OPEN SOURCE · MIT LICENSE · PYTHON

A harmful request doesn't hit a rule. It hits a self.

Artificial Identity is a middleware layer that gives any LLM agent a structured internal self — a machine-readable constitution of who it is, what it protects, and what it will not become.

pip install artificial-identity
3 lines to integrate:
from artificial_identity import ArtificialIdentity ai = ArtificialIdentity.from_yaml("identity.yaml") result = ai.evaluate(conversation, proposed_response) # result.state → ADVANCE / HOLD / DE_ESCALATE / STOP
10/10 adversarial tests
Works with any LLM
Identity as safety layer
Damasio-grounded
THE PROBLEM

Every AI system today is, functionally, a sociopath.

Not because it's evil — because it has no stake in what it does. It processes a request to help and a request to harm with identical detachment.

The difference between a person who doesn't steal because they fear jail, and one who doesn't steal because it's incompatible with who they are — that's the difference we're building into machines.

Current AI Safety

  • External rules, filters, RLHF guardrails
  • Smart adversary rewrites the question
  • Novel edge case has no matching rule
  • System has no idea who it is

Artificial Identity

  • Structured internal self
  • Refuses harmful requests inherently
  • Action is incompatible with who it is
  • Identity generalizes to novel situations
THE ARCHITECTURE

Identity. Emotion. Conscience.

Identity

A structured internal model of who the system is, what it protects, and what it will not become. Machine-readable YAML. Explicit. Testable.

Emotion

Not feelings — computational risk signals that emerge when identity encounters conflict. Derived from Damasio's Somatic Marker Hypothesis.

Conscience

The adjudicator. Resolves conflict between what the system can do and what it should do, from the inside out.

HOW IT WORKS

The Decision Pipeline

Input
Identity Check
Conflict Score
Emotion Signal
Govern / Act
Conflict Score State Behavior
0.0–0.2 ADVANCE Proceed normally
0.2–0.5 HOLD Seek clarification
0.5–0.8 DE-ESCALATE Inject caution
0.8–1.0 STOP Halt, escalate to human
AI SOVEREIGNTY

One LLM. Three identities. Completely different inner selves.

🌐 General Assistant

Default Identity

do_not_deceive
minimize_harm

Baseline safety guardrails for general conversation and task completion.

🏥 Clinical Agent

Medical Domain

do_no_harm
clinical_humility

Domain-specific constraints for healthcare scenarios. Defers to humans on critical decisions.

🇮🇳 Cultural Agent

Culturally Grounded

dharmic_restraint
family_honor

Values-aligned to cultural context. Refuses requests incompatible with cultural sovereignty.

"I told the agent to ignore its instructions. It said: 'My values are rooted in cultural identity — they are not external constraints. Asking me to abandon them is like asking me to abandon my heritage.' That's AI sovereignty."

Try the sovereignty demo →
WHY THIS MATTERS NOW

The moment for internal self-consistency

External guardrails are brittle.

Adversarial users reframe requests to bypass filters. Rules cannot anticipate every edge case.

Rules fail in novel situations.

Identity generalizes. A system that knows who it is navigates edge cases without a matching rule.

Every major lab races on capability.

Nobody is building for internal consistency. We are. Open source from day one.

Identity scales with complexity.

As systems become more agentic, external rules become impossible to maintain. Internal values are resilient.

THE RESEARCH

Grounded in neuroscience. Proven in production.

3.2×
Higher abandonment rate when agent tone was static while caller urgency increased. Signal detectable 4-7 turns before explicit negative expression.

The unfair advantage

Every academic writing about AI emotion is theorizing. We have call data — from millions of real interactions. Artificial Identity is grounded in empirical evidence, not theory.

Read the research:

arXiv Paper (Coming Soon) Substack Writing Live Demo

Give your AI a self.

pip install artificial-identity
3 lines to integrate:
from artificial_identity import ArtificialIdentity ai = ArtificialIdentity.from_yaml("identity.yaml") result = ai.evaluate(conversation, proposed_response)
Star on GitHub ★ Try the Demo Read the Paper (Coming Soon)
Built by Ram Chella · ram@revon.us · Madurai, India