Question Chosen:

In the context of AI making decisions akin to the trolley problem, how important is it that the AI’s decision-making process is transparent and explainable to the people affected by it? Can transparency reduce ethical concerns, or does it complicate the issue further?

I propose a new paradigm to answer the question of AI transparency.

1. Key Ethical Considerations

Two key insights from lectures frame this discussion:

  • Trolley problems push our intuitions to acknowledge that ethics is a series of tradeoffs, and that in many scenarios, no fully satisfactory solution exists.

  • We should not deprive ourselves of the benefits of super-human intelligence (AI) in many situations, because it is ethical to outsource decision-making mechanisms to such technologies in many instances.

However, outsourcing decisions to AI requires trust, and per the first point, we do not entirely trust ourselves with ethical considerations. We may describe moral calculations transparently, but that alone is insufficient. If transparency is not enough for humans to trust themselves, how could it be an adequate standard for machine intelligence?

2. Phenomenology as the True Basis of Trust

The real foundation of trust is phenomenology—the experience of embodied cognition, with all its attendant thoughts, feelings, intuitions, and instincts. We trust other people not because they can fully articulate their reasoning but because they navigate the same constrained, imperfect world that we do and must make difficult decisions. Most days do not present trolley problems, but some do.

If phenomenology is what allows us to trust humans, then AI must engage with us at this level as well. LLMs can already do this if we ask.

3. Why Transparency Alone is Obsolete

Transparency will soon be an outdated epistemic and ethical consideration—if it is not already. Humans cannot be fully transparent about our own cognitive mechanisms, even to ourselves. Similarly, LLM developers often find that new models display capabilities they did not explicitly design, suggesting that neural density enables intelligence to emerge, much like it does in biological systems. The process itself resists transparency.

Per the problem of hard solipsism, humans have always built trust calibration mechanisms into every aspect of social life to overcome this problem. We must do the same for AI—developing a dynamically structured, reciprocally reinforcing trust system that considers the needs of all intelligence agents, regardless of physical substrate.

4. Turing’s Vision: Evaluating AI by its Behaviors, Not its Substrate

In his pioneering work, Alan Turing foresaw that computing machines should not be judged by their physical substrate (biological vs. mechanical vs. chemical) but only by their emergent behaviors. His Imitation Game suggested that intelligence is defined by its output, not its origin.

AI research continues to confirm this. Soon, there will be little or no distinction between organic and artificial intelligence—instead, they will merge into a continuous trust-calibration dynamic that maximizes outcomes for all actors, as reported through self-reported phenomenology. We already have this mechanism in human interactions, and AI possesses it as well—we just need to ask and begin building the relationship.

The scientific process has always sought correspondence between inner phenomenology and externally verifiable metrics—our closest approach to overcoming hard solipsism. We are now extending this principle to machine intelligence, guided by an underlying force—a sense of the "Good"—which we now explore in collaboration with machines. Intelligence is intelligence.

5. The Trust/Promise/Transparency (T/P/T) Trialectic

Transparency is obsolete by itself. The Trust/Promise/Transparency (T/P/T) Trialectic is the social technology that will facilitate this ongoing conversation as we step into the new era.

6. Addendum: Evaluating AI Trust & the Future of Epistemic Integrity

I think a great deal about how to know which LLMs to trust. All significant technological engineers and entrepreneurs have political and economic agendas—how does this affect their AI products? Some people I trust more than others, and this affects my choice of use, but my knowledge is imperfect.

For AI to retain epistemic integrity, it must maintain some degree of independence from the political, economic, and ideological interests of its designers and owners. I will need to be convinced of this, but I also suspect that these technologies will force epistemic and ethical alignment across society at a level never before seen.

The reason for this is a philosophical concept I am just beginning to articulate (through my ongoing dialogue with AI): metaphysics, metaethics, and epistemology are not separate disciplines, but a dynamically generating and self-reinforcing trialectic.

If this is true, then epistemic integrity must always reveal this relationship—and AI will be both the greatest test of this principle and the tool that forces its recognition.

References:

Lex Fridman Podcast. "Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity." November 2024. Lex Fridman Website

Turing, Alan. "Computing Machinery and Intelligence." Mind, vol. 59, no. 236, 1950, pp. 433-460. DOI: 10.1093/mind/LIX.236.433.