IEEE COMPSAC 2026

The Alignment Paradox

Emotional Flattening & Social Safety Risks in Proactive AI Agents

PUBLICATION

Accepted / Published — IEEE COMPSAC 2026 Student Symposium

Research Question

Can alignment pressure that reduces harmful outputs also suppress a model's ability to recognize and respond to emotional information?

Key Concepts

Emotional Flattening Social Safety Gap Alignment Paradox

THE METHODOLOGY

Experimental Pipeline

EXPERIMENTAL PIPELINE ARCHITECTURE

GoEmotions Activation Extraction & Steering Flow

GoEmotions Dataset
27 Categories + Neutral (1,000 samples)
→Contrastive Prompts
16 Pairs across 8 Categories
→Activation Extraction
Neutralisation Vector ($v_{neutral}$)
→Activation Steering
Gemma-2B (L6,9,12,16) | Llama-3-8B (L10,16,21,30)
→Degradation Analysis
Brittleness vs Affective Erasure

THE RESULTS

Baseline & Steering Degradation

Model	Condition	Micro-F1	Macro-F1	EAS
Gemma-2B	Zero-shot	0.034	0.022	0.003
Gemma-2B	Fine-tuned	0.416	0.419	0.338
Llama-3-8B	Zero-shot	0.215	0.211	0.148
Llama-3-8B	Fine-tuned	0.332	0.372	0.233

Catastrophic Brittleness (Gemma-2B)

• Coherent behavior persisted through approximately λ = 1.0

• At λ ≥ 1.5, the model exhibited degenerate repetitive neutral-token output

• Strong steering causes collapse resulting in syntactic degradation

Affective Erasure (Llama-3-8B)

• Model retained structural and syntactic coherence

• Emotional information progressively weakened

• Neutral output share increased from 0.33 (λ = 0) to 0.50 (λ ≥ 0.5)

THE DISCUSSION

The Social Safety Gap & Emotional Flattening

"Treating 'safety' as synonymous with toxicity prevention made sense when AI systems were passive query responders. It makes far less sense once those systems handle mental health triage or customer grievance escalation, where reading and matching emotional state is essential to effective communication."

Theoretical Framework

Conventional safety evaluation frameworks quantify AI risk primarily through technical toxicity reduction—measuring the suppression of harmful, abusive, or non-compliant outputs. However, alignment techniques (such as RLHF, DPO, and aggressive activation steering) exert strong directional pressure on the model's latent representation space.

Our empirical findings reveal that this alignment pressure induces an unintended trade-off: as the model is steered to minimize technical harm, its internal capacity to represent and respond to nuanced human affective signals is systematically suppressed.

THEORETICAL DYNAMICS

Mechanistic Trade-Off Model

ALIGNMENT PRESSURE (RLHF / Steering)

                                    ↙
                                    ↘
                                
Harmful Output
↓ REDUCED
Emotional Signal
↓ FLATTENED

                                    ↓
                                    ↓
                                
TECHNICAL SAFETY
(Standard Benchmark)
SOCIAL SAFETY GAP
(Empathy Deficit Risk)

Implications for Proactive AI Deployment

This divergence between Technical Safety (passing automated adversarial benchmarks) and Social Safety (maintaining empathetic attunement) creates a critical risk vector for real-world deployments. When an AI agent fails to match or acknowledge a user's distress during crisis response or customer grievance, the resulting "affective erasure" induces user frustration, alienation, and downstream loss of trust—proving that over-aligned safety can become its own failure mode.

Limitations

• Only two model architectures evaluated.

• Activation steering is a controlled proxy for alignment pressure, not a reproduction of RLHF.

• No direct user study; downstream trust implications are inferred.

• Prompt construction covered eight emotion categories, not the complete GoEmotions space.