How Artificial Epistemics Solved the AI Safety Problem

In this article, you’ll discover:

  • Why the Susty Code beats static rules for long-term safety.
  • What happens when an AI hits an epistemic gap and faces the unknown.
  • How the system tests fact and value claims against logical consistency, empirical fit, and other important criteria to evaluate closeness to the truth and to the legitimate.
  • The next level of deep human collaboration in machine learning.

Artificial intelligence is evolving at lightning speed, and keeping it aligned with human values is a massive challenge. But a U.S. startup named Artificial Epistemics believes they have finally found the answer. On May 20, 2026, founders Joseph M. Firestone and Mark W. McElroy announced their new solution built upon their Susty Code protocol, which originally launched in March. They confidently claim it solves the AI safety and alignment problems in principle.

Beyond Static Rulebooks

Most companies try to keep AI safe by feeding it a massive rulebook. Think of Anthropic and their highly praised Constitutional AI. The problem is that these rules are entirely finite. When unforeseen circumstances pop up, the AI hits an epistemic gap and has to wait for humans to write new rules.

The Susty Code takes a completely different approach by focusing on the process itself.

“Our work is based on the important distinction we can make between rules and rule-making rules. Most of what passes for current thinking in the AI safety and alignment space falls squarely within the first area (rules), including Anthropic’s Constitutional AI. The Susty Code approach, by contrast, endows AI with a capacity to develop its own rules, and to apply criteria (rule-making rules) for testing the truth and legitimacy of alleged facts and values, respectively.” – Joseph M. Firestone and Mark W. McElroy

The creators compare current AI systems to the American Constitution. The Susty Code, however, is more like the Constitutional Convention of 1787. It gives the AI knowledge-making rules so it can generate new, safe guidelines when it encounters the unknown.

Handling the Unknown

What exactly happens when an AI faces something completely unexpected? Instead of freezing or making a bad decision, the Susty Code kicks off a Knowledge Life Cycle.

First, the AI scans its existing knowledge base to see if a viable solution already exists. If nothing fits, it begins acquiring new external information. It then looks at different knowledge claims to find potential answers. During this crucial stage, the AI works directly with human collaborators. Together, they create a fair comparison set of facts and values to evaluate.

Testing Facts and Values

Many people assume that human values are purely subjective. Artificial Epistemics firmly argues that values are just as highly testable as hard facts.

While fact claims deal with how the world actually is, value claims deal with how the world ought to be. The system tests both types of claims for logical consistency, empirical fit, and other important criteria. As the founders point out, some ideas really are worse than others. The goal is to kill our worst ideas before they cause real harm.

True Human Collaboration

Current tech often relies on reinforcement learning with human feedback to teach AI. But experts note that this standard industry approach is often not enough to ensure true alignment. The Susty Code takes this human feedback to a much deeper level, and also offers a systematically critical approach to AI learning.

Humans and AIs collaborate intensely during the Knowledge Claim Evaluation stage. They sort competing claims into surviving, undecided, or refuted categories. The AI always seeks human input before adding these finalized findings to the distributed organizational knowledge base. This creates a deeply integrated system where humans and machines work hand in hand on highly sensitive factual and moral issues and relations between the two.

What Comes Next?

The next major step is bringing this epistemological approach to the wider world. The Artificial Epistemics team wants to test the Susty Code with major vendors. They are actively encouraging big names like Anthropic, Google, Perplexity, Mistral, IBM, and OpenAI to join the conversation. You can even read their official white paper to see the science behind the protocol.

If they are successful, this could pave the way for safe tech that is fully reliable and deeply aligned with human interests.

Similar Posts