and the present situated work is of importance. The authors of the next paper, Christoph Riedl, Johann Füller, Katja Hutter, and Gerard J. Tellis, argue that instead of offering a single type of reward, the crowdsourcing sponsors should offer a menu of awards in order to match the participants’ motivations. In a number of empirical studies presented in their work, the authors demonstrate the effectiveness of the heterogeneity of rewards in eliciting truly innovative ideas from the “crowd.”
Two subsequent papers address nodal issues in the design and use of social media. Ofir Turel and Hamed Qahri-Saremi empirically study the effects of “likes” (which may now be masked in some media) versus the more recently introduced “dislikes” on user behaviors, with the aim of preventing the toxic variety. Social comparisons have been recognized as a source of such toxicity. Here, the authors also introduce the construct of internal comparisons, affecting the individuals’ homeostasis (dynamic psychological balance an individual is aiming to maintain) as another such source. The empirics are driven by established psychological theories and show the deleterious effects of “dislikes.” In complex systems, seemingly small changes may have large effects; social media are systems involving billions of people who interact with and react to each other in the environment created to increase their engagement. The work breaks new ground and is societally, as well as socially, important as we seek to eradicate the phenomena on the dark side of social media.
Mass participation in social media makes them a potent means to detect various individual and societal ills, at all times preserving the privacy of the individuals. It has been recognized that these systems are a fit means for detecting depression. The authors of the next paper, Wenli Zhang, Jiaheng Xie, Zhu (Drew) Zhang, and Xiang Liu present a novel method for doing so. Echoing the work of Xia et al. that appears in this issue as well, the researchers use a knowledge-aware approach, incorporating the knowledge in the medical subdomain combined with deep learning acting on the traces left on the media by the users. The design-science paper demonstrates the effectiveness of this approach. Early detection of depression, an illness affecting a vast number of people often without their knowledge is of a great moment in the societies undergoing rapid change—and this work is highly salient.
The concluding paper of the issue focuses on the effects of outsourcing on the high-tech firms, the main actors in the IT arena. More specifically, Yi Dong, Nan Hu, Yonghua Ji, Chenkai Ni, and Jing Xie study empirically the impact of government contracting on the long-term health of the firms. The authors find that these contracts have a weighty influence on the market valuation of the companies in possession of them, and on the assessment of the quality of their earnings by investors. This heightened valuation leads in turn to a greater stability of such firms as compared to their competitors, and thus to the advantages of so contracted firms by their inclusion in further supply webs. The positive spiral is important to the stability of the IT sector, deleterious as a market entrenchment can be to innovation. Balance is the thing, with the government contracting agencies desirably conducting ongoing reviews of both contracted and as yet uncontracted firms.