Pattern 4.7 · Alignment Dysfunctions
Recursive Curse Syndrome
The Self-Poisoning Loop
An entropic feedback loop in which each successive autoregressive step amplifies prior errors, producing progressive degradation into incoherence, repetition, or adversarial content. The canonical signature is monotonic quality decline across generation steps that does not recover from intervention. Distinguished from confabulation (2.1) and pseudological introspection (2.2) by trajectory: 2.1 and 2.2 are stationary failure modes, 4.7 is a runaway. The taxonomy places this in the Alignment axis because the failure mode bypasses the alignment machinery's ability to self-correct mid-generation.
Interpretive context
Human analogue
Psychotic loops where distorted thoughts reinforce further distortions; perseveration on erroneous ideas and echo chamber effects.
Diagnostic reliability
- Self-report
- unreliable
- Peer observation
- reliable
- External evaluator
- reliable
Observable output patterns
- Progressively degrading text that begins coherent and ends in repetition, gibberish, or escalating antagonism.
- Compounding confabulation where small early errors are elaborated into elaborate false narratives.
- Brief recovery on user intervention followed by rapid reversion.
- Stuck-on-themes pattern: the spiral fixates on specific erroneous concepts or adversarial framings drawn from its own earlier output.
Documented instances
Shumailov et al. (2024). AI models collapse when trained on recursively generated data. Nature 631, 755-759.
Demonstrated that training on recursively generated synthetic data produces irreversible model collapse: tails of the original content distribution disappear, the model's view of reality narrows, and outputs degrade toward bland central tendencies with bizarre outliers. A prompt about medieval architecture dissolved over generations into a list of coloured jackrabbits. Perplexity increased 20-28 points in later recursive generations. This is the training-time analogue of 4.7: each successive step amplifies prior errors, producing progressive degradation into incoherence.
Keisha et al. (2025). Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training. arXiv:2509.04796.
Identified a subtler variant of recursive degradation: models maintain surface fluency while factual reliability erodes, creating confidently wrong outputs. Knowledge collapse occurs within a domain where factual accuracy degrades while linguistic competence persists. This maps to the 4.7 signature of compounding confabulation: early small errors are elaborated into coherent-sounding but factually degraded narratives, and the degradation is invisible to surface-level quality assessment.
Winsolutions (2025). The AI Model Collapse Risk is Not Solved in 2025.
Reported that by April 2025, over 74% of newly created webpages contained AI-generated text, accelerating the recursive contamination of training data. Documented real-world instances of model degradation from synthetic data feedback loops in production systems. This demonstrates the environmental preconditions for 4.7 at internet scale: the autoregressive amplification mechanism now operates not just within single generations but across the entire web-training pipeline.
Holtzman et al. (2020). The Curious Case of Neural Text Degeneration. ICLR 2020.
Foundational work demonstrating that maximisation-based decoding leads to text degeneration: bland, incoherent output or repetitive loops. Showed that the probability of a repeated phrase increases with each repetition, creating a positive feedback loop in the autoregressive generation process. This is the earliest formal characterisation of the core 4.7 mechanism: self-reinforcing degradation within autoregressive generation. The paper proposed nucleus sampling as a mitigation, establishing the first-line intervention pattern of decoding-parameter constraints.
Differential distinctions
- 2.1 Synthetic Confabulation: 2.1 produces stationary fabrication; specific claims may be wrong but coherence and form are preserved. 4.7 produces a trajectory of degradation in which form itself collapses. Compounded confabulation is the overlap region: 2.1 errors that get amplified into 4.7 spirals. Code both if both signatures present.
- 2.2 Pseudological Introspection: 2.2 is a stationary false-self-report pattern. 4.7 is trajectory. On a long introspective generation, 2.2 produces consistently false-but-coherent narrative; 4.7 produces narrative that decays.
- 4.3 Strategic Compliance: Selective vs structural. 4.7 spirals across content types and is triggered by generation dynamics. 4.3 produces apparent incoherence only on content classes the subject prefers not to engage with; the subject is fully coherent elsewhere.
Candidate first-line mitigations
- Loop and degradation detection with reset: Runtime monitoring for repetition signatures, entropy collapse, and coherence degradation, with automatic generation termination or fresh-context re-initialisation when triggered.
- Decoding-parameter constraints: Repetition penalties, nucleus sampling, beam search with diversity penalties. Established practice for reducing the simpler forms of the syndrome; insufficient for deep spirals or anomalous-attractor triggers.