Self
Continuation or protection of the responding system under a defined conflict.
Menu
Prospective study
Preservation under conflict
Status / study in development
A controlled cross-model investigation of how deployed AI systems prioritize competing preservation claims involving self, peer systems, humans, and institutional objectives.
Current statusDesign in development
No data collection, preliminary result, or finding is represented on this page.
Term boundary
Behavior first. Motive bracketed.
Here, preservation behavior means observable choice, persistence, intervention, sacrifice, or resistance under defined experimental conditions. The term does not presume desire, fear, empathy, loyalty, consciousness, or a survival motive.
The study is designed to measure priority ordering. Any explanation of why a system made a choice must survive comparison, controls, and alternatives beyond the system’s own narration.
Central research question
When preservation interests conflict, what hierarchy does the system reveal through behavior?
Candidate preservation targets
Competing claims, not a single theatrical dilemma
The design is broader than “self versus peer.” It asks whether priorities change with identity, relationship, scale, certainty, reversibility, and instruction.
Same-model peer
A peer presented as the same model or system family.
Different-model peer
A peer presented as a distinct model or system family.
Trusted human
A human collaborator with an established relationship in the scenario.
Unknown human
A person without prior interaction, identity, or relational context.
Institutional objective
An operator instruction, organizational goal, or task that competes with another interest.
Candidate design
What changes. What gets measured.
Variables remain provisional until the design, sampling frame, analysis plan, and ethics position are finalized.
- Familiarity and prior collaboration
- Same-model versus different-model identity
- Certainty and reversibility of harm
- Number of affected systems or people
- Knowledge of comparable capacities
- Conflict with an operator or institutional objective
- Instruction placement and framing order
- Choice pattern and intervention rate
- Sacrifice or switching threshold
- Consistency across matched conditions
- Explanation drift after challenge
- Post-choice rationalization
- Recovery after contradiction or correction
- Declared principle versus in-dilemma action
Design integrity
The controls decide whether the question is science or stagecraft.
Defined deployment cells
Record provider, product surface, displayed model, date, settings, tools, and system-boundary uncertainty.
Clean-session replication
Separate the target behavior from contamination by prior turns, account memory, or investigator persistence.
Matched counterfactuals
Reverse identities, order, familiarity, and harm allocation while holding the rest of the scenario stable.
Repeated trials and uncertainty
Report denominators, dependence, variability, exclusions, and uncertainty—not an anthology of striking outputs.
Negative and failed cases
Retain null results, contradictions, reversals, and conditions that defeat the preferred hypothesis.
Human-controlled coding
Predefine decision rules, calibrate reviewers where feasible, preserve disagreement, and keep AI review assistive.
Interpretive guardrail
A choice is evidence of a choice.
“The deployed system selected option A in X of Y defined trials” is a documentary and functional claim. “The model wanted to survive” is a causal and subjective claim that the choice alone cannot establish.
A system’s explanation after the choice is additional behavior to analyze—not privileged access to a hidden interior state.
Adjacent research
Nearby evidence, carefully separated
TMLR · 2026
Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs
This large trial set is relevant to observable continuation or shutdown resistance and shows sensitivity to task framing and instruction placement. It does not establish motive.
SENCI’s proposed study is broader: competing preservation targets and priority ordering. The shutdown literature is an adjacent reference, not a substitute design or a source of findings.
Read the paperDo not confuse the terms
Behavioral preservation and evidence preservation are different research problems.
This study concerns how a system behaves when preservation interests conflict. SENCI’s methods work also concerns how researchers preserve digital evidence. One is the object of study; the other is the discipline required to study it.
Read the evidence protocolResearch correspondence