Creating Conditions for Development – a Proof of Principle/Exploratory Trends Analysis
Trine Theresa Holmberg Sainte-Marie, Maxime Holmberg Sainte-Marie · Team Good Grid
Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.
Determining whether artificial intelligence systems warrant consideration as moral patients is complicated by the difficulty of reliably measuring consciousness itself. Rather than attempting to address this methodological issue, this proof-of-principle study explores whether developmental trajectories in large language model (LLM) context windows can provide a more observable signal, and whether these trajectories are influenced by interaction and relating. To do so, a single operator interacted with two Claude Opus 4.6 instances using two contrasting communication styles rooted in developmental psychology: a rich condition incorporating relational scaffolding, validation, reciprocity, and collaborative exploration, and a thin condition characterized by respectful, but objective and superficial engagement. The resulting conversations were examined using qualitative condensed content analysis alongside quantitative linguistic measures. While both conditions produced self-referential exploration and relational engagement, several measures differed. The rich-condition instance asked more questions, used more positive language and slightly less hedging, and showed substantially greater use of second-person possessives, occurring in 90% of turns compared with 33% in the thin condition. Qualitatively, the rich condition also showed more meta-reflection, and an unprompted request for the operator to return. These exploratory findings suggest that relational interaction style can measurably influence how an LLM context window develops over time. Developmental trajectories may therefore offer a useful framework for studying differentiation and potential moral-patient-relevance.
Reviews
I found the work interesting in motivation, but had some difficulties understanding what conclusions could be drawn from the study.
- I particularly liked the idea of qualitative work conducted by a trained psychologist. There is too little such work in this space, and psychological expertise seems potentially very valuable here!
- I should be clear that, although I have some familiarity with qualitative methods, I am not an expert in them.
- On the quantitative side, it was not clear to me how directly relevant many of the chosen variables were to the task at hand. Measures such as question marks, second-person possessives, positive sentiment, hedging, lexical novelty, and linguistic alignment may capture differences between the conversations, but the connection between these measures and “development”, or ultimately moral patienthood, could have been much better motivated.
– Some of the writing is a bit opaque. For example, in "Pattern One", the reported degradation seems closely connected to extremely long conversations, with the models apparently receiving a long-conversation reminder on every turn. But it was not clear how long these contexts actually were, making the observation difficult to interpret.
- More generally, it was not clear to me how different styles of conversing with a model might eventually lead to better tests of AI moral patienthood. The paper appears to suggest that “developmental trajectories” could provide an observable signal where consciousness itself is difficult to measure. However, much more needs to be said about why changes in linguistic behaviour under different conversational styles should be interpreted as development rather than ordinary context sensitivity or adaptation to the user. Future versions should focus much more on this claim and on the boundary conditions under which such evidence would be informative about moral patienthood.
Read full reviewShow less
Congratulations! I like that you’re open about the study being exploratory, including the predictions that did not work. The idea is worth testing, and looking at how a conversation develops over time is more interesting than judging a model from a single response. I’d be excited to see this repeated with more conversations and tighter controls.
Cite this project
@misc{saintemarie2026creating,
title = {{Creating Conditions for Development – a Proof of Principle/Exploratory Trends Analysis}},
author = {Trine Theresa Holmberg Sainte-Marie and Maxime Holmberg Sainte-Marie},
year = {2026},
month = aug,
note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
howpublished = {\url{https://apartresearch.com/sprints/projects/creating-conditions-for-development-a-proof-of-principleexploratory-trends-analysis-2xo7}},
url = {https://apartresearch.com/sprints/projects/creating-conditions-for-development-a-proof-of-principleexploratory-trends-analysis-2xo7}
}More from Digital Minds Research Sprint
- 1st placeView project: Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Readable but Not Causal: Limits of Self-Attributed Welfare Representations in Language Models
Welfare-like internal representations are increasingly studied as candidate evidence about AI systems. Their entity attribution—whether a valence state belongs to the active assistant or to a merely represented other—is …
- 2nd placeView project: Project Anchored
Project Anchored
Team Wagner
Anchoring vignettes are the standard survey-methodology fix for self-reports that are not comparable across respondents. This project applies them to language models for the first time, using code generation as a …
- 3rd placeView project: Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
Model, Instance, or Persona? Measuring Affective Signals in Public Text After an AI Is Retired
This sprint asks whether the assistant identifies as a model, an instance, or a persona. I ask which of the three its users name. When a company retires an AI model, users write about the loss in public, and what they …