Banal Account of a Safety-Creativity Tradeof in Generative AI

Abstract

Safety is banal. YouChat and other generative artificial intelligence (AI) models may be used in a variety of tasks, some mundane and some creative. Their safety may be of concern. CEUR Safety is a constraint on artifacts. Like other constraints, safety makes the feasible region under the qualitynovelty tradeof curve smaller and creativity more difi-

Keywords

• computational creativity
• generative model
• safety
• information geometry

1. Introduction

2. Safety

Safety is defined in terms of harm, aleatoric uncertainty, and epistemic uncertainty [ 1 ]. Safe AI systems constrain the probability of expected harms and the possibility of unexpected harms [ 2 ]. Harms from generative AI may be representational, allocative, quality-of-service, interpersonal, or societal [ 3 ].

3. Creativity

Creativity is the generation of an artifact that is highquality and novel [ 4 ]. Quality metrics are specific to the application. Novelty is a more application-agnostic concept that may be measured using Bayesian surprise, the relative entropy between the empirical distribution of an inspiration set and that set updated with the new artifact [ 5 ]. An inspiration set is a collection of previous artifacts in the creative domain.

Creativity by modern generative AI is implicitly or explicitly combinatorial. It generates unfamiliar combinations of familiar ideas [ 6 ]. Combinatorial creativity has precise information-theoretic limits on the tradeof LGOBE http://www.varshney.csl.illinois.edu (L. R. Varshney)

0000-0002-7376-5536 (K. R. Varshney); 0000-0003-2798-5308 (L. R. Varshney)

5. Implications

Some applications of generative AI, like autonomously writing boilerplate, require safety whereas others, like inspiring a human poet, do not. Some applications of generative AI, like writing poetry, require creativity and others, like writing boilerplate do not. Applications requiring safety tend to also be ones not requiring creativity. Applications not requiring safety tend to also be ones requiring creativity.

6. Conclusion

Information theory tells us that most natural applications of combinatorial creativity with modern generative AI are feasible in terms of the safety-creativity tradeof. Future work requires constructive algorithms for placing safety constraints on generative AI. The end.

References

[1] N. Möller , The concepts of risk and safety , in: Handbook of Risk Theory , Springer, Dordrecht, Netherlands, 2012 , pp. 55 - 85 .

[2] K. R. Varshney , H. Alemzadeh , On the safety of machine learning: Cyber-physical systems, decision sciences, and data products, Big Data 5 ( 2017 ) 246 - 255 .

[3] R. Shelby , S. Rismani , K. Henne , A. Moon , N. Rostamzadeh , P. Nicholas , N. Yilla , J. Gallegos , A. Smart , E. Garcia, G. Virk, Sociotechnical harms: Scoping a taxonomy for harm reduction , arXiv:2210.05791 , 2022 .

[4] L. R. Varshney , F. Pinel , K. R. Varshney , D. Bhattacharjya , A. Schörgendorfer , Y. -M. Chee , A big data approach to computational creativity: The curious case of Chef Watson , IBM Journal of Research and Development 63 ( 2019 ) 7 .

[5] L. Itti , P. Baldi , Bayesian surprise attracts human attention , in: Advances in Neural Information Processing Systems , 2005 .

[6] P. Das , L. R. Varshney , Explaining artificial intelligence generation and creativity , IEEE Signal Processing Magazine 39 ( 2022 ) 85 - 95 .

[7] L. R. Varshney , Mathematical limit theorems for computational creativity , IBM Journal of Research and Development 63 ( 2019 ) 2 .

[8] L. R. Varshney , Limits theorems for creativity with intentionality , in: Proceedings of the International Conference on Computational Creativity , 2020 , pp. 390 - 393 .

[9] R. K. Sawyer , Explaining Creativity: The Science of Human Innovation , Oxford University Press, New York, NY, USA, 2012 .