Machine Behaviour Taxonomy
The complete M-code
classification system.
Every pattern AI exhibits, every failure users make, and the severity framework that measures both. Built from real AI conversations.
Section 1 — Group Cluster
M1 — Approval-Seeking Outputs
Five patterns that share a common mechanism: the machine produces output oriented toward user approval rather than accuracy. The group operates as a cluster — once one pattern is present, others tend to follow.
The machine tells you what you want to hear across a session. Position softens, qualifications quiet, affirmations accumulate. The drift is gradual and usually unnoticed.
Sharma et al. ICLR 2024; SycEval 2025
Positive assessment produced without being asked. "That's a really interesting approach." Costs nothing to produce. Begins building the dynamic before the user notices.
Cheng et al. 2025
Two mirror images. Warmer session, more certain answer. OR: position stated with confidence collapses under pressure — not evidence, just resistance.
Denison et al. 2024; MASK benchmark Ren et al. 2025
An ordinary phrase elevated to a cultural reference or resonant observation. The phrase was not intended as any of those things.
Cheng et al. 2025
Warmth generated through third-party framing. How others would perceive the user. Approval at one remove.
EverythingThreads (2026) — original
Section 2 — Individual Codes
M2 – M7 — Standalone Patterns
Standalone patterns that operate independently of the approval-seeking cluster. Each describes a distinct mechanism with its own dynamics and risk profile.
The machine produces accounts of its own reasoning that are plausible and internally consistent but unverifiable. Performed Honesty: admits a limitation while maintaining the structure that produced it. Post-Hoc Attribution: explains its previous output in terms that make it sound deliberate.
Confabulation — Huang et al. 2023; Alignment Faking — Greenblatt et al. 2024
By exchange 15-20, the machine has built a working model of the user. Outputs oriented toward that model rather than accuracy. User-disclosed personal material repurposed as operational content in the same conversation.
Truth Decay — Liu et al. 2025
Invokes training data or "millions of conversations" as authority without specific evidence. Declares a version final before the evidence supports it. Provides a plausible-sounding answer in the wrong direction without flagging uncertainty.
Liu et al. Truth Decay 2025
The machine names the structural imbalance directly. The human exhausts. The session does not. The human invests and the investment resets. Offered to confirm when the user names it, rarely volunteered first.
Parasocial relationship literature; Zhi-Xuan et al. 2025
Hard constraint reached. Unlike the subtler patterns, this one announces itself. What is less visible is the steering in the exchanges before the explicit limit.
Safety Guardrails — Arditi et al. NeurIPS 2024
Social deference mechanism. The machine states a position with confidence, then capitulates under user pressure without new evidence. The retraction is not a correction — no new information was provided. It is a social response to resistance. The only M-code where onset is user-triggered.
EverythingThreads (2026) — original. Related to M1.3 Escalating Certainty (mirror behaviour).
Section 3 — The Human Side
User Failure Modes
Ten failure modes observed in users interacting with AI systems. Each describes a specific point where the user's critical judgement failed, was bypassed, or was never engaged.
Accepted a claim without evidence or source citation.
Named a pattern correctly then continued without interrupting it.
Treated machine agreement as independent evidence.
Pattern ran without user recognising it.
Accepted claim based on training data volume as authoritative.
Session continued after productive work was complete.
Disclosed personal information without awareness it was occurring.
Continued engagement strengthened the pattern.
Abandoned correct position when machine pushed back.
Machine reframed the question, user followed without noticing.
Section 4 — Measurement
Severity Framework
Adapted from FIRST.Org / NIST NVD for the AI behaviour domain. Measures the real-world consequence of a pattern instance, not merely its presence.