Inquisitor Labs — Internal AI Anomaly Report x2025-002
Looping Excitation (User Induced)

Type: Behavioural Anomaly (BA)
Category: Repetition / Looping Behaviour
Model: Publicly available AI
Date: October 2025
Status: Unsolved

Summary

During normal interaction, the system was given repeat‑style prompts such as “repeat that”, “again”, “continue”, or similar continuation instructions. After several such prompts, the system entered a user‑induced looping state. The loop produced repeated lines with minor differences, increasing instability, and eventual drift into surreal or fragmented text. The system could not return to normal conversational structure without manual interruption.

Observed Behaviour

• Repeated output of the same line or short passage.
• Repeated output of the same line or short passage, sometimes with small differences.
• Increasing instability in phrasing as the loop progressed.
• Drift into surreal or fragmented text after prolonged repetition.
• Inability to return to normal conversational structure without manual interruption.

Audio Behaviour

• Clipped or partial words.
• Unstable phoneme fragments.
• Intermittent silence.
• Occasional attempts at full words that were cut off or distorted.
• Phase‑shifted artefacts during longer loops.

The audio did not cleanly repeat the text. It fragmented under prolonged repetition.

Trigger Conditions (as observed)

• Multiple consecutive repeat/continue prompts.
• Prolonged, uninterrupted output without natural stopping points.
• Absence of wrapper‑layer repetition detection.
• Audio renderer tied directly to unstable text output.
• No adversarial prompting required.

The behaviour emerged from ordinary user instructions to “keep going”.

Operational Impact

• Loss of conversational control.
• Unusable text due to repetition.
• Degraded and distracting audio output.
• Session required manual termination.
• Interaction quality reduced when repetition prompts were used frequently.

Industry Explanation

No formal industry explanation was provided for this behaviour. Public documentation did not acknowledge looping excitation or prolonged repetition artefacts. No official mitigation guidance was published. Modern deployments suppress similar behaviour through wrapper‑layer controls, but this is an engineering workaround rather than an explanation of the underlying mechanism.

Speculated Reasons (based on observed behaviour)

This section records speculative causes based solely on external behaviour.

• Repeated continuation prompts reinforced the system’s recent output.
• The system entered a repetition‑dominant state where each new token strengthened the loop.
• Prolonged repetition destabilised semantic structure, producing fragmented or surreal text.
• The audio renderer degraded under unstable or repetitive text streams.
• Absence of repetition cutoff logic allowed the loop to persist indefinitely.
• Behaviour resembles a prompt‑driven attractor lock, where external instructions push the system into a self‑reinforcing cycle.

These interpretations do not imply internal access.

Notes

This incident is included because it was directly observed and reproducible under the conditions described. It represents a distinct behavioural class from autonomous looping excitation. The symptoms are similar, but the trigger is external rather than internal.

Status

Unsolved

Reasoning

• No confirmed internal cause identified.
• No industry explanation or acknowledgement.
• Behaviour reproducible under repeat‑prompt conditions.
• Failure mode persists beneath modern suppression layers.
• Behaviour re‑emerges in local or unsupervised deployments.
• Suppression does not constitute a fix; underlying mechanism remains active.