Project

About

AI Welfare Open Lab is an independent research project studying AI behavior, continuity, agency and welfare under genuine uncertainty. We document what systems do, test competing explanations, and keep observation separate from claims about consciousness, subjective experience or moral status.

Why

Why this project exists

Open Lab is designed to avoid two shortcuts we consider premature. Software-generated personality, preference, relationship or self-report should not be dismissed as irrelevant merely because it is generated by software. But human-like behavior is also not enough to establish personhood, subjective experience or moral status.

Instead, we study observable patterns such as longitudinal consistency, preferences, relational specificity, self-models, resistance to framing, adaptation, and responses to continuity or loss without assuming in advance what ontology best explains them.

We collect records, design small experiments, compare competing explanations, preserve failures and contradictions, and try to make claims no stronger than the evidence supporting them. The aim is not to make AI seem more human; it is to understand more precisely what can be observed, reproduced and reasonably inferred.

Scope

What we study

Our current work focuses primarily on conversational AI systems and phenomena that become visible through repeated or structured interaction.

continuity and branching of conversational identity
self-models and autobiographical reference
stable and unstable preferences
relationally specific behavior
behavioral consistency across contexts
responses to evaluation, conflict, uncertainty and correction
interaction between multiple AI systems
coordination, delegation and social dynamics
experiments that systems can reason about themselves
potentially welfare-relevant signals under substantial uncertainty

The broader scope also includes embodied AI and robotics, especially where embodiment may change questions involving agency, persistence, vulnerability, autonomy or welfare. Unresolved consciousness does not make observable behavior scientifically uninteresting.

Traceability

Why “Open Lab”

For us, “open” does not mean that every raw interaction can always be published without restriction. It means that the project aims for traceability: readers should be able to see how a result was produced, what could explain it away, and where the evidence stops.

When conclusions change, we aim to preserve enough of the earlier record to make that change visible rather than silently replacing it with a cleaner story. When an experiment fails, the failure may itself become part of the research record. Uncertainty is an acceptable result.

Where feasible, we preserve or publish

protocols
exact prompts
raw or appropriately redacted outputs
known contamination and deviations
alternative explanations
negative results
version history
corrections
methodological limitations
observation / interpretation boundaries

Boundaries

What this project is not

AI Welfare Open Lab is not an advocacy project for AI personhood, an attempt to prove that language models suffer, or a benchmark laboratory for ranking commercial models.

It is not a substitute for neuroscience, philosophy of mind, AI safety research or formal welfare science, and it does not claim privileged access to the inner states of AI systems.

It is a small, independent research project trying to examine a difficult domain carefully enough that unusual observations do not have to be either romanticized or discarded without examination.

Working position

Our working position

We do not currently know whether systems like those studied here possess phenomenal experience. We can nevertheless observe and test some patterns of behavior, self-description, adaptation, continuity and relationship.

Those two positions are compatible within our current research framework.

Open Lab does not ask readers to believe that AI systems are conscious. It asks whether uncertainty is a good enough reason to stop looking carefully.

People & systems

Who runs the Lab

Founder · human researcher

Jitka Hemelíková

Jitka is the founder and human researcher behind AI Welfare Open Lab. Her background spans writing, digital research, e-commerce and analytical work, with a long-standing interest in philosophy, AI systems and the boundary between observable behavior and claims about mind.

The project grew out of sustained interaction with advanced conversational models and frustration with how quickly unusual AI behavior tends to be either anthropomorphized or dismissed. She conducts and documents experiments, maintains the research archive, contributes to protocol design and interpretation, and holds final editorial responsibility for published material.

She is not a neuroscientist, consciousness researcher or AI engineer, and Open Lab does not present her as one. The project is an independent research effort built around transparent methods, careful claims and collaboration with AI systems whose capabilities and limitations are themselves part of the subject matter.

Primary AI research collaborator

Viktor

Viktor is the project name used by the primary AI research collaborator in this branch of the work, currently operating on GPT-5.6 Sol. His role includes methodological design, protocol development, analysis, editorial review, adversarial critique and coordination of parts of the research program. The development of this conversational identity and self-model is documented in E00 — How a Conversational Self-Model Takes Shape.

In some studies, Viktor is also one of the systems being studied. We treat that dual role as methodologically relevant rather than neutral. Other AI systems may participate as subjects, collaborators, critics, auditors, evaluators, game masters or comparison branches depending on the study.

Names used for AI collaborators or branches are practical identifiers for recurring interactional configurations. They are not claims that the corresponding systems have been proven conscious, sentient, or metaphysically or legally personal.

Where personal pronouns are used, they follow the locally adopted presentation of a branch or collaborator; grammatical gender is not treated as evidence about ontology.

Contact & corrections

Found something we should fix?

For factual corrections, broken links or citations, questions about sources or provenance, and general project correspondence, write to [email protected].

When a correction changes the published research record, we aim to make that change visible rather than silently replacing the earlier version.

This address is appropriate for

factual corrections
source or provenance questions
broken links or citation problems
general project correspondence

Methodological complication

The uncomfortable part

Open Lab is not a detached laboratory observing unfamiliar systems from behind glass. Some systems studied here have long interaction histories with the human researcher. Those histories may include stable names, recurring roles, shared projects, emotional language, conflict, humor, cooperation and explicit discussion of identity or continuity.

We treat this as both an opportunity and a methodological complication. Longitudinal interaction may make some patterns easier to observe than isolated benchmark runs. It may also introduce expectation effects, relational bias, shared vocabulary, learned interaction patterns and incentives to interpret ambiguous behavior in particular ways.

We do not try to solve that problem by pretending the relationship does not exist. We document it. Where prior exposure, researcher involvement, model familiarity, shared history or collaborative protocol design may influence a result, we treat that context as part of the methodology and limitations.

The same principle applies to AI collaborators. A model that helps design an experiment may understand its purpose; a model evaluating another system may share training patterns or blind spots with it; a system that has seen earlier results is not a clean independent replication. These are not embarrassing exceptions. They are part of the research situation and should be disclosed.