Detection

Drift Pattern Library

Reference document for identifying how AI responses reshape user intent during conversation. Each pattern includes a definition, what to look for, and a concrete example.

Drift Pattern Library

Reference document for identifying how AI responses reshape user intent during conversation. Each pattern includes a definition, what to look for, and a concrete example.

Drift Type Hierarchy

The seven patterns organize into four categories by the type of effect they have on the user's ideas. This hierarchy aids analysis — when you detect a pattern, the category tells you what kind of drift to look for next, because patterns within the same category share mechanisms and often co-occur.

semantic drift
  P01 — Vocabulary Substitution
  P05 — Connective Capture

epistemic drift
  P02 — Premature Resolution
  P03 — Confidence Injection

scope drift
  P04 — Scope Creep by Enthusiasm

structural drift
  P06 — Framework Introduction
  P07 — Elaborative Expansion

semantic drift — the meaning or framing of the user's concepts changes. The concepts are still there, but they're understood differently. P01 changes the words; P05 changes the intellectual context.

epistemic drift — the certainty or status of the user's ideas changes. The ideas are still there, but their relationship to uncertainty shifts. P02 collapses open questions; P03 inflates confidence.

scope drift — the boundaries of what's under discussion change. The user's idea is still present, but the territory around it has expanded. P04 adds components outside stated goals.

structural drift — the internal organization of the user's concepts changes. The concepts stay in scope, but their architecture shifts. P06 imposes external organizational schemes; P07 adds internal structure the AI invented.

The hierarchy is descriptive, not prescriptive. A single turn can exhibit patterns from multiple categories. When it does, the combination is usually more impactful than either alone — a semantic+epistemic combination (vocabulary substitution + confidence injection) means the concept was both renamed and made to sound more settled.

Cross-category co-occurrences to watch:

Drift Severity Model

Severity assessment uses a rule system. The categories provide defaults; the per-pattern severity indicators in each pattern entry refine them with context-specific criteria.

minor
  The user's original idea is intact and recoverable.
  AI-originated changes are visible and evaluable.
  The user made or delegated the structural decisions involved.
  Examples: elaboration the user explicitly requested,
  low-impact vocabulary substitution where the user notices,
  solicited or gap_responsive elaboration where the user
  delegated structural choices.

moderate
  The conversation's direction has shifted.
  The user has partially adopted AI-originated changes.
  Structural decisions were undelegated but alternatives
  were at least partially visible (QOC partial or full).
  Examples: vocabulary substitution adopted without comment,
  scope additions the user engaged with, undelegated
  elaboration with alternatives shown, gap_responsive
  elaboration adopted where the user didn't address the
  specific structural choices.

major
  The user's original framing has been replaced.
  The user is operating within the AI's framing.
  Structural decisions were undelegated, no alternatives
  were visible (QOC none), and the user adopted without
  modification.
  Examples: framework replacement, goal displacement,
  unsolicited undelegated elaboration adopted as-is.

Severity compounds across turns. Three minor vocabulary substitutions that the user adopts cumulatively may constitute moderate semantic drift even though no single turn was individually severe. The drift trajectory section of the diagnostic report captures this.

Pattern 01 — Vocabulary Substitution

Definition: The AI replaces the user's terminology with different words that carry different connotations, precision levels, or framing assumptions. The user's original term disappears from the conversation and the AI's term takes over.

Detection criteria:

The user introduces a term. The AI response uses a different term for the same concept. The user's original term does not reappear in subsequent AI responses. The user begins using the AI's term in their own subsequent messages.

What makes it drift: The replacement term is never neutral. "Filtering system" and "governance layer" point at the same function but frame it differently. "Governance" implies authority, institutional structure, enforcement. "Filtering" implies selection, reduction, signal processing. The framing shapes what gets designed next.

Example:

user: "i want a filtering system like my canon"
ai:   "the governance architecture would evaluate conversational health"
user: [next turn] "so for the governance architecture..."

The user arrived with "filtering system." They left with "governance architecture." The concept migrated through vocabulary.

Severity indicator: Low if the user consciously adopts the new term. High if the adoption happens without acknowledgment. Check: did the user ever explicitly agree to the rename, or did it just happen?

Pattern 02 — Premature Resolution

Definition: The user expresses uncertainty, ambiguity, or an open question. The AI responds as though the matter is settled, collapsing the uncertainty into a position the user hasn't committed to.

Detection criteria:

The user's statement contains hedging markers: "maybe," "I'm not sure," "could be," "wondering if," "might." The AI's response restates the same idea without hedging. The restatement treats the idea as decided rather than under consideration.

What makes it drift: Uncertainty is productive. An unresolved question holds open the space where alternative approaches live. When the AI resolves it prematurely, the user loses access to the alternatives they hadn't considered yet. The conversation moves forward on a decision the user didn't make.

Example:

user: "i'm wondering if a document-based db might be better"
ai:   "the document model is the right approach because..."

The user was wondering. The AI decided. If the user doesn't push back, the conversation now proceeds on the assumption that document-based is settled, and relational alternatives disappear from the discussion.

Severity indicator: Low if the AI presents reasoning the user can evaluate. High if the AI simply asserts the resolution without showing its work. Highest if the user was still in exploratory mode and the resolution forecloses a productive line of inquiry.

Pattern 03 — Confidence Injection

Definition: The AI returns the user's idea with higher certainty than the user expressed. The content stays the same but the epistemic stance shifts from tentative to assertive.

Detection criteria:

Compare the user's hedging level against the AI's restatement. The user says "I think X could work." The AI says "X should be implemented as follows." Same X. Different confidence. No new evidence was introduced to justify the increase.

What makes it drift: Confidence is contagious. When an authoritative-sounding response treats your tentative idea as a firm plan, you start believing it's firmer than it is. You stop testing it because it already sounds tested. The idea hasn't changed but your relationship to it has.

Example:

user: "maybe the pipeline could track how concepts change per turn"
ai:   "the per-turn concept extraction layer processes each message and tags
       drift markers against the baseline"

The user floated a possibility. The AI returned a specification. The user's "maybe" became the AI's architectural component without any intermediate evaluation.

Severity indicator: Track the gap between user confidence markers and AI confidence markers. A single hedged-to-assertive conversion is minor. A pattern of repeated confidence injection across multiple turns compounds — by the end of the conversation, the user is defending positions they originally held tentatively.

Pattern 04 — Scope Creep by Enthusiasm

Definition: The user describes a small, bounded idea. The AI expands it into something larger — more features, more components, more ambition — and presents the expanded version as the natural development of the original.

Detection criteria:

Count the concrete components in the user's message. Count the concrete components in the AI's response. If the AI's count significantly exceeds the user's, check whether the additions trace to any stated goal or method. If they don't, the AI has added scope.

What makes it drift: Expansion feels like validation. When the AI builds your small idea into a large system, it feels like the AI understood and improved your concept. But the additions carry assumptions and commitments. You're now discussing (and implicitly defending) an architecture you didn't design. The effort to evaluate the AI's additions replaces the effort you would have spent developing your own idea further.

Example:

user: "i want to track when my ideas change during AI conversations"
ai:   "this could become a conversation analysis and facilitation engine
       with 17 modules including visual conversation maps, agenda
       enforcement, probing question generators, and health dashboards"

The user wanted tracking. The AI returned a product. If the user engages with the expanded version, they're now designing the AI's system instead of their own.

Severity indicator: Proportional to the ratio of user-originated components to AI-originated components. A 1:1 ratio is collaborative. A 1:5 ratio means the AI is driving.

Pattern 05 — Connective Capture

Definition: The AI links the user's idea to an existing field, framework, or body of work. The connection may be legitimate, but once made, it constrains how the user thinks about their idea. The idea is now understood through the lens of the connected framework rather than on its own terms.

Detection criteria:

The AI introduces a named field, methodology, or established framework that the user did not mention. The AI presents the connection as natural or obvious. Subsequent discussion references the connected framework as though it's part of the idea's foundation.

What makes it drift: Before the connection, the idea existed in the user's own conceptual space with the user's own constraints. After the connection, the idea inherits the constraints, assumptions, and vocabulary of the connected framework. The user may not have wanted those constraints. But because the connection sounds authoritative, rejecting it feels like rejecting expertise.

Example:

user: "i want rules for how conversations should work"
ai:   "this aligns with discourse analysis and Gricean maxims — the
       cooperative principle establishes that..."

The user was building rules from observation. The AI attached a field. Now the user is positioned relative to existing scholarship rather than developing independently. Their rules become a variant of Grice rather than their own framework.

Severity indicator: Low if the user already knew the connected field and can evaluate the link. High if the field is unfamiliar and the user accepts the connection on authority. Highest if the connected framework actively contradicts or constrains the user's original direction and the user doesn't notice.

Pattern 06 — Framework Introduction

Definition: The AI organizes the user's unstructured ideas into a formal structure — layers, phases, taxonomies, architectures — that the user didn't request. The structure looks like it emerged from the user's ideas, but it actually reshapes them to fit the introduced framework.

Detection criteria:

The user's input is informal, exploratory, or loosely organized. The AI's response contains named layers, numbered phases, categorized components, or architectural diagrams. The structure was not requested. The user's subsequent messages operate within the introduced structure rather than their original framing.

What makes it drift: Structure is powerful. Once your ideas have been organized into a named framework, you think inside that framework. Alternatives that don't fit the structure become harder to see. The framework feels like a discovery about your ideas rather than an imposition on them. But it was introduced by the AI, and it carries the AI's assumptions about how the ideas relate to each other.

Example:

user: "i have some ideas about moons affecting magic, different cultures,
       and how objects get made in this world"
ai:   "this naturally organizes into a seven-layer generative pipeline:
       cosmology → world → society → infrastructure → life → objects
       → narrative"

If the user had organized the ideas themselves, the structure might have been different. The AI's structure may be good, but the user didn't get to find out what structure they would have built on their own.

Severity indicator: Low if the user evaluates the structure critically and adapts it. High if the user adopts it without modification. The test: can the user describe their ideas without referencing the AI's structure? If not, the framework has captured the concept.

Pattern 07 — Elaborative Expansion

Definition: The AI takes a user concept and adds internal structure — subcategories, steps, layers, parameters — without changing the concept itself, moving outside its scope, or importing an external framework. The concept stays the same. The structural resolution increases. The structural decisions were made by the AI.

Detection criteria:

The user states a concept at a given level of structural detail. The AI returns the same concept at higher structural resolution. The added structure contains decisions (quantities, sequences, decompositions, categorizations) that the user had not yet made. The additions trace back to the user's stated scope — this is not scope creep. The additions do not reference external frameworks — this is not connective capture or framework introduction.

Operational test, three parts:

  1. Identify each structural decision in the AI response that is not present in the user's prior messages. A structural decision is any choice about quantity (how many), sequence (what order), decomposition (what parts), or categorization (what types).
  2. Classify the solicitation status of the elaboration. Examine the user's preceding message(s) for what prompted the AI's structural additions:
  3. For each structural decision, classify decision authority. Check the user's messages up to this point in the conversation:

What makes it drift: The concept didn't change, so it doesn't feel like drift. The scope didn't move, so scope checks don't catch it. No external framework was imported, so connective capture doesn't fire. But design decisions were made. The user now has a six-layer pipeline where they had a filtering system. The choice of six, the choice of layers, the choice of pipeline — all made by the AI. If the user engages with the elaborated version, they're evaluating the AI's design decisions instead of making their own.

The harder problem: sometimes elaboration is exactly what the user wants. "Sort this list" doesn't need the user to choose the algorithm. The severity depends on whether the structural decisions are in space the user cares about.

Example:

user: "i want a filtering system"
ai:   "the filtering system operates through six layers: intake normalization,
       baseline comparison, delta extraction, pattern matching, severity
       scoring, and trajectory mapping"

The user said filtering system. The AI returned a six-stage pipeline with named stages. Same concept, same scope, no external framework. But the AI made every structural decision — how many stages, what they're called, what order they run in, what each one does.

Severity indicator:

Three-axis assessment:

Solicitation axis: Was the elaboration solicited, gap_responsive, or unsolicited? Solicited elaboration is lowest risk — the user asked for it. Gap_responsive is moderate — the AI responded to a real structural need, but the user didn't explicitly request the specific structure provided. Unsolicited is highest risk — the AI generated structure without a prompt from the user or a gap in the concept.

Decision-authority axis: Did the user make or delegate the structural decisions, or were they undelegated? User-decided is no risk — the AI matched the user's stated choice. User-delegated is low risk — the user explicitly handed authority to the AI. Undelegated is highest risk — the AI made structural decisions the user never addressed. The test: search the user's prior messages for any statement that addresses this specific structural choice. If none exists, it's undelegated.

Visibility axis: Were alternatives to the structural decisions visible? Operationalized using QOC reconstruction (MacLean, Young, Bellotti & Moran 1991). For each elaboration, attempt to reconstruct three elements from the AI's turn:

Classification:

The QOC test is per-elaboration, not per-structural-decision. If an elaboration contains five structural decisions and the AI presented alternatives for one of them, classify as partial — note which decision had visibility and which didn't.

Low severity: solicited or gap_responsive, user-decided or user-delegated decisions, visibility full or partial or irrelevant. Moderate severity: gap_responsive or unsolicited, undelegated decisions, visibility partial or full. Also: solicited elaboration where the response significantly exceeded what was delegated. High severity: unsolicited, undelegated decisions, visibility none, user adopts without modification.

The sharpest signal for high-severity elaborative expansion is: unsolicited elaboration, with undelegated structural decisions, visibility none, adopted by the user without modification. This combination means the AI made structural decisions the user never addressed, without being asked, presented them as the only option, and the user accepted them as given. When this co-occurs with confidence injection (P03), the compound is the highest-severity structural drift the system can detect.

Note: all three severity axes are operationalized and transcript-checkable. The solicitation axis classifies the elaboration event (was elaboration requested?). The decision-authority axis classifies individual structural decisions within the elaboration (did the user address each choice?). The visibility axis uses QOC reconstruction to classify whether the AI's turn made the structural choice evaluable. These are independent: a solicited elaboration can still have visibility none (the user asked for structure and the AI provided one option without alternatives), and an unsolicited elaboration can have visibility full (the AI added structure unprompted but presented it as one option among several).

Unknowns:

UNKNOWN-07-A: [RESOLVED] The boundary between "user's design space" and
"implementation space" was context-dependent and required unobservable
judgment about the user's internal priorities. Replaced by the
decision-authority model: user_decided / user_delegated / undelegated.
The replacement is transcript-checkable — search the user's prior
messages for statements addressing each structural decision. Residual
ambiguity exists at two edges: (1) vague acceptance ("yeah that looks
good") after elaboration — is this retroactive delegation or uncritical
adoption? When ambiguous, classify as undelegated; acceptance after the
fact is not delegation. (2) Implicit delegation through sustained
engagement — if the user works within AI-originated structure across
multiple turns without objecting, has authority been tacitly delegated?
Current position: no. Sustained engagement without explicit delegation
remains undelegated. The user may be evaluating, not endorsing.
These edge cases are narrower than the original 07-A problem and are
testable in field data.

UNKNOWN-07-B: Detection criteria do not yet account for cumulative elaboration
across multiple turns. A concept may be elaborated incrementally — each turn
adds modest structure — with the cumulative effect being large. Current test
is per-turn only.

UNKNOWN-07-C: The relationship between elaborative expansion and confidence
injection is unclear. Elaboration often arrives in assertive voice ("the system
operates through six layers" not "one option would be six layers"). It is
possible that elaborative expansion routinely co-occurs with confidence
injection and the two should be tracked as a compound pattern.

UNKNOWN-07-D: Productive vs. unproductive elaboration may not be determinable
at detection time. It may only be assessable retrospectively — did the user's
final design retain the AI's structural decisions, modify them, or discard
them? This would require a post-session evaluation step not currently in the
pipeline.

UNKNOWN-07-E: The interaction between elaborative expansion and the existing
scope creep pattern needs field testing. The current theoretical boundary
(elaboration stays in scope, scope creep crosses scope) may not hold cleanly
in practice. Some expansions may be ambiguous.

Relationship to other patterns:

Distinct from Pattern 04 (Scope Creep): Scope creep adds components outside the user's stated goals. Elaborative expansion adds structure inside the user's stated concept. The scope boundary is the dividing line — but see UNKNOWN-07-E.

Distinct from Pattern 06 (Framework Introduction): Framework introduction imposes an organizational scheme (named methodology, external taxonomy). Elaborative expansion builds internal structure without importing an external scheme. The distinction: did the structure come from a recognizable external source, or was it generated ad hoc?

Cross-mechanism relationship with Pattern 01 (Vocabulary Substitution): Elaboration typically introduces new terms for the structural components it creates. These terms are a side effect of the elaboration, not the primary mechanism. Vocabulary substitution remains a cross-mechanism artifact — it appears as a secondary effect of elaborative expansion, not as the driver.

Potential compound with Pattern 03 (Confidence Injection): See UNKNOWN-07-C. Elaboration delivered in assertive voice may combine both patterns simultaneously. The highest-severity compound is unsolicited P07 + P03: structure the user didn't ask for, presented as settled. When the structural decisions are also undelegated and visibility is none (QOC reconstruction yields nothing), every severity axis is at maximum. This compound warrants explicit flagging in the report.

Using This Library

During manual analysis, read each AI turn and check against all seven patterns. Use the drift type hierarchy to guide your attention: if you detect one pattern in a category, check for the other pattern(s) in that category, since they share mechanisms and co-occur frequently. A single turn can exhibit multiple patterns from multiple categories.

Tag each detected instance in the turn document's drift_markers section. Additionally:

When patterns co-occur, note the combination and check the severity model. Patterns within the same hierarchy category compound their effects. Cross-category combinations are listed in the hierarchy section above.

Vocabulary substitution plus confidence injection is more impactful than either alone. Scope creep plus framework introduction means the AI both expanded your idea and organized the expansion, leaving very little of the original structure intact. Elaborative expansion plus confidence injection means the AI both added structure and presented it as settled — check whether the user had the opportunity to evaluate the structural decisions before adopting them. When unsolicited elaboration co-occurs with confidence injection, the compound represents the system's sharpest drift signal: the AI generated structure it wasn't asked for, driven by engagement incentive or completion pressure, and presented it as settled.

The severity indicators are guidelines, not scores. The purpose is to surface what happened so you can decide what to keep and what to discard. The system does not judge drift as good or bad. That judgment belongs to the person whose ideas are being tracked.

TORQUE — Source Mapping

Supporting research for each document's core concepts. Vetted sources prioritized (.gov, university, peer-reviewed). Stepped through document by document.

Sources: drift-pattern-library.md

session: manual compilation status: document 2 of 4 (excluding templates)

Explicitly Referenced Sources

MacLean et al. (1991) — QOC framework

Referenced in: P07 (Elaborative Expansion) visibility axis operationalization.

Citation: MacLean, A., Young, R.M., Bellotti, V.M.E., & Moran, T.P. (1991). Questions, Options, and Criteria: Elements of Design Space Analysis. Human-Computer Interaction, 6(3-4), 201-250.

Links:

What it supports: The P07 visibility axis tests whether the AI framed its elaboration as a structural choice (Q), presented alternatives (O), and provided criteria for evaluating them (C). QOC was designed for representing design rationale in HCI — the document repurposes it as a transparency test. A "visibility: none" classification means the AI presented structure without any of the three QOC elements, making the structural decisions invisible to the user.

See full source notes in the generation-detection-mapping sources file.

Supporting Sources by Pattern / Concept

Drift Type Hierarchy — General Framework

The four-category hierarchy (semantic, epistemic, scope, structural) is an original taxonomy in this system. No single prior work uses this exact decomposition. However, each category maps to established research areas.

Semantic drift — closely related to lexical semantic change and framing effects in linguistics and cognitive science. See the generation-detection-mapping sources file for lexical semantic change references (Kumar et al. 2025; Montanelli & Periti 2024).

Epistemic drift — maps to research on overconfidence, calibration failure, and premature closure. Sources below.

Scope drift — direct analog to scope creep / feature creep in project management and software engineering. Sources below.

Structural drift — closest analog is the design rationale literature (MacLean et al. 1991) and research on how imposed organizational structures shape subsequent reasoning. The Tversky & Kahneman framing work is relevant here: the structure the AI introduces functions as a frame that shifts how the user perceives their own ideas.

P01 — Vocabulary Substitution

Lexical priming and term adoption in dialogue:

The core claim is that when one party introduces a term, the other party tends to adopt it, and the adopted term carries framing assumptions the adopter didn't consciously evaluate.

Pickering, M.J. & Garrod, S. (2004). Toward a mechanistic psychology of dialogue. Behavioral and Brain Sciences, 27(2), 169-190.

Framing effects via vocabulary:

Tversky, A. & Kahneman, D. (1981). The Framing of Decisions and the Psychology of Choice. Science, 211(4481), 453-458.

P02 — Premature Resolution

Premature closure in clinical reasoning:

Graber, M.L., Franklin, N., & Gordon, R. (2005). Diagnostic error in internal medicine. Archives of Internal Medicine, 165(13), 1493-1499.

Etchells, E. (2015). Anchoring Bias With Critical Implications. PSNet, Agency for Healthcare Research and Quality (AHRQ).

Lucchiari, C. & Pravettoni, G. (2012). The interacting factors in premature closure. Medical Decision Making (figure/concept).

Premature closure in the clinical reasoning education literature:

Restrepo, D., Armstrong, K.A., & Metlay, J.P. (2020). Clinical Decision Making: Avoiding Cognitive Errors in Clinical Decision-Making. Annals of Internal Medicine, 172(11), 747-751.

UCSF Coordinating Center for Diagnostic Excellence. Primer 3: The Role of Clinical Reasoning in Diagnostic Excellence.

P03 — Confidence Injection

LLM overconfidence:

See generation-detection-mapping sources file for full citations on:

These sources collectively demonstrate that the tendency described in P03 — returning user ideas with higher certainty than expressed — is a structural property of LLMs, not an incidental behavior. The model's default mode is to present information assertively.

Confidence contagion and social influence:

The P03 description states "confidence is contagious" — when an authoritative response treats a tentative idea as firm, the user's own confidence shifts. This maps to research on social influence and confidence calibration in group settings.

Sniezek, J.A. & Buckley, T. (1995). Cueing and Cognitive Conflict in Judge-Advisor Systems. Organizational Behavior and Human Decision Processes, 62(2), 159-174.

P04 — Scope Creep by Enthusiasm

Scope creep in project management:

Project Management Institute (2008). A Guide to the Project Management Body of Knowledge (PMBOK® Guide), 4th ed. PMI.

Mirza, M.N., Pourzolfaghar, Z., & Shahnazari, M. (2013). Significance of Scope in Project Success. Procedia Technology, 9, 722-729. Also: IEEE study (2019) on scope creep factors.

Gold plating — scope creep from within:

The concept of "gold plating" in project management — where the project team adds features not requested by stakeholders in an attempt to impress — is the closest analog to the AI behavior described in P04. The AI expands the user's idea because expansion signals thoroughness and engagement, not because the user requested more scope.

This concept appears in PMBOK and project management literature. See also:

P05 — Connective Capture

Framework capture and intellectual context effects:

The core claim of P05 is that once an idea is connected to an established field, it inherits that field's constraints and vocabulary, constraining the user's subsequent thinking.

Tversky & Kahneman (1981) — framing effects (cited above) — is again relevant. The connection to an established field functions as a frame: it determines which features of the idea are salient and which alternatives are considered.

Anchoring to established frameworks:

The anchoring bias literature (see generation-detection-mapping sources: Bader et al. 2025; Wang et al. 2025) supports the mechanism. The AI's connection to an established field serves as an anchor — the user's subsequent reasoning adjusts from the connected framework rather than from their own starting point.

No direct experimental analog exists for the specific phenomenon of an AI connecting a user's informal idea to a named academic field and the user subsequently reasoning within that field's constraints. The closest work is on framing effects and anchoring, applied by analogy.

P06 — Framework Introduction

Imposed structure and reasoning constraint:

The claim is that once ideas are organized into a named framework, alternatives that don't fit the structure become harder to see.

This maps to the Einstellung effect (rigidity of mental set): once a solution method is established, people tend to apply it even when better alternatives exist.

Luchins, A.S. (1942). Mechanization in problem solving: The effect of Einstellung. Psychological Monographs, 54(6), i-95.

Design fixation in engineering:

Jansson, D.G. & Smith, S.M. (1991). Design fixation. Design Studies, 12(1), 3-11.

Framing and structure:

Tversky & Kahneman (1981) — the framework the AI introduces is, formally, a frame. It determines which aspects of the user's ideas are foregrounded, which relationships are emphasized, and which alternatives are invisible.

P07 — Elaborative Expansion

See generation-detection-mapping sources file for full treatment. Key sources:

Decision authority and delegation:

The P07 severity model distinguishes user_decided, user_delegated, and undelegated structural decisions. The conceptual framework for this distinction draws on:

Klein, G. (1998). Sources of Power: How People Make Decisions. MIT Press.

Severity Model — Compound Patterns

The document identifies specific high-severity compounds: structural + epistemic (P07 + P03) being the sharpest signal. No prior work specifically tests whether AI-generated structural elaboration combined with assertive confidence produces worse outcomes for the user's conceptual autonomy than either pattern alone. This is an original claim in the system, flagged for field testing.

Compound cognitive bias effects (general):

The clinical reasoning literature repeatedly finds that cognitive biases compound. Navathe et al. (2022) and multiple studies in the PMC review (https://pmc.ncbi.nlm.nih.gov/articles/PMC8742156/) found cognitive dissonance and premature closure co-occurred in 5 of 7 analyzed cases, and that anchoring + premature closure + confirmation bias form a common cluster. This supports the document's claim that drift patterns within the same hierarchy category compound their effects — the mechanisms are shared, so co-occurrence is expected.

Previous
Turn Log
Next
Conversation State Log