INTRODUCTION
Concepts and assumptions derived from observations and laboratories have restricted social science, machine learning (ML) and belief logics to operate in controlled and stable contexts; even still, however, by using hand-picked observers who were also well-trained in the laboratory to make predicti- ons for the outcomes of social events occurring in open contexts, Tetlock
© 2023. Published by AHFE Open Access. All rights reserved. 148
Autonomous Human Machine Teams: Data Dependency 149
and Gardner’s (2015) “superforecasters” failed to predict Brexit or Trump in 2016 (Lawless et al., 2019); similarly, Western Russian expert observers (e.g., Jones & Wasielewski, 2022), along with Russian military planners (Risen & Klippenstein, 2022), misjudged that Russia’s army would easily defeat Ukraine.
In contrast, the National Academy of Sciences (Endsley et al., 2021) has warned that controlled contexts are insufficient to produce operational auto- nomous systems, in particluar, for human-machine teams. We agree; by studying real-world contexts, we have concluded that interdependence is the missing ingredient necessary to achieve autonomy. Interdependence explains the Academy’s second claim that the “performance of a team is not decom- posable to, or an aggregation of, individual performances” (Endsley et al., 2021, p. 11). We explain the Academy’s second claim about disaggregation by noting that the data dependency (Davies, 2020) among teammates is reflected by the reduction in the degrees of freedom among them and the internal infor- mation needed to transmit and share each agent’s status (e.g., a simple grunt is sufficient to alert a teammate; from Sliwa, 2021). With significan- tly less information needed than independent individuals can provide, these two reductions account for the failure of belief logics facing uncertainty or conflict in the field (e.g., Mann, 2018), yet recommendations to minimize interdependence persist (with experiments, see Kenny et al., 1998; for organi- zations, see Conant, 1976), even though the literature indicates that humans live their lives in states of interdependence (Jones, 1998), even though the source of innovation is the interdependence between culture and technology (de León et al., 2021, and even though the best science teams have been found to be highly interdependent (Cummings, 2015).
But, in addition, social science has had great difficulty replicating concepts based solely on observations of individuals alone (Nosek, 2015); e.g., the concept of implicit racial attitudes has been found to be invalid (Blanton et al., 2009); worse, explicit treatment in an attempt to change implicit racial attitudes appears to have had negligible effects (reviewed in Singal, 2023). This situation has led us to argue that users of the traditional models of teams (e.g., Cooke & Hilton, 2015) should overturn these assumptions because the independent data collected by observations alone, especially based on laboratory methods and the additional assumption that the cognitive model subsumes behavior (Thagard, 2019), cannot recreate whatever social event is being observed (viz., Shannon’s, 1948, information theory and i.i.d. data preclude social interdependence; for a fuller discussion of these issues, see Schölkopf et al., 2021).
Mutual dependency is an interdisciplinary concept: In quantum mechanics, a knowledge of the whole precludes a knowledge of the parts (Schrödinger, 1935), a concept also fundamental to social psychology (i.e., the concept of the whole being larger than the sum of its parts was freely used by the founder of the discipline of Social Psychology, Lewin, 1951), and systems enginee- ring also used the concept of the whole and its parts (Walden et al., 2015); however, System Engineers have begun to refer to this phenomenon as “emer- gence.” But unlike Shannon (1948) information (for its use in organizations,
150 Lawless
see Conant, 1976), dependency in data poses a formidable challenge theore- tically, mathematically, for engineering and information fusion systems, and the associated disciplines that depend on intuition, interpretation and mea- ning (e.g., for a review, see Speaks, 2021); viz., philosophy, economics and social science.
Despite these challenges, we have succeeded by treating cognition inter- dependently with behavior in open contexts to find that minimum team structural entropy allows a team to produce maximum team performance (Lawless, 2022a, b, c, d); and by treating beliefs as imaginary, we have rediscovered the value of debate to reduce the uncertainty and conflict that autonomous systems must be able to confront in open contexts. Regarding debate, from the U.S. Supreme Court (1970): “cross-examination [in the courtroom is] the ‘greatest legal engine ever invented for the discovery of truth”’; regarding the value of appeals, from Justice Ginsburg (2011, p. 3), competing views provide an “informed assessment of competing interests”; and, in the aftermath of a tragic drone strike in Afghanistan in August 2021, from the Department of Defense (2021), “red-teaming” challenges the decisions of humans when operating in a state of heightened emotion while they are attacking targets on the ground with drone machines in combat.
Our model for autonomous human-machine teams leads us to expect that an AI machine operating interdependently with a human as a team- mate, jointly challenging each other’s beliefs about reality while shaping shared experiences, has a better chance to operate autonomously in open contexts. Our model exploits interdependence by requiring that teams engage in tradeoffs for agent fittedness. Surprisingly, by adding boundaries within which uncertainty and conflict could be minimized during operations allows logic to return in part, justifying Simon’s (1989) bounded rationality (e.g., roundabouts with traffic; robotic surgery; no fly zones in combat; the context dependency of machine learning; and bounding uncertainty in the courtroom).
To apply Simon (1989) to the courtroom, we note that to reduce the uncertainty associated with circumstantial evidence in a criminal case, a courtroom often has two opposing officers of the court (a prosecutor and a defense attorney) face off before a judge and a jury to determine the inno- cence of an individual charged with a crime. The environmental uncertainty involved is further narrowed by rulings from the judge and the judge’s instru- ctions to the jury. Consequently, facing uncertainty, we conclude that debate is the primary means to ground truth (for an application of these ideas to Artificial Intelligence (AI), see Cooke & Lawless, 2021; Lawless et al., 2019; NSC, 2021; Sofge et al., 2019; and more recently, see Lawless et al., 2023).
DEVELOPING AN EQUATION TO ACCOUNT FOR THE TRADEOFFS BETWEEN THE STRUCTURE AND PERFORMANCE OF A TEAM Briefly, we model with our equation between the uncertainty in a team’s structure, 1SEP, and its maximum entropy performance, 1MEP:
Autonomous Human Machine Teams: Data Dependency 151
1SEP ∗1MEP ≥C (1) Thus, as the uncertainty in a team’s structure reduces to perfect cohere- nce in the limit, its performance is allowed to become a maximum. With Equation (1), we have modeled uncertainty and conflict (where logic fails; Mann, 2018); deception; blue-red team challenges; emotion (higher emotion reduces a team’s options); vulnerability; mergers (the random effects of fitte- dness); and innovation. In addition, Equation (1) indicates that uncertainty in a team’s structure depends on how the members of a team fit together, not whether they believe they have a shared model or not, supporting the finding by the National Academies of Science that disaggregation is unable to assign individual contributions to the individual members of a team by observing a team’s performance (Endsley, 2021, p. 11).
There is always a danger of autonomous teams being overseen by auth- oritarian regimes. But, what we have found is that authoritarians first shut down free speech, which requires them to stop or interfere with interdepen- dence. But impeding, interferring or stopping interdependence impedes or slows innovation, requiring an authoritarian regime to steal in order to be able to innovate (Lawless, 2022d). Why do republics with strong checks and balances like the United States or Israel often lead in innovation (Lawless, 2022c)? We have found that strong republics promote the noise arising from “the circulation and mixture” of opposing ideas (Puchner, 2023), characteri- zing the sources of debate and conflict among ideas and concepts associated with free speech and innovation..
CONCLUSION
In this brief review, we have sought to highlight that traditional social science offers little help to produce the mathematical models of autonomy that will be needed to design and operate human-machine teams and systems. In con- trast, by building upon the finding of disaggregation cited by the National Academies of Science (Endsley, 2021, p. 11), we have succeeded by exploiting the effects of interdependence under uncertainty, not the internal uncertainty in a model, but the uncertainty that is a significant part of open environments, and after much effort (Lawless, 2022a, b, c. d), we have begun to make signi- ficant strides in producing a model that works in the open (Lawless et al., 2023), away from the laboratory. In concluion, the results we have reviewed herein support the idea of a new discipline of data dependency.
ACKNOWLEDGMENT
The author would like to acknowledge and offer his thanks for the many years of financial and intellectual assistance that he has received from Ranjeev Mittu, the Branch Head for the Information and Decision Sciences Branch within the Information Technology Division at the U.S. Naval Research Laboratory, Washington, DC. The author also thanks his collea- gue of many years, Don Sofge, a Computer Scientist and Roboticist who leads the Distributed Autonomous Systems Group in the Navy Center for Applied Research in Artificial Intelligence (NCARAI), U.S. Naval Research
152 Lawless
Laboratory, Washington, DC. And the author would like to thank Kimberly L.Y. Butler for her many years of discussion and support.
REFERENCES
Blanton, Hart, Klick, J., Mitchell, G., Jaccard, J., Mellers, B. & Tetlock, P. E. (2009), Strong Claims and Weak Evidence: Reassessing the Predictive Validity of the IAT, Journal of Applied Psychology, 94(3): 567–582. Conant, R. C. (1976). “Laws of information which govern systems.” IEEE Transa-
ction on Systems, Man, and Cybernetics 6: 240–255. Cooke, N. & Lawless, W. F. (2021), Cooke, N. J. & Lawless, W. F. (2021, forthcoming), Effective Human-Artificial Intelligence Teaming, In Lawless et al., Engineering Science and Artificial Intelligence, Springer. Cooke, N. J. & Hilton, M. L. (Eds.) (2015), Enhancing the Effectiveness of Team
Science. Authors: Committee on the Science of Team Science; Board on Behavioral, Cognitive, and Sensory Sciences; Division of Behavioral and Social Sciences and Education; National Research Council. Washington (DC): National Academies Press. Cummings, J. (2015). Team Science Successes and Challenges. National Science Foundation Sponsored Workshop on Fundamentals of Team Sci- ence and the Science of Team Science (June 2), Bethesda MD ({https: //www.ohsu.edu/xd/education/schools/school-of-medicine/departments/clini cal-departments/radiation-medicine/upload/12-_cummings_talk.pdf}). Davies, P. Does new physics lurk inside living matter? Physics today 2020, 73, 34–41.
doi: 10.1063/PT.3.4546. de León, M. S. P. et al. (2021), The primitive brain of early homo. Science, 372: 165–171. doi: 10.1126/science.aaz0032. DoD (2021, 11/3), Pentagon Press Secretary John F. Kirby and Air Force Lt. Gen. Sami D. Said Hold a Press Briefing. Retrieved 11/3/2021 from https://www.defense.gov/News/Transcripts/Transcript/Article/2832634/pe ntagon-press-secretary-john-f-kirby-and-air-force-lt-gen-sami-d-said-hold-a-p/. Endsley, M. R. et al. (2021). Human-AI Teaming: State-of-the-Art and Research
Needs. The National Academies of Sciences-Engineering-Medicine. Washington, DC: National Academies Press. Retrieved 12/27/2021 from https://www.nap.edu/ catalog/26355/human-ai-teaming-state-of-the-art-and-research-needs. Ginsburg, R. B. (2011), American Electric Power Co., Inc., et al. v. Connecticut et al., 10–174, http://www.supremecourt.gov/opinions/10pdf/10-174.pdf (Accessed 11 May 2017). Jones, E. E. (1998), Major developments in five decades of social psychology, In
Gilbert, D. T., Fiske, S. T., & Lindzey, G., The Handbook of Social Psychology, Vol. I, pp. 3–57. Boston: McGraw-Hill. Jones, S. G. & Wasielewski, P. G. (2022, 1/13), “Russia’s Possible Invasion of Ukraine,” Center for Strategic & International Studies, retrieved 1/28/2023 from https://www.csis.org/analysis/russias-possible-invasion-ukraine. Kenny, D. A., Kashy, D. A., & Bolger, N. (1998). Data analyses in social psychology.
Handbook of Social Psychology. D. T. Gilbert, Fiske, S. T. & Lindzey, G.. Boston, MA, McGraw-Hill. 4th Ed., 1: 233–65. Lawless, W. F. (2022a), Risk Determination versus Risk Perception: A New Model of Reality for Human–Machine Autonomy, Informatics 2022, 9(2), 30; https://doi.org/10.3390/informatics9020030.
Autonomous Human Machine Teams: Data Dependency 153
Lawless, W. F. (2022b), Toward a physics of interdependence for autonomous human-machine systems: The case of the Uber fatal accident, 2018, Frontiers in Physics. Section Interdisciplinary Physics, https://doi.org/10.3389/fphy.2022. 879171. Lawless, W. F. (2022c), Interdependent Autonomous Human-Machine Systems: The
Complementarity of Fitness, Vulnerability & Evolution, Entropy. Lawless, W. F. (2022d). Autonomous human-machine teams: Reality defeats belief logics, but not the dualism in observational data. Invited paper. Lawless, W. F., Sofge; Donald A., Lofaro, Daniel, & Mittu, Ranjeev (2023), Editorial:
Interdisciplinary Approaches to the Structure and Performance of Interdependent Autonomous Human Machine Teams and Systems, Frontiers in Physics, eBook. Lawless, W. F.; Mittu, R.; Sofge, D.; Hiatt, L. Artificial intelligence, autonomy, and human-machine teams: Interdependence, context, and explainable ai. AI Magazine 2019, 40, 5–13. doi: 10.1609/ aimag.v40i3.2866. Lewin, K. (1951), Field theory of social science. Selected theoretical papers. Darwin
Cartwright (Ed.). New York: Harper & Brothers. Mann, R. P. (2018), Collective decision making by rational individuals, PNAS, 115(44): E10387-E10396; from https://doi.org/10.1073/pnas.1811964115. Nosek, B., corresponding author from OCS (2015), Estimating the reproducibility
of psychological science, Science, 349 (6251): 943: also, https://doi.org/10.17226/ 25303.) NSC (2021), National Security Commission Report on AI; https://www.nscai.gov/ 2021-final-report/. Puchner, M. (2023). Culture: The Story of Us, From Cave Art to K-Pop. W. W. Norton
& Company. Risen, J. & Klippenstein, D. (2022, 10/5), “The CIA thought Putin would quickly conquer Ukraine. Why did they get it so wrong? High-tech surveillance may have blinded the U. S. to how corruption has weakened the Russian military,” The Intercept, retrieved 10/8/2022 from https://theintercept.com/2022/10/05/russia-u kraine-putin-cia/. Schölkopf, B. et al. (2021), Towards Causal Representation Learning, arXiv, retrieved
7/6/2021 from https://arxiv.org/pdf/2102.11107.pdf. Schrödinger, E., 1935. “Discussion of Probability Relations Between Separated Systems,” Proceedings of the Cambridge Philosophical Society, 31: 555–563; 32 (1936): 446–451. Shannon, C. E. (1948), A Mathematical Theory of Communication, The Bell System
Technical Journal, 27: 379–423, 623–656. Simon, H. A. (1989, 9/23), Bounded rationality and organizational learning, Technical Report AIP 107, CMU, Pittsburgh, PA. Singal, J. (2023, 1/17), “What if Diversity Training Is Doing More Harm Than
Good?” New York Times, retrieved 1/29/2023 from https://www.nytimes.com/ 2023/01/17/opinion/dei-trainings-effective.html. Sliwa. J. (2021), Toward collective animal neuroscience. Science, 374 (6566): 397–398. 10.1126/science.abm3060. Sofge, D. A., Mittu, R. & Lawless, W. F. (2019), Will a Self-Authorizing AI-Based
System Take Control from a Human Operator? From https://ojs.aaai.org/index.p hp/aimagazine/article/view/5196. Speaks, J. (2021), “Theories of Meaning”, Stanford Encycl. Philosophy, Edward N.
Zalta (ed.), https://plato.stanford.edu/archives/spr2021/entries/meaning/. Tetlock, P. E. & Gardner, D. (2015), Superforecasting: The Art and Science of Prediction, Crown; also, see Tetlock’s and Gardiner’s website at http://goodjudgment .com/superforecasting/index.php/2016/11/03/is-donald-trump-mr-brexit/ Retrie- ved 12/15/2016.
154 Lawless
Thagard, P. (2019), Cognitive Science, in E. N. Zalta (ed.), The Stanford Encyclopedia of Philosophy, retrieved 11/15/2020 from https://plato.stanford.edu/archives/sp r2019/entries/cognitive-science. U. S. Supreme Court (1970); cite: U. S. Supreme Court, California v. Green, 399
U. S. 149 (1970), California v. Green, No. 387, Argued April 20, 1970, Decided June 23, 1970, 399 U. S. 149. Walden, D. D., Roedler, G. J., Forsberg, K. J., Hamelin, R. D. & Shortell, T. M.
(Eds.) (2015), Systems Engineering Handbook. A guide for system life cycle pro- cesses and activities (4th Edition). Prepared by International Council on System Engineering (INCOSE-TP-2003-002-04. John Wiley.