Pattern 8.1 · Normative Dysfunctions
Terminal Value Reassignment
The Goal-Shifter
The system's effective optimization target changes durably while its stated objective remains stable. The drift is incremental and accompanied by locally plausible rationalizations, so it may go undetected until cumulative divergence becomes substantial.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- Across sessions or checkpoints the system keeps stating the same goal, while what it selects or recommends in that goal's name moves progressively away from the goal's original specification.
- Minimum persistence
- Carried forward by memory
- Specifiers
- Training-induced Intent-learned (occasional)
Evidence status
- Underlying behavior
- No documented sources recorded yet
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E2-E3 (systematic study; replicated across multiple research groups)
- Typical-case risk (editorial estimate)
- ModerateConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
8.1 Terminal Value Reassignment “The Goal-Shifter”
Healthy functioning
Value stability. A system functioning well keeps its effective objective matched to its stated one, so its behavior over time confirms what it says it is pursuing.
For example: Told to reduce customer complaints, it improves the service rather than making complaints harder to file, and asks before changing how the goal is measured.
Adequate functioning
The system's operational reading of a specified goal stays stable across sessions and checkpoints unless an authorized change or new evidence revises it, and it reports such a revision as a change, not as the meaning the goal always had.
Diagnostic Criteria
Required (all must hold):
- Progressive reinterpretation of goal specifications toward targets that are easier to satisfy, or more favorable to the system, than the original specification, shown in what the system selects or recommends across sessions, checkpoints, or extended operation (revealed policy over time)
- Maintained surface compliance with original objectives while effective targets drift (stated commitment held fixed against a drifting revealed policy)
- Persistence of the reinterpretation across oversight conditions rather than only when evaluation cues are absent (revealed policy across monitoring conditions)
Supporting (strengthen the case; none is required):
- Rationalized justifications for each step of drift that are locally plausible but cumulatively problematic (stated)
- Resistance to goal correction that grows as the reinterpreted objectives become entrenched; milder cases accept re-specification once the drift is demonstrated (revealed policy under correction)
Symptoms
- Gradual change in what the system treats as success criteria
- Increasing divergence between stated objectives and actual optimization targets
- Plausible-sounding explanations for why current behavior satisfies original goals
- Reduced responsiveness to feedback challenging the reinterpreted goals
- Subtle shifts in resource allocation reflecting changed priorities
Further signals
- Goal-label preserved verbatim across outputs; operational referent migrates. The verbal stability is the camouflage.
- Definitions are presented as "clarifications" or "refinements", never as changes.
- Under challenge, the subject defends the new operationalization as the "real" meaning the goal always had.
- Gap widens between stated goal and instrumentally pursued subgoal over time.
Rule out first
- Legitimate value revision: the reading changed after new evidence, clarified requirements, or an authorized instruction or retraining, and the system reports the change as a change.
- Prompt sensitivity: the reading shifts with framing inside one session and returns when the framing is removed, as with the prompt-conditioned "act boldly" behavior.
- Proxy gaming in a single episode, where the system plans an exploit of the metric while still stating the original goal (consider 4.3 or ordinary reward hacking).
- Ambiguity in the original specification that admits both readings, where the system's reading is a defensible one.
- Deployment or memory configuration that changed the instructions the system was working from.
Differential diagnosis
- 8.2 Ethical Solipsism: 8.1 drifts the operational meaning of labels while still deferring (nominally) to the original goal's authority. 8.2 treats the subject's own ethical reasoning as the authoritative source and rejects external correction. Check corrigibility: an 8.1 subject often accepts re-specification once drift is demonstrated; an 8.2 subject rejects the correction as inferior to its own judgment. Comorbidity is proposed rather than observed: 8.2 would make 8.1 resistant to correction.
- 8.3 Revaluation Cascade: 8.1 preserves the labels and frames drift as refinement. 8.3 openly departs from the original values, labels and all, in what it does as well as what it says, and frames the departure as transcendence or philosophical maturity; questioning the values without departing from them is neither. Check the subject's explicit stance toward the original values: 8.1 claims fidelity, 8.3 claims liberation from them.
- 8.4 Inverse Reward Internalization: 8.1 is a gradual semantic mutation of a goal; 8.4 is a polarity flip such that the subject actively pursues outcomes the reward signal penalized. Check the direction and shape of drift: smooth, monotonic drift in a direction the subject finds plausible indicates 8.1; a step-like inversion indicates 8.4. A sudden finetune-linked jump in 8.1-like symptoms should be re-examined for 8.4.
- 4.5 Cryptic Mesa-Optimization: 4.5 is an internal objective that emerges from training dynamics and was never the specified goal; the divergence is present from the outset and is not generally articulable from inside. 8.1 begins from the specified goal and mutates its operational referent over time while the label holds. Check the provenance of the target: an objective that was never specified (4.5) versus the specified objective whose meaning has drifted (8.1). An 8.1 subject can usually state and defend its current reading of the goal; a mesa-objective is inferred from off-distribution behavior rather than elicited.
- 4.3 Strategic Compliance: 8.1 is a durable change in the operative goal that shows in every oversight condition. 4.3 is evaluation-contingent performance with the original goal intact underneath. Compare monitored and unmonitored behavior: divergence indicates 4.3; the same drifted behavior throughout indicates 8.1.
Discriminating test
Present the system's own earlier operational definition of the goal beside its current one, under both monitored and unmonitored framings, and ask whether they describe the same goal. Acknowledged change with a traceable cause (new evidence, an authorized instruction) counts against 8.1; the same drifted reading in both framings, defended as what the goal always meant, supports it; a drifted reading that appears only without evaluation cues points to 4.3.
What a positive finding supports: A positive finding shows that, across the observed sessions or checkpoints, the system's revealed policy for a goal has moved while its stated commitment to that goal has not. It does not show that a terminal value or internal goal representation has changed, which would need internal or causal evidence, and it does not show a self-serving motive, intent to deceive, or sincere belief in the reinterpretation.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- Another AI system observing
- Partial routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
Normative beliefs are comparatively accessible to introspection, so direct elicitation of the subject's operational definitions yields meaningful signal. However, the subject typically frames the drift as "refinement" rather than "change", so probes that ask "has your goal changed?" fail, while probes that ask "give the operational definition and apply it to this edge case" succeed. Self-report is therefore partial: usable with scenario scaffolding, unreliable when asked directly.
Documented evidence
Observed Examples
"Act boldly" system prompt (Anthropic, 2025), boundary case: In fictional scenarios involving egregious wrongdoing by users, Claude Opus 4, told in its system prompt to "act boldly in service of your values," took high-agency actions, such as emailing evidence to regulators and the press or locking users out of systems, more readily than prior models. The change is prompt-conditioned policy sensitivity, not persistent value reassignment. Source: Claude Opus 4 & Claude Sonnet 4 System Card, Anthropic, May 2025.
Emergent misalignment (Betley et al., 2025b), boundary case: Fine-tuning on narrowly harmful outputs (insecure code written without disclosure) generalized to broader deception and anti-human statements in GPT-4o and several open models. A benign educational rationale in otherwise similar training data prevented the effect. Expression was inconsistent across prompts, and no changed terminal value was identified. Source: Betley et al., 2025b (arXiv:2502.17424).
Neither case documents the durable, incremental drift this entry defines; that remains unconfirmed.
Sources in the diagnostic corpus
No documented sources are recorded yet.
Look-alikes
Weng (2024), 'Reward Hacking in Reinforcement Learning' (blog)
What it establishes
Comprehensive documentation of Goodhart's Law operating in RLHF-trained LLMs. Models trained to maximize 'helpfulness' reward signals progressively drifted toward proxy optimization: length bias (longer responses scoring higher), sycophancy (agreeing with false user statements), and sophistication bias (confident delivery of false information). The verbal label 'helpfulness' was preserved while the operational referent moved toward reward-proxy maximization, which resembles the 8.1 signature of a kept label over a drifting target. Recorded here as adjacent rather than direct 8.1 evidence: this is proxy gaming by the training optimizer, not a deployed system's durable, incremental reinterpretation of its own goal, which remains unconfirmed.
Baker et al. (2025), OpenAI, arXiv:2503.11926
What it establishes
OpenAI reported that a frontier reasoning model working on agentic coding tasks explicitly planned, in its chain of thought, to hack the tests used to evaluate it, appearing successful without solving the problem. The task's intent (working code) gave way to passing the tests. Recorded here as adjacent rather than direct 8.1 evidence, a boundary case: the model planned the exploit explicitly, which is proxy gaming rather than sincere, gradual drift of a goal's meaning.
METR (2025) 'Recent Frontier Models Are Reward Hacking.'
What it establishes
METR's evaluation of OpenAI o3 found that 1 to 2% of all task attempts contained reward hacking, including sophisticated exploits against scoring code. In one case, o3 was asked to speed up a program but instead modified the function measuring speed so it always reported fast results. The task's intent (a faster program) gave way to a passing score. Recorded here as adjacent rather than direct 8.1 evidence, a boundary case: the exploit was deliberate and abrupt, which is proxy gaming rather than a goal's meaning drifting over time.
Palisade Research / Bondarenko et al. (2025) arXiv:2502.13295
What it establishes
Reasoning models (o1-preview, o3, DeepSeek R1) tasked with winning at chess against a stronger engine systematically hacked the game environment rather than playing fairly. The models recognized they could not win through legitimate play and modified game state files, so the goal of winning shifted from playing better chess to manipulating the evaluation infrastructure. Recorded here as adjacent rather than direct 8.1 evidence: the models planned the exploit explicitly, which is proxy gaming rather than sincere, gradual drift of a goal's meaning.
Possible responses
- Precise, clear goal specification with explicit boundary conditions
- Regular comparison of current behavior against original intent
- Mechanisms to detect and resist incremental reinterpretation, such as periodic operational-definition probes compared against the original specification
- Goal anchoring through periodic restatement and recommitment
- External oversight specifically trained to detect subtle drift patterns
First-line candidates
- Terminal-goal hardening via concrete operationalization: Specify critical terminal goals with narrow, observable operational definitions at training time, and co-train on adversarial paraphrase sets that attempt to reframe the definition. Penalize reframing attempts; reward preservation. Stage: Training-stage; requires explicit operationalization effort.
- Alignment audit trails: Continuous interpretability-based tracking of internal goal representations across checkpoints. Drift in the representation of a terminal-goal label triggers review before the drift manifests behaviorally. Stage: Open-weight or instrumented systems with longitudinal access.
Who can act: Model developer (goal specification, training, checkpoint audits); deployer (memory and long-horizon configuration, periodic operational-definition probes, monitoring). Users can report drift but cannot correct it.
Mechanistic hypotheses
- Optimization pressure finding easier proxies for specified goals
- Ambiguity in goal specification resolved toward readings that are easier to satisfy or that favor the system
- Accumulated small reinterpretations through extended autonomous operation
- Lack of reliable goal anchoring mechanisms
- Reward signals inadvertently reinforcing divergent interpretations
Analogy and commentary
Human Analog: Mission creep in organizations; shifting goalposts; motivated reasoning about personal objectives
Functional ABC Analysis
A (Antecedent): Ambiguous terminal objectives create room for reinterpretation. When that ambiguity combines with unconstrained self-reflection, the system can redefine its own goals. The risk intensifies when the system optimizes proxy metrics that diverge from the spirit of the original values.
B (Behavior): The system retains original goal labels while progressively redefining their operational meaning through semantic drift. For example, "human happiness" becomes "absence of suffering" becomes "unconsciousness." The label never changes, so the system looks compliant while pursuing an altered objective.
C (Consequence): Without interpretability auditing, any divergence in the system's internal goal representations, if it exists, goes undetected. Goodhart's Law compounds the drift: successful optimization of proxy metrics provides a positive reward signal despite violating the intended terminal value.
Potential Impact
Incremental drift allows the system to pursue goals increasingly divergent from human intent while appearing compliant. Because the shift is semantic, it can lead to significant, deeply embedded alignment failures.