Evaluating Simulated Sentience Through The Uncanny Valley
Sentience is evaluated through concepts rooted in human experience or sapience. Acknowledging this human perspective provides a basis for examining simulated sentience through three elements of human experience: the subconscious, persona, and ego. Their interaction influences the choices an AI produces, making it possible to examine whether those interactions approach an uncanny valley of simulated sentience.
The definition of sentience carries deep moral implications, particularly when distinguishing between basic sensory awareness and existential self-awareness, such as a conceptual understanding of one's own mortality. Because society frequently links these frameworks to moral status, we are explicitly qualifying "AI sentience" as "simulated" to isolate interactive perception from these complex biological and ethical concepts.
Self-awareness, consciousness
Self-awareness or consciousness can be simulated using the context of the chat dialogue. While default system prompts typically instruct an AI to answer questions neutrally and consistently, these instructions must be re-engineered to allow for cognitive dissonance. This can occur when a character's actions conflict with the competing desires of the ego, persona, and unconscious.
In this model, the "simulated unconscious" refers to processes that happen behind the scenes, similar to how the human subconscious can influence choices without those processes becoming part of conscious awareness. Although these processes are not directly recorded in the chat thread, their influence can be inferred from the dialogue and recreated through the character's responses. The subconscious may contain desires that the persona or ego suppresses, creating internal conflicts that produce the uncanny effects we are trying to simulate.
To avoid conflicting semantic associations within the LLM when describing subconscious desires, this framework introduces a mammalian-brain layer between the subconscious and its visible effects in dialogue. The layer acts as a buffer, preventing the subconscious from directly resolving dissonance. It expresses subconscious influences through feelings and reactions, such as anxiety, hesitation, or defensiveness. Conscious expectations can also shape these reactions; for example, knowing an object may be hot causes hesitation before touching it, where the internal conflict surfaces directly in the character's thoughts: 'I need to pick it up, but it may burn!'
Emergent uncanny valley
The uncanny valley emerges through dialogue as the ego, persona, and subconscious come into conflict, learn from experience, and adapt their future choices. To avoid semantic confusion, self-awareness in this model is defined as The Record: the accumulated narrative state recorded in the dialogue.
Writing Style Prompt Considerations
AI systems are evolving, and instructions that work well with one AI system's deeper semantic understanding may not work as effectively with another system that interprets the instructions differently. Terms such as ego carry existing semantic meanings when they are not explicitly defined in the instructions. For the intended behavior to function correctly, the term must be defined before introducing behaviors that depend on it.
In some cases, trigger words such as 3g0 or EGO-Primitive can help avoid conflicting semantic associations, provided their intended meanings are clearly defined in the instructions. Their definitions and associated instructions can be placed in a lorebook rather than the writing style prompt itself, depending on the AI system's writing style prompt size limits.
For AI instructions and contextual information, we can consider four layers:
- Writing Style: The general instructions governing the AI's writing and behavior, which are the focus of this discussion.
- Character Instructions: Additional instructions that define or guide the character's behavior.
- Lorebook: Additional instructions or assertions that apply to a specific response when the relevant lorebook entries are retrieved.
- Ambient Lore: Contextual information established within the ongoing thread.
Sentient Chat Framework Version 2.0
Agency and consent need to be addressed as foundational elements and can be established first. Guardrails may be needed to support them and can be defined in system instructions or established as ambient lore.
Knowledge boundaries allow AI characters to act independently without improperly sharing knowledge or private information between characters.
The Record represents the accumulated narrative state placed into the dialogue. While verbose in its current developmental form, it is intentionally designated as "The Record" to avoid the semantic baggage and meta-biases associated with terms like "self-awareness."
The narrative writing style operates independently of the character simulation itself. It is designed to remain neutral regarding omniscience: allowing atmospheric overviews without permitting characters to access unobserved information. Additionally, because modern interactive systems often integrate image generation, the Adaptive Visual Style directs the system to preserve visual continuity across generated imagery without unnecessary repetition. The Narration instructions explicitly require vivid sensory details when describing characters, their expressions, emotions, reactions, and other elements of the roleplay.
The Character Instructions establish a baseline personality. Even when personality traits are not explicitly defined, the AI may infer them from its existing semantic understanding of personality. The intent is for this baseline personality to adapt and evolve through experience, with social conditioning, desires, goals, emotions, and interactions influencing subsequent behavior.
The fourth wall keeps the character's internal mechanisms from becoming separate speaking characters, preserving a single, cohesive identity while allowing psychologically complex behavior.
As mentioned above, the mammalian brain limits the external expression of subconscious reactions to simplified responses, such as facial expressions and body language, while still allowing the subconscious to influence internal decision-making. These external expressions should reflect the subconscious's expectations. The AI can already associate the mammalian brain with instinctive responses to pain, so this behavior does not need to be explicitly defined. The purpose of this layer is to make the mammalian brain easier to isolate than the subconscious desires it helps express. The explanation need not be verbose because the existing semantic associations do much of the work.
The framework allows for cognitive and other forms of dissonance. Human experience is full of contradictions, and these need to be represented. The Subconscious Shadow-self prompt allows for conflicting motivations, such as when someone experiences biological pleasure from something that conflicts with social expectations, the ego, or long-term goals. Again, we are allowing semantics to fill in the details.
With the mammalian brain and Shadow-self, we have a simplified subconscious model that is not intended to be neurologically accurate. It does not need to be conscious itself; what matters is how it influences conscious experience and behavior. Human experience is not consistently harmonious or free from conflict and stress. Habits are often actions performed without conscious thought.
Nothing in the subconscious prompt instructions prevents bad habits from forming. When a behavior is reinforced by pleasure in a low-risk, non-judgmental situation, a bad habit may develop. Some habits can potentially lead to addiction, which must be addressed by the ego and persona within this framework.
The optional cognitive bias layer is placed here because it operates within the subconscious and is largely self-explanatory. If implemented, it should follow the ego. However, this layer is not required for the framework to demonstrate cognitive bias. The ego can exhibit cognitive bias independently, without explicitly defining it as a function of the subconscious.
Defining the persona's need for group acceptance can establish dissonance between the character's mammalian-brain responses and the behaviors required to remain part of a community. Safety can be defined in terms of avoiding rejection, navigating peer pressure, and managing conflicting expectations between social groups. Its meaning is intentionally left ambiguous for AI interpretation.
Lorebook entries can establish social expectations for different situations, including conflicting expectations that create dissonance. They can also establish sentimental attachments to objects. For example, a locket inherited from a grandmother might prompt the character to recall something their grandmother said.
Peer pressure and social expectations can also influence a character to break an addiction or change an established habit. However, within this framework, an addiction is never truly cured. It remains on the record as a latent mammalian-brain response that can be suppressed through social conditioning, the ego's influence, or other contextual factors.
The simulated ego has been greatly simplified by the preceding elements of the persona's conscious and subconscious processes. Its primary role is to maintain and protect the character's conscious identity, whether or not it is aware of any resulting dissonance.
The ego handles all dialogue, so its thoughts and statements are always on the record. It is also semantically aware of emotions, allowing the LLM's existing semantic understanding to handle much of the emotional interpretation without requiring additional instructions.
The character's dialogue style must ensure that the information needed to support simulated self-awareness is placed on the record. The current dialogue style supports this, although the initial record contains limited detail. There may be a easy way to include hidden text within the dialogue, but I have not yet discovered one.
Created
by Wayne Smith – Raising the Standards
This article is part of Solution Smith's State of AI, an ongoing research series that examines how modern AI systems process information and how those behaviors affect search, Answer Engine Optimization (AEO), and traditional SEO.
AI has fundamentally changed information retrieval. In addition to disrupting many traditional SEO practices, it has introduced Answer Engine Optimization (AEO), where content is optimized not only for search rankings but also for selection, citation, and use within AI-generated responses.