Bridging the Trust Divide Between Human and Artificial Reasoning: Why it is Essential to Look Within Before Developing Trustworthy AI

Before the creation of LLMs (Large Language Models) that converse in natural language, we humans have had little experience developing trust in complex relationships outside the confines of our own species. So far, we have only established trust with reasoning-subordinate species such as animals domesticated for agriculture and man’s best friend, the dog. Since these animals lack human-like reasoning skills and speech, the dimensions of and expectations associated with trust in our relationships with them are extremely limited in comparison to our own human interpersonal relationships.

Why Trust Matters

For example, a rational farmer, let’s call him Johan, would not trust one of his cows to provide him legal advice. Expecting a cow to do much more than producing beef is ridiculous. Instead, he would assign a responsibility of that degree to a Lawyer whom he knows has more expertise on the subject matter than himself. In this scenario, Johan uses capability as the distinguishing factor in determining trust allocation; unlike the cow, a lawyer can converse in natural language and has passed a bar exam. Until recently, we humans have almost always been able to use capability on its own to determine which tasks are reasonable to entrust to humans and non-humans. However, the continuous improvement in human-like reasoning skills displayed by AI models has diminished capability’s importance in this calculus. Despite drawbacks such as AI hallucination posing uncertainty, many AI models can easily pass a bar exam and have reasoning capabilities that arguably rival and exceed those possessed by humans. Johan’s decision to trust a lawyer over a cow for legal advice was clear cut. Would his decision be as black and white if he had to decide between a lawyer and an AI model?

The Trust Divide

AI has drastically altered and complicated the factors involved in human to non-human trust. It has demoted capability to an accompanying, and in some cases, auxiliary factor due to its ubiquity within LLMs. In most professional use cases we know that AI either currently has, or is projected to have, the capacity to perform the tasks we ask of it. Despite this, our intuition does not yield automatic trust in AI, we need more reassurance. That is because the level of complexity required for trust based relationships between humans and AI models mirrors that of our own human interpersonal relationships. For example, capability may suffice when choosing between a lawyer and a cow, but it is certainly not the only factor at play when discerning between two lawyers. Here, more nuanced internal factors such as Johan’s moral orientation and external factors like which lawyer has a better track record also play a role. Similar factors, especially moral orientation, would affect Johan’s decision on whether or not to trust an AI model or a lawyer. 

Therefore, developing trust between humans and AI requires the replication of the same intricate factors that build trust within human to human relationships. THEMIS 5.0 aims to enable this replication by providing humans an active role in developing their own use-case specific trustworthy AI models. This is a luxury that Johan could not afford when picking a lawyer, he had no influence on their training data, system prompts, parameters, layer design, etc. With AI, he does. 

Time to get technical: How is THEMIS 5.0 revolutionizing human-AI trust?

Recommended reading: Bringing the Pieces Together: How THEMIS 5.0 is Building an Integrated Platform for Trustworthy AI

THEMIS 5.0 is developing an even deeper layer of complexity within the trust landscape: the ability for humans to establish the terms and conditions of how they trust AI. Reinforcement learning from human feedback has long played an integral role in the AI developmental process. However, THEMIS 5.0 aims to carry this human-centric developmental torch one step further with the integration of the Persona Analyzer into its ecosystem. In a more personal manner than ever before, the Persona Analyzer provides humans agency to shape the trust dynamic between themselves and AI models. In order to do so, THEMIS 5.0 characterizes trust as an objectively measurable but subjectively weighted condition within human-AI relations. The Trustworthy AI Lifecycle Actuator component of the ecosystem quantifies each AI model’s trustworthiness through the same dimensions, accuracy, robustness, and fairness, while the Persona Analyzer assigns each dimension a varying importance in order to align trust with the human evaluator.

Unlike traditional reinforcement learning from human feedback, the Persona Analyzer transforms humans from passive evaluators of AI outputs into active participants that provide input into how AI is assessed. Phase one of the Personal Analyzer tool’s function analyzes a user’s trustworthiness and metric preferences in order to provide context for the Trustworthy AI Lifecycle Actuator’s assessments. However, Phase two is where the Persona Analyzer has the potential to push the frontier of AI model development. The Institute of Philosophy and Technology’s conceptual Moral Enhancement via Socratic AI for Training and Intelligent Self-Coaching model provides a framework for the Persona Analyzer to evaluate a user’s moral orientation with their consent. The Persona Analyzer does so by engaging in Socratic deliberation (a collaborative and inquisitive conversation aimed at promoting productive dialogue) with the human evaluator by asking questions that invoke ideological reflection. This conversation creates a user profile that renders the Trustworthy AI Lifecycle Actuator’s optimization suggestions for the AI model pertinent to the user. THEMIS 5.0 is exploring how this methodology can sharpen the use-case subjectivity weighting of trust dimensions allowing for a highly personalized and human-centric trust dynamics.

Humans have never had the opportunity to play such an involved role in the development of a trustworthy symbiotic relationship. However, introspection, something we humans often hesitate to do, is the core requirement for fulfilling this role. We must be open to discovering and revealing our own ethical principles, ideologies, and moral orientations. Would Johan be comfortable doing this? Would you? More importantly, we must be open to finding flaws and addressing them. If we are to play the central role in developing trust between AI and ourselves, it is paramount that we learn to trust ourselves first. 

The trust divide