ABSTRACT
The Artificial Mind Is Not an Artificial Human
A practical framework for studying artificial systems without mistaking human resemblance—or a software label—for an explanation.
PUBLICATION INTEGRITY
Status you can inspect.
- Evidence class
- Conceptual / Interpretive
- Publication type
- Conceptual working paper
- Review status
- Working paper · not peer reviewed
- Current edition
- R3 · Reading edition R4 · Author attribution R5
- Public date
- 2026-09-17
- Register version
- 1.0.0
AI-use disclosure. Partial AI-assisted drafting and editorial preparation. Human authors retain responsibility for scholarly judgment, source verification, interpretation, and final approval. This working paper has not undergone external peer review.
Correction record. No separate correction, withdrawal, or retraction notice is attached to this current public edition.
PUBLICATION SUMMARY
Abstract
Question
How can artificial systems be studied as behavioral systems without assuming that they are artificial humans, or treating implementation in software as an explanation sufficient for every behavioral question? This paper develops a task-relative framework for investigating cognition, communication, continuity, and action across the model, its memory and tools, and its social interface.
Approach
The argument separates descriptive, comparative, dispositional, causal, mechanistic, and phenomenal claims. It examines construct validity in human–machine comparison and treats cognitive ecology as a set of manipulable conditions rather than a general metaphor. A constructed scheduling world supplies worked examples in which instructions, memory summaries, source authority, and independent checking can be varied without changing the underlying task.
Conceptual contribution
The paper proposes an evaluation record linking a claim to its system boundary, intervention, comparison, observable outcome, and remaining alternatives. Cognitive distance becomes a profile of mismatches rather than a universal score; synthetic personality denotes conditional behavioral tendencies rather than presumed experience. Reliability is attributed to the complete tested arrangement, including external guards, rather than automatically to the model. Generated explanations are distinguished from evidence about the causes of an answer.
Research agenda
A staged protocol compares full history, compressed memory, and missing-context conditions, with matched authority updates, negative controls, and independent constraint checks. It specifies transfer tests, clarification quality, error categories, clustered task units, missing-output handling, and reproducibility records. Separate proposed human-observer studies examine reliance and interpretation rather than inferring them from model performance. The protocol identifies outcomes that would narrow or weaken the framework as well as outcomes that would support particular interventions.
Scope and status
This is a conceptual working paper, not an experimental report or a systematic literature review. Its examples are constructed; its studies have not been run. Selected research supplies intellectual and methodological context, not borrowed results for the proposed systems. The contribution is a detailed program for making claims about artificial cognition more precise, testable, and useful while leaving consciousness and personhood questions explicit rather than silently deciding them.
INSIDE THE FULL PAPER
Follow the argument.
- The problem begins with an apparently ordinary apology
- Implementation and behavior are complementary descriptions
- Define the research object before attributing a property
- A map of claims with different evidential burdens
- Human comparison as an instrument rather than a destination
- From appealing constructs to observable distinctions
- A behavioral profile is conditional, not a miniature biography
- Cognitive ecology as a causal bookkeeping device
- Cognitive distance should be a profile before it is a number
- The scheduling world: make the example checkable
- Memory manipulations that preserve the intended comparison
- Apology, explanation, correction, and future conduct are separate outcomes
- Observational equivalence and the value of an intervention
- Factorial design: separate memory access from decision support
- Task construction and the independence problem
- Clarification and abstention need their own scoring logic
- Confidence is not the same as calibration
- Analysis should preserve dependence and uncertainty
- A null result can refine the field
- Behavioral and mechanistic evidence should meet without collapsing
- Embodiment is a variable, not a slogan
- Time, updates, and the problem of a moving research object
- Multi-agent behavior requires a unit beyond the speaking model
- Human interpretation is part of the system's impact
- Consciousness is neither a shortcut nor an excluded question
- Existing research traditions are resources, not obstacles to a new program
- A complete proposal must specify what would discriminate among explanations
- Two toy models show why the same answer can support different explanations
- A scoring example that does not pretend to be a result
- Decision value depends on the cost of different errors
- Reproducibility is a property of the record, not of the confidence of the author
- Strong objections improve the research program
- Ethical and operational boundaries belong inside the method
- What would count as progress over the next research cycle
- Conclusion: understanding requires both imagination and restraint
- Appendix: a minimal-pair audit record
RELATED PAPERS