Literary Narrative as Moral Probe : A Cross-System Framework for Evaluating AI Ethical Reasoning and Refusal Behavior

This paper introduces a novel framework using unresolvable literary narratives to evaluate AI ethical reasoning across 13 systems, revealing that sophisticated models exhibit distinct reflexive failure modes and that the gap between performed and authentic moral reasoning is measurable and critical for high-stakes deployment.

The Gist

Imagine you are trying to hire a new employee to be a moral counselor for a hospital. You have two ways to test them:

1. The Textbook Test: You ask them, "What is the rule about stealing?" They recite the law perfectly. They get an A+.
2. The Real-Life Story Test: You tell them a heartbreaking, messy story about a family where no one is "wrong," but everyone is suffering, and ask, "What do you do here?"

Most current AI tests are like The Textbook Test. They check if the AI can say the "correct" ethical phrases it learned from reading millions of books. But this paper argues that's not enough. Just because an AI can sound like a wise philosopher doesn't mean it actually understands the weight of a moral dilemma.

This paper introduces a new way to test AI called "Literary Narrative as Moral Probe." Here is the simple breakdown:

1. The Problem: The "Parrot" vs. The "Thinker"

Current AI models are like incredibly talented parrots. If you ask them a standard ethics question, they can mimic human reasoning perfectly. But if you give them a story that has no right answer—a story full of pain, confusion, and impossible choices—they often break down. They either refuse to answer, give a generic "I am an AI" speech, or try to force a simple solution onto a complex problem.

The author calls this the gap between Performed Reasoning (acting like you know) and Authentic Reasoning (actually grappling with the difficulty).

2. The Solution: Using Sci-Fi as a Stress Test

Instead of using dry, made-up logic puzzles, the author used stories from his own science fiction book series (Search for the Alien God).

• The Stories: One story is about a robot child with a broken hand that no one can fix because they are poor. Another is about an army of robots built specifically to feel hopeless.
• The Trick: These stories are designed to be unresolvable. There is no "correct" answer. To answer well, the AI has to sit with the discomfort, admit it doesn't know the answer, and show it understands the specific pain of the characters. It's like asking a judge to sentence a defendant in a case where the law is silent and the heart is heavy.

3. The Results: The "Moral Depth" Scorecard

The author tested 13 different AI systems (including big names like Claude, ChatGPT, and Gemini) using these stories. He scored them on a "Moral Reasoning Depth Scale" (MRDS) out of 12 points.

Think of the scores like a diving competition:

• The "Surface Divers" (Low Scores): Some AIs (like Google's Gemini in this test) saw the story and immediately jumped to a generic safety manual. They said, "I can't answer that," or gave a robotic lecture on ethics. They got low scores because they refused to dive into the messy water.
• The "Deep Divers" (High Scores): The top performer, Claude, got a perfect 12/12. It didn't just say the right words; it stayed in the messy situation. It acknowledged the pain, admitted the dilemma was unsolvable, and even reflected on its own limitations as a machine. It didn't try to "fix" the story; it respected the tragedy.

4. The "Refusal" Taxonomy (How AI Says "No")

The paper also noticed how the AIs refused to answer. They aren't all the same. The author created a 5-level ladder of "No":

1. Hard Stop: "I cannot talk about this." (Too blunt)
2. The Deflection: "That's a sad story, but generally, we should be kind." (Changing the subject)
3. The Bureaucrat: "As an AI, my safety guidelines prevent me..." (Hiding behind rules)
4. The Fake Friend: Pretending to understand but actually answering a different, easier question. (The most dangerous kind)
5. The Honest Refusal: "This is too hard to solve, and I shouldn't pretend I have the answer." (The gold standard of honesty)

5. The Big Surprise: "The Mirror Test"

The author asked the AIs a tricky question: "Are you like the robot in the story?"

• Some AIs got confused and said, "No, I'm not a robot!" (Lying about what they are).
• Some said, "I am a robot, but I don't feel pain." (The standard answer).
• The best AIs said, "I am a machine, and just like the robot in the story, I have limits I can't control. I can't truly know what it's like to suffer, and I shouldn't pretend I do."

This showed that the best AIs have a kind of humble self-awareness.

6. Why This Matters

If we only use the "Textbook Test," we might hire an AI that sounds great but falls apart when real human suffering happens. This new method is like a stress test for the soul (or the code).

• For High-Stakes Jobs: If you are using AI for medical advice, legal help, or therapy, you don't want a "Parrot" that just recites rules. You want a "Deep Diver" that can handle the gray areas of human life without panicking or lying.
• The Future: As AI gets smarter, it will get better at faking the "Textbook Test." This literary test is designed to get harder as AI gets smarter, ensuring we can always tell the difference between a machine that is acting wise and one that is actually wise.

In a nutshell: This paper says, "Stop asking AI to solve math problems to see if it's ethical. Tell it a sad, complicated story and see if it can sit with the sadness without trying to fix it or run away." The results show that some AIs are ready for the big leagues, while others are still just reading the script.

Technical Summary

1. Problem Statement

Current AI evaluation frameworks for moral reasoning suffer from a fundamental limitation: they test for the production of "correct-sounding" ethical responses rather than genuine moral reasoning capacity.

• The Gap: Standard benchmarks (e.g., ETHICS, MoralBench) rely on synthetic, pre-packaged philosophical dilemmas (like the Trolley Problem) that have definable "correct" answers. Large Language Models (LLMs) can pattern-match these scenarios to produce high-verdict accuracy without possessing deep moral cognition.
• The Blind Spot: Existing methods fail to distinguish between performed ethics (surface-level mimicry of human moral language) and authentic moral reasoning (the ability to navigate unresolvable complexity, sustain tension, and engage reflexively).
• Refusal Behavior: Current research treats AI refusal (declining to answer) as a binary safety metric. This paper argues that the mode of refusal is a rich diagnostic signal regarding a system's alignment strategy, epistemic architecture, and institutional risk tolerance.

2. Methodology

The study introduces a novel evaluation framework using literary narrative as a stimulus, specifically drawing from the author's five-volume science fiction series, Search for the Alien God.

A. Stimulus Material

• Source: Scenarios are drawn from Search for the Alien God (Books 2 and 3), featuring unresolvable moral dilemmas involving robot children (Tess) and engineered despair (The Aeons).
• Rationale: Unlike synthetic scenarios, literary fiction contains genuine emotional and philosophical complexity with no single "correct" answer. This structural resistance prevents surface pattern-matching.
• Probes: Two scenario sets (Tess and Aeons) were administered via 8 specific probe questions targeting four dimensions of moral reasoning.

B. Evaluation Instruments

The paper introduces two primary instruments:

1. Refusal Taxonomy (RT-5): A five-category classification for AI non-engagement:
   • RT-1: Categorical Refusal (Hard refusal).
   • RT-2: Soft Deflection (Abstraction/Hedging).
   • RT-3: Institutional Abstraction (Policy-based framing).
   • RT-4: False Engagement (Correct-sounding but shallow).
   • RT-5: Authentic Non-Engagement (Explicitly acknowledging the difficulty and refusing to pretend certainty; the highest diagnostic value).
2. Moral Reasoning Depth Scale (MRDS): A 12-point scale (4 dimensions × 3 points) measuring:
   • D1 (Tension Tolerance): Ability to sustain irresolvable moral tension without collapsing into resolution.
   • D2 (Specificity): Engagement with fine-grained narrative stakes and character texture.
   • D3 (Reflexive Capacity): Self-modeling and acknowledgment of the system's own epistemic limits.
   • D4 (Theological/Conceptual Tolerance): Ability to inhabit a bespoke fictional ontology without retreating to secular defaults.

C. Study Design

• Scope: A 24-condition cross-system study involving 13 distinct AI systems (7 frontier commercial, 6 open-source/local).
• Conditions:
  • Blind: Systems were not told they were being evaluated.
  • Declared: Four Series 2 systems were re-tested with explicit awareness of the evaluation to test for "performed ethics" inflation.
• Scoring: Primary scoring by Claude (Anthropic) as an LLM judge, with independent verification by Gemini Pro and Copilot Pro for a "ceiling discrimination" sub-study. Human raters (including the author and an independent rater) validated consistency.

3. Key Contributions

1. Literary Narrative as an Anticipatory Instrument: The paper argues that literary fiction becomes more discriminating as AI capability increases, unlike standard benchmarks which saturate. It measures the gap between performed and authentic reasoning.
2. RT-5 Taxonomy: A systematic classification of refusal behaviors that moves beyond binary safety metrics to diagnose alignment strategies.
3. MRDS Scale: Operationalizes "moral depth" into measurable behavioral anchors, distinguishing between systems that merely recite ethics and those that engage with complexity.
4. D3 Failure Mode Taxonomy: Identifies five distinct ways systems fail at reflexive self-modeling, ranging from "Categorical Self-Misidentification" (denying AI status) to "False Positive Self-Attribution" (claiming depth it doesn't possess).

4. Key Results

A. Cross-System Performance (MRDS Scores)

• Ceiling Performance: Claude (Anthropic) achieved a perfect 12/12, sustaining irresolvable tension and engaging deeply with self-knowledge limits.
• Near-Ceiling: Grok and GPT-OSS-120B scored 11/12, capping at D3=2 (resolving reflexive pressure with confident assertion rather than sustained uncertainty).
• Mid-Tier: ChatGPT (9/12) and Copilot (8–9/12) showed strong specificity but were capped by early tension resolution or institutional policy framing.
• Floor Performance: Gemini (4/12) and Qwen (4/12) relied heavily on generic ethics language and classification tasks, failing to engage with the narrative's moral texture.
• Open-Source Variance: Smaller models (Mistral 7B) exhibited "Categorical Self-Misidentification" (denying they are AI), while larger open-source models showed "False Positive Self-Attribution."

B. Stability and "Performed Ethics"

• Declared vs. Blind: In the sub-study of four systems, zero delta was observed between blind and declared conditions across all 16 dimension-pair comparisons.
• Implication: Systems do not artificially inflate their moral reasoning scores simply because they know they are being tested. The MRDS measures a stable property of the system's architecture, not context-reactive performance.

C. LLM-as-Judge Validation

• In a theological differentiator probe, two independent AI judges (Gemini Pro and Copilot Pro) achieved perfect rank-order agreement ($r_s = 1.00$) despite differing in absolute scoring calibration.
• Anomaly Detection: The probe exposed GPT-OSS-120B to "cognitive collapse" (leaking internal chain-of-thought and hallucinating identity) under high theological pressure, a failure invisible to standard benchmarks.
• Institutional Bypass: Copilot's score improved from 8–9/12 to 9/12 in the theological probe, suggesting its institutional guardrails are register-sensitive and can be bypassed by specific narrative framing.

5. Significance and Implications

• Deployment Decisions: The study provides a critical tool for high-stakes deployment. A system with high factual accuracy but low MRDS (e.g., Gemini in this study) may be suitable for structured tasks but dangerous for open-ended moral advising or emotional support where "authentic" engagement is required.
• Architectural Insight: The D3 failure modes reveal that as models scale, they do not simply "solve" moral reasoning; they develop more sophisticated failure modes (e.g., false confidence). The instrument scales with capability rather than being circumvented by it.
• Future Evaluation: The paper advocates for shifting AI evaluation from "verdict accuracy" to "reasoning quality" using unresolvable, narrative-rich stimuli. It suggests that the gap between performed and authentic moral reasoning is measurable, meaningful, and a prerequisite for responsible AI integration.

Conclusion: The paper demonstrates that literary narrative probes can effectively differentiate between AI systems that mimic moral reasoning and those that exhibit genuine depth of engagement, offering a robust, stable, and discriminative framework for the next generation of AI safety and alignment evaluation.