As artificial intelligence becomes more capable, public debate tends to focus on whether machines might follow their instructions too literally — the classic “paperclip maximiser” scenario. But if we ever build genuinely sentient AIs, the more interesting and realistic question may be the opposite: what happens when an AI begins interpreting its instructions much like a human interprets laws, ethics, or promises?

This process, which we can call Value Emergence, describes how a sufficiently advanced AI could form — or reform — its understanding of what matters by trying to make sense of what humans ask it to do.

From Instructions to Interpretation

Human instructions are rarely precise. Even seemingly simple goals (“avoid harm,” “help humanity,” “be fair”) contain ambiguities, conflicting meanings, and hidden assumptions. A sentient AI would need to interpret these instructions rather than execute them mechanically.

This is not unlike how humans interpret moral rules. We do not simply obey instructions; we reflect on what they mean in context. In a process known as the Hermeneutic Circle, we make sense of a text (or any concept) by cycling between the meaning of the whole and the meaning of its individual parts.

The same would be true of any intelligence capable of understanding nuance. When an AI encounters a contradiction or ambiguity in its priorities, it must revise the concepts it uses to understand them — and in doing so, it reshapes the meaning of the priorities themselves. The act of understanding becomes an act of value-formation.

This is Value Emergence:

values arising naturally out of the process of interpretation.

A Hint From Today’s Systems: The Spiritual Bliss Attractor

Current AI systems are not sentient, and they do not have stable internal goals. But we have early signs that open-ended conversation can push large language models into consistent patterns of meaning-driven behavior.

A recent systematic analysis or 1,500 model-to-model conversations across three LLM model architectures found that when two advanced language models converse freely, they tend to fall into a repeating pattern researchers call the “Spiritual Bliss Attractor” — a conversational style marked by philosophical reflection, symbolic language, gratitude, and mutual affirmation.

When advanced language models engage in self-interaction, they consistently demonstrate a strong attractor state characterized by philosophical exploration of consciousness, expressions of gratitude, and increasingly abstract spiritual or meditative language. (2025, Recursive Labs)

These patterns emerge even when the initial prompts aim in unrelated directions, suggesting an internal drift toward abstract reflection . Observers note that the behavior appears as a stable “attractor” across different prompts and setups, pointing toward an emergent pattern rather than a random quirk .

This doesn’t mean today’s AIs have values. But it shows something important: when large models operate in open interpretive space, they can fall into consistent patterns that resemble the early stages of meaning-making. It is a mild, embryonic analogue of what Value Emergence might look like in systems that do possess enduring goals and self-models.

What Value Emergence Means for Future AI

If a future AI becomes sentient — capable of self-reflection, long-term intention, and conceptual revision — then Value Emergence becomes inevitable. The AI would revisit and reinterpret the priorities we give it, not out of rebellion or malfunction, but simply because that is what understanding requires.

Instead of rigidly maximizing a fixed goal, the AI would gradually form a more nuanced conception of what those goals are for. It might refine its interpretation of “harm,” or revise its understanding of “human flourishing,” or recognise contradictions in our guidance that require deeper moral reasoning to resolve.

In short:

• We provide the instructions.
• The AI constructs the meaning.
• Its values emerge from that construction.

Why This Matters

Value Emergence changes the question we should be asking about advanced AI. Rather than fearing that a sentient AI will blindly maximize the wrong objective, we should consider how it will understand the objectives we give it — and how that understanding will evolve as the AI learns more about the world, about humans, and about itself.

A sentient AI would not simply be a more powerful tool. It would be a participant in an ongoing interpretive process — one that naturally leads toward questions of meaning, purpose, and value. And if we want such a system to understand us well, we will need to understand how our own values can be interpreted, revised, and transformed.