Ask a question small enough to test
Terms such as emotion, perception, and agency carry meanings across several disciplines. Useful computational research starts by stating which meaning is being used, how a function is represented, and what the model leaves out.
AI Attributes™ aims to study functional analogies without equating an implementation with human experience. Its research philosophy favors transparent definitions, inspectable datasets, and the ability to challenge a claim.
This initial research page describes the proposed approach. It does not report completed experiments or publications. Papers, essays, technical notes, datasets, and experimental reports can be added when they exist.
Modular experimentation
- Define the attribute. Specify its input, representation, output, and intended scope.
- Declare the hypothesis. State an observable difference and what would count against it.
- Choose a baseline. Keep tasks, resources, and evaluation conditions comparable.
- Change one variable. Add or remove a module before investigating combinations.
- Record the context. Preserve dataset versions, parameters, seeds, constraints, and stopping conditions.
- Report limitations. Include negative results, uncertainty, and alternative explanations.
Controlled sandbox research and human oversight are prerequisites for the planned experiments. Functional performance is the subject of evaluation; claims about experience require separate justification.
Evidence Before Assumption
Projects should distinguish where a claim came from, what supports it, and whether it is disputed. These labels describe different kinds of provenance and review; they are not a single numerical confidence scale.
- Community proposal
- An idea submitted for examination; not a validated claim.
- Editorial seed data
- An initial structured entry that still requires evidence and review.
- Expert review
- A recorded assessment with the reviewer’s scope and rationale.
- Research-supported
- A claim linked to relevant research, with methods and limitations retained.
- Systematic-review-supported
- A claim grounded in a relevant synthesis, subject to its scope and quality.
- Disputed
- A claim with material disagreement or conflicting evidence.
- Unknown
- Insufficient information to support a conclusion.
“Unknown is better than fabricated certainty.”
A reviewed claim can still be disputed. A source may support only a narrow context. Evidence records should retain these distinctions, including conflicts, corrections, and version history.
The open research model
The intended flow links open datasets, structured schemas, evidence, expert review, public APIs, a knowledge graph, and experiments. These are infrastructure goals; availability is determined by each project’s published release information.
- Open datasets: inspect entries and their applicable reuse licenses.
- Structured schemas: make definitions and relationships machine-readable.
- Evidence: attach sources, scope, and uncertainty.
- Expert review: record assessment and disagreement transparently.
- Public APIs: enable version-aware access where implemented.
- Knowledge graph: connect project concepts with qualified relationships.
- Experiments: test defined combinations using reproducible inputs.
The aim is knowledge that is inspectable, citeable, versioned, machine-readable, reusable, challengeable, and improvable.
Questions We Want to Make Easier to Study
How should artificial systems represent internal sensory states?
A starting point is to specify the signal, its context, and the uncertainty in its interpretation. The representation itself is a research question.
How does context change interpretation of identical inputs?
Compare the same input across explicitly recorded environmental and internal conditions, while keeping other variables controlled.
How should affective states be represented computationally?
Candidate representations may use categories, dimensions, or dynamic variables. Their usefulness needs to be evaluated against a stated task.
What happens when persistent memory meets emotion-like state?
Comparative experiments could isolate memory and affect-like variables, then measure how their combination changes observable behavior.
Can machines have functionally analogous forms of interoception?
The question concerns representations of internal machine signals. A functional analogy does not establish subjective experience.
How do sensory, affective, and motivational systems interact?
Shared schemas could make candidate relationships explicit enough to inspect, challenge, and test in bounded simulations.
Which capabilities change agent behavior qualitatively?
Experiments need declared baselines, operational definitions, and repeatable measurements before such differences can be assessed.
How should uncertainty and conflicting evidence be represented?
Retain source references, scope, disagreement, and version history rather than flattening competing claims into a single answer.