Empirical Reordering of Horn Clauses in Logic Programs; Style as Property: Protecting Creative Expression from Generative AI 92 views

In recent years, generative artificial intelligence has exploded in both popularity and capability. While generative AI excels at replicating creative work, it suffers from a lack of generalizability and a tendency towards hallucination that makes it unfit for high-risk or precise tasks. The end result is that generative AI is being used to automate away enjoyable creative work while doing little to automate high-risk or monotonous work. This is far cry from the ideal world where AI takes care of the drudgery and leaves humanity to pursue the arts, so there must be steps taken to improve the capabilities of the next generation of AI and to prevent the current generation from devaluing creative work. My technical paper tackles this problem by developing novel algorithms for optimizing neural symbolic AI, which improves generative AI with logical programs to be more generalizable and precise. My STS research explores the protections creatives have from their work and personal styles being replicated and consequently devalued by generative AI. My technical research seeks to fix some of the fundamental limitations of generative artificial intelligence. A generative AI can store vast amounts of knowledge but has limited ability to generalize outside of its training distribution and can often hallucinate faulty information. Logic programming is an older form of artificial intelligence that works by representing knowledge as logical facts and then using logical rules to make conclusions. Neuro-symbolic AI is a paradigm where knowledge is generated by generative AI and stored in a logical programming language. It combines the robustness of generative AI with the precision and generalizability of logical programming, though these generated logic programs are often severely unoptimized. Logic programs are represented with mathematical structures called Horn clauses, and the order of the conditions in a Horn clause can severely impact the performance. My technical research proposed a novel algorithm that would automatically reorder Horn clause conditions to more optimized forms. The algorithm works by rating each condition by their expected total matches and improves over previous methods by tracking the number of cyclical logical dependencies. My work further improves on previous methods by using the rating algorithm as only an initial approximation, and instead tracking the average output time for each condition empirically and improving the ordering over time. I implemented a new logical programming language to test my novel algorithm as well as various other algorithms found in literature. My technical research advances the field of neuro-symbolic AI by allowing AI generated logic programs to be optimized for real-world usage. My socio-technical paper explored how generative artificial intelligence has the capability to devalue creative work by replicating creatives’ personal styles. I made the argument that personal creative style is a vital resource that is often the only thing that gives creative work value, so creatives have an incentive to protect it. I showed how generative AI is able to replicate personal creative style and is causing economic losses to artists. I explored various protections creatives have against their work being used as training data for generative AI and discovered that there is currently no legal protection against AI being trained on copyrighted work. I explored various preventive measures including data-scraping blockers like the Robots Exclusion Protocol and opt-out options for data scraping on social media platforms. I argued that the post-training opt-out procedures offered by AI companies aren’t very effective and that people should be able to boycott AI content. I explore methods of marking AI generated content, so users are better able to identify it. The artificial intelligence of the future should be able to automate away a wide variety of tasks while also allowing human creative expression to be preserved. My technical research advanced this goal by allowing generative AI to produce optimized logic programs. My sociotechnical paper explores the various legal, social, and technological measures needed to protect human creative expression in the age of AI. Future researchers should look into further ways of optimizing generated logic programs like just-in-time compilation or the socio-technical steps required to prevent data from being scraped from people who are trying to protect it. I would like to thank Professor Caitlin Wylie for guiding me in writing my socio-technical paper and Professor Rosanne Vrugtman for helping me construct my technical paper.

BS (Bachelor of Science)

Generative AI; LLM; Logic Programming; Symbolic Logic; Data Privacy

School of Engineering and Applied Science Bachelor of Science in Computer Science Technical Advisor: Rosanne Vrugtman STS Advisor: Caitlin Wylie

English

2025-05-04