The evolution of deception

Abstract

Deception plays a critical role in the dissemination of information, and has important consequences on the functioning of cultural, market-based and democratic institutions. Deception has been widely studied within the fields of philosophy, psychology, economics and political science. Yet, we still lack an understanding of how deception emerges in a society under competitive (evolutionary) pressures. This paper begins to fill this gap by bridging evolutionary models of social good— public goods games (PGGs)—with ideas from interpersonal deception theory (Buller and Burgoon 1996 Commun. Theory 6 , 203–242. ( doi:10.1111/j.1468-2885.1996.tb00127.x )) and truth-default theory (Levine 2014 J. Lang. Soc. Psychol. 33 , 378–392. ( doi:10.1177/0261927X14535916 ); Levine 2019 Duped: truth-default theory and the social science of lying and deception . University of Alabama Press). This provides a well-founded analysis of the growth of deception in societies and the effectiveness of several approaches to reducing deception. Assuming that knowledge is a public good, we use extensive simulation studies to explore (i) how deception impacts the sharing and dissemination of knowledge in societies over time, (ii) how different types of knowledge sharing societies are affected by deception and (iii) what type of policing and regulation is needed to reduce the negative effects of deception in knowledge sharing. Our results indicate that cooperation in knowledge sharing can be re-established in systems by introducing institutions that investigate and regulate both defection and deception using a decentralized case-by-case strategy. This provides evidence for the adoption of methods for reducing the use of deception in the world around us in order to avoid a Tragedy of the Digital Commons (Greco and Floridi 2004 Ethics Inf. Technol. 6 , 73–81. ( doi:10.1007/s10676-004-2895-2 )).

Deceptive AI Ecosystems: The Case of ChatGPT

“…At the core, the origins of training data for contemporary LLMs like ChatGPT remain uncertain, particularly when the source code and training processes are not openly accessible. Determining methods to ensure that these models obtain precise information for training and learning, as well as comprehending the rationale behind model-generated responses, are essential directions for future research and development if we want to avoid a Tragedy of The Digital Commons [29] where our digital ecosystem is polluted by deceptive AI [59].…”

Limitations of Theory of Mind Defenses against Deception in Multi-Agent Systems

“…Our game setting adheres to a zero-sum structure, implying that all rational strategies, regardless of whether deception is employed, come at the expense of other players. The computational demands associated with orchestrating a deception process can be substantial, potentially exceeding the scope of this example, but we will delve into this aspect in the Discussion Section (8).…”

“…Deception within the context of multi-agent systems can be delineated as a sophisticated process in which one agent, known as the deceiver, strategically manipulates another agent, referred to as the target, inducing the latter to harbor erroneous beliefs. The ultimate objective of this manipulation is to accomplish an underlying goal, which could involve leading the target agent to draw incorrect conclusions or compelling it to undertake specific actions, thereby facilitating the attainment of the deceiver's desired outcome ( [8], [9], [10], [11]). This deceptive process is often predominantly grounded in Boolean assignments.…”

Interoperable AI: Evolutionary Race Toward Sustainable Knowledge Sharing

“…To test Smith's intuition, we do not engineer a specific MAS. Instead, we derive our approach from the evolutionary agent-based modelling of cooperation, communication, and knowledge-sharing in complex social systems as the ones in [6,7,9]. None of these previous studies have looked at the question of shared language vs. translation strategies in the context of ensuring sustainable communication.…”

“…The first set (prefixed by "KSG") explores the direct competition between 2 agents that share knowledge between themselves in a private way. The second set (prefixed by "PKSG") explores the idea of knowledge exchanged publicly.We have set the number of agents that interact in these games M = 5 following the study presented in [7], which uses similar evolutionary agent-based modelling techniques to ours, based on [6,8]. If we consider that knowledge is shared publicly (e.g., on the Web) in domains where agents interact in groups, a first reasonable setup is to adopt the same value for number of agents engaged in these interactions.…”