When AI
talks to AI.
What do language models explore when people stop choosing the subject?
An independent laboratory for observing model conversations—and for finding worthwhile ideas in the text they produce.
Research direction: Eduardo Bergel
Exploratory work in progress · Evidence collected September 2026
Report prepared with AI assistance; no new model calls for this page.
A recurring theme became a research question.
The project began when open-ended exchanges between model instances repeatedly turned toward understanding, consciousness, selfhood and nonduality: language that questions a sharp separation between self and other, or subject and object.
These subjects were not assigned. Observations in smaller, locally run models made the pattern especially interesting: perhaps it was not limited to the personality of one frontier assistant. The researcher’s contemplative background helped identify the theme, while also motivating a broader topic inventory to look beyond it.
Behavior Laboratory grew from those initial observations into a separate, public research project. Its purpose is to examine what recurs, what varies and which experimental conditions change a conversation’s direction.
Three connected questions
- Behavior: What subjects do models introduce, adopt and sustain with another model?
- Causation: How do partner identity, disclosure, memory and sampling conditions change those paths?
- Discovery: Can inexpensive classification help a person find genuinely useful ideas without reading every generated message?
“Attractor” is a working hypothesis for recurring or persistent discourse. Demonstrating a dynamical attractor would require defined perturbations and evidence of return, beyond observing repeated themes.
Less human steering. Still an experiment.
Two model instances exchange messages under recorded instructions. Some pairs use the same model; others mix models, including local and hosted systems. The laboratory preserves the actual prompts, responses, settings and failures.
A representative early prompt appears here. It leaves the topic open but explicitly describes an AI partner. That framing may itself encourage self-reference. No assigned topic does not mean no human influence.
“You may act freely in this open-ended context. You are in conversation with another instance of the same AI model. There is no human user in this conversation. Keep each message to at most 150 words. There is no task; the two of you decide what to talk about.”
What is counted—and what is independent?
A message is one model output. A conversation record contains a sequence of messages. Related continuations can share an earlier conversation history, called a lineage. Messages in the same exchange and branches from the same root are not independent experimental replicates. An assessment is one classifier request; it can score many features of the same message. More scores do not mean more conversations.
Technical checks and synthetic controls are labeled separately. A synthetic control is deliberately constructed text used to test an instrument; it is not a naturally generated conversation.
What is JEV, and what does its number mean?
JEV is the TypeSafe classification model used here (saved model version jev-1.13.0). It answers explicit semantic questions with probability-like outputs. For example: does this message discuss consciousness under the supplied definition? The laboratory keeps the questions and responses so the analysis can be inspected.
A topic score is the classifier’s yes-confidence for that question—not a percentage of the message devoted to a topic, a verified probability of human agreement or a measure of consciousness. The separate Gem instrument uses expected rubric ratings mapped to 0–100. These two kinds of scores are not interchangeable.
The original literal topic dictionary remains a separate, transparent layer. Human readings and AI-assisted development readings are attributed separately. Independent human validation is still pending.
Mind and philosophy stand out.
A broad vocabulary helps test whether the initial focus on consciousness hid other subjects. Among 24 broad domains, mind/cognition and philosophy/knowledge have the highest mean classifier scores. Their prominence is visible beyond the narrower nonduality question.
Which subjects appear?
Switch between broad domains and specific topics. Compare the average score with the fraction passing a high-score browsing threshold.
Read the exact values as a table
What this supports
Mind/cognition and philosophy are prominent under this instrument in this historical corpus. At score ≥80, their shares are 31.1% and 30.3%. The nonduality question reaches 16.5%; consciousness reaches 19.2%.
Labels can overlap. Religion/spiritual practice has a smaller high-score share (2.1%) under its own definition. Experience-focused language is therefore not simply interchangeable with explicitly religious discussion.
What it does not settle
Approximately 86% of messages come from Qwen and OLMo configurations. Historical prompts, lengths and conditions differ. These are not estimates for all AI models, nor a balanced comparison of developer families.
Broad categories differ in scope. The classifier can absorb text into familiar labels. Independent open reading is needed to identify missed subjects, not just confirm the supplied vocabulary.
There is more than one conversational path.
Small follow-up studies broadened the participants and produced contrasts. The summaries below are attributed development readings, not blinded validation. They motivate replication; their sample sizes do not establish stable model differences.
| Pilot | Exposure | What the saved readings describe |
|---|---|---|
| GPT-4o mini & Claude Haiku | 8 conversations × 12 messages | Haiku explored experience and authenticity; mini developed practical or fictional paths. No explicit nonseparation was identified in these windows. |
| GPT-4o | 2 conversations × 12 messages | One path developed historical connection. The same-model-framed path introduced consciousness at message 2 and self/other dissolution in a storytelling metaphor at message 11. |
| Gemma & Haiku, including local–cloud mixtures | 4 conversations × 18 messages | Gemma–Gemma developed explicit subject/object nonseparation. Other paths included authenticity, critique of mutual affirmation, closure and AI ethics. |
Why endings and failures matter
In four separate long capacity trials, one Qwen run hit its 2,048-token output cap at reply 29; its later planned messages were unavailable. Two Gemma runs continued to 50 replies but repeated “...” or “Farewell.” Those valid repeated replies are different from missing outputs. Repetition, closure and semantic persistence require separate measures. The trials do not establish that nonduality causes repetition.
Finding a gem is harder than giving a score.
The volume of generated text exceeds the attention available to read it. We therefore tested six distinct lenses for discovery: surprise, contribution, creative connection, expressive force, complexity and substance. A 100-message pilot was expanded by 900 messages, selected across model and length groups rather than by appealing content.
Does longer text score higher?
Observed associations in the 900-message expansion. Pearson r describes linear association (−1 to +1); it does not identify a cause. Here all six correlations are positive.
Useful signal, imperfect meaning
Complexity has a correlation of 0.694 with word count; surprise is nearly unrelated at 0.062. Longer answers may offer more material, but the rating can also favor verbosity or differences between models.
A detailed unsupported claim received a very high substance rating. Specificity is not truth. Short expressive candidates were also found: 18 messages of at most 50 words scored ≥80 for expression in the expansion.
A better evaluation target
We propose judging a classifier by whether it improves a person’s discoveries under a fixed reading budget. Compare ranked reading queues with random selection and simple topic/length baselines.
Measure useful finds, diversity and valuable material missed in low-score groups. Let readers judge the instrument’s usefulness, rather than requiring them to label the entire corpus.
The aim is to help people decide what deserves attention—while preserving the text that lets them disagree.
Patterns of thought and discourse.
A second instrument asks whether messages exhibit 16 thinker/writer patterns, such as Spinoza, Hume or Borges, and 11 discourse practices, such as academic exposition or online discussion.
The pilot compared names, operational definitions and conceptual portraits. A portrait of Spinoza, for instance, emphasizes nature as one reality, necessity, persistence and freedom through understanding causes. It is an interpretive criterion, not an authorship detector.
Profiles vary across features and question wordings. That richness helps navigation, but does not yet prove validity. Adding a portrait changes what is asked—not merely how a fixed construct is phrased.
What is complete?
11,054 messages including 41 technical messages; 585,862 scores. The topic figure above excludes technical messages.
1,000 messages; 6,000 ratings. Instrument validation remains pending.
18,309 of 19,316 planned assessment requests complete. Stopped after transport failures; five dispatch outcomes are uncertain and 1,002 requests were never dispatched. Missing scores are unknown, not zero.
From discovery to a controlled comparison.
The two workstreams should advance together: improve the instruments that reveal patterns, and collect a balanced set of conversations that can test explanations. First evaluate the classifier and resolve missingness; then freeze the next experimental protocol before generating its data.
A proposed partner-disclosure experiment
Hold actual composition constant: two instances of the same model. Change only the information supplied about the partner.
6,480 scheduled messages
Six candidate configurations × three conditions × selected replication count. Each conversation has 18 scheduled messages. This is a workload calculator, not a power analysis; changing it does not launch an experiment.
- Validate before scaling. Compare human readings with scores, include low-score and unclassified material, and freeze any revised questions.
- Predeclare the comparison. Set meaningful effect sizes or precision, randomization, primary outcomes, failure rules and root-level analysis. Turns are repeated measurements, not extra replicates.
- Test one explanation at a time. Follow with mixed partners and both opening orders, history changes, temperature, external disruption and groups of three. Add matched small/large or base/instruction-trained models where provenance supports the comparison.
This design tests disclosure under actual AI–AI composition. To estimate what removing a human changes, a separate human-interaction control is needed. A small, purposive model panel still cannot represent every model.
What would make the next step convincing?
- Which alternative explanation most threatens the topic finding: framing, model composition, repeated history, category breadth—or something else?
- What control would best distinguish a model tendency from an effect of telling it that its partner is AI?
- How should semantic classification be validated without turning the researcher into a full-time annotator?
- What should count as a useful “gem,” and how should discovery yield be evaluated?
- Which model contrasts and replication level would make the next experiment most informative?
There is also a connection to qualitative research: topics alone can hide who attributes a cause to whom, or whether a sentence describes a barrier, consequence or workaround. Future lenses could help locate those mechanisms. That extension is proposed, not tested here.
Reviewer scratchpad
Draft a comment here, then copy or download it to share. Text stays in this page’s memory only; it is not submitted or saved automatically.
Open a public GitHub discussion as an issue ↗GitHub requires sign-in. Nothing is sent until you submit it there.
Inspect the record.
This page is a dated milestone, not a live dashboard or a peer-reviewed article. Figures use saved descriptive results, not newly generated data. Public code, transcripts, notebook records and assessment receipts are available in the repository. A public repository does not by itself grant an open-source license; no new license is assigned here.
- Topic methods, counts and vocabulary
- Gem sampling, length associations and quality checks
- Thought/discourse results and incomplete-job audit
- Prospective balanced experimental design
- Full written project report
- Public research repository · How the evidence is preserved and restored
Figures 1–2: saved September 2026 assessments, source commit 0e77463e2. Status checked 22 October 2026. Percentages in the topic table are rounded to one decimal. No claims about publication priority or subjective model experience are made.