Assistive AI for Augmenting Human Decision-making

Natabara Máté Gyöngyössy, Bernát Török, Csilla Farkas, Laura Lucaj, Attila Menyhárd, Krisztina Menyhard-Balázs, András Simonyi, Patrick van der Smagt, Zsolt Zodi, András LorinczarXiv 2024 · DOI · arXiv

Introduction: The Regulatory Lag

As artificial intelligence continues to evolve at an exponential rate, traditional regulatory frameworks are struggling to keep pace. Current efforts, such as the EU AI Act, are inherently reactive, often addressing threats like deepfakes, automated phishing, and misinformation only after they have proliferated. This paper proposes a paradigm shift: rather than viewing AI solely as an object to be regulated, we should leverage it as an instrument of control to verify the reality of the communications we consume.

The Shift from Regulation to Assistance

The core challenge in the current digital landscape is how individuals and communities can obtain reliable information to make decisions aligned with their values and the law. The authors argue for a move toward "Assistive AI"—a framework designed to augment human decision-making rather than replace it. This approach focuses on activating individuals and communities to protect themselves, using AI as a tool to rely on during complex decision-making processes.

Comparison between traditional regulatory approaches and the proposed Assistive AI framework.

The "Board" as a Governance Unit

At the heart of this framework is the concept of the "Board." A Board is a human-led, legally accountable decision-making body—such as an editorial board or a consumer protection agency—that oversees AI-driven trust networks. These Boards are hierarchical; a "higher" Board can set overarching policies or overrule decisions made by a "lower" one, creating a structured chain of accountability.

The operational mechanism of a Board, showing the interaction between AI proposals and human decision-making.

As illustrated in the mechanism above, the Assistive AI (AAIfDM) generates proposals based on established policies. For simple, routine tasks, the system can automate decisions based on these policies. However, for complex or high-stakes "hard cases," the system escalates the proposal to a human decision-maker, ensuring that ultimate accountability remains with the Board.

Technical Components: d-credit and Knowledge Bases

To support this governance structure, the paper introduces several novel technical components:

  • d-credit (Domain-dependent Credit): This is a credibility score specific to a particular field. Unlike generic reputation scores, d-credit recognizes that an expert in physics may not have the same authority in financial advice. This score propagates through networks of authors and claims to estimate reliability in real-time.
  • Knowledge Base (KB): The framework utilizes a dual-structured KB containing "Factual Knowledge" (e.g., scientific constants) and "Rules" (e.g., statutory provisions and community policies). This ensures that AI reasoning is grounded in verifiable facts and legal standards.
  • Anonymity with Accountability: To protect privacy while maintaining trust, the framework uses cryptographic methods to allow authors (like whistleblowers) to remain anonymous unless a Board determines that revealing their identity is necessary for legal accountability.
  • Transparent Reasoning: The authors suggest integrating Multimodal Foundation Models with Relation Extraction (RE) and First-Order Logic (FOL). This combination allows the AI to process diverse data types while ensuring its reasoning remains transparent and verifiable against the Knowledge Base.

Practical Applications and Proof-of-Concept

The paper demonstrates the framework's utility through several qualitative use cases:

  1. Pharmaceutical Advertising: The system can extract metrics from multimodal ads and flag deceptive claims by cross-referencing them with a medical Knowledge Base.
  2. Academic Integrity: By analyzing citation networks, the framework can update the d-credit of researchers and flag retracted findings in real-time.
  3. Financial Services: In credit scoring and product recommendations, the framework ensures that automated decisions align with ethical banking guidelines and remain free from "black box" biases.

Conclusion and Future Challenges

While the Assistive AI framework offers a proactive solution to AI-generated threats, significant hurdles remain. The implementation complexity of integrating LLMs with formal logic solvers is high, as are the operational costs of maintaining domain-specific Knowledge Bases. Furthermore, the system's integrity depends entirely on the integrity of the Boards themselves. Future research must focus on the legal accountability of Board members and the technical challenge of mitigating AI "hallucinations" to ensure the system remains a reliable proxy for trust.

Summary updated

Abstract

Regulatory frameworks for the use of AI are emerging. However, they trail behind the fast-evolving malicious AI technologies that can quickly cause lasting societal damage. In response, we introduce a pioneering Assistive AI framework designed to enhance human decision-making capabilities. This framework aims to establish a trust network across various fields, especially within legal contexts, serving as a proactive complement to ongoing regulatory efforts. Central to our framework are the principles of privacy, accountability, and credibility. In our methodology, the foundation of reliability of information and information sources is built upon the ability to uphold accountability, enhance security, and protect privacy. This approach supports, filters, and potentially guides communication, thereby empowering individuals and communities to make well-informed decisions based on cutting-edge advancements in AI. Our framework uses the concept of Boards as proxies to collectively ensure that AI-assisted decisions are reliable, accountable, and in alignment with societal values and legal standards. Through a detailed exploration of our framework, including its main components, operations, and sample use cases, the paper shows how AI can assist in the complex process of decision-making while maintaining human oversight. The proposed framework not only extends regulatory landscapes but also highlights the synergy between AI technology and human judgement, underscoring the potential of AI to serve as a vital instrument in discerning reality from fiction and thus enhancing the decision-making process. Furthermore, we provide domain-specific use cases to highlight the applicability of our framework.
BibTeX
@misc{gyongyossy2024assistive,
  author = {Natabara Máté Gyöngyössy and Bernát Török and Csilla Farkas and Laura Lucaj and Attila Menyhárd and Krisztina Menyhard-Balázs and András Simonyi and Patrick van der Smagt and Zsolt Zodi and András Lorincz},
  title = {{Assistive AI for Augmenting Human Decision-making}},
  year = {2024},
  eprint = {2410.14353},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.14353},
}
Figure