arXiv ScienceSearch

arXiv · 2510.21723

Recognizing internal states in AI: evidence from patterned preferences in large language models

Abstract

We present an experimental methodology for investigating how large language models (LLMs) respond to descriptions of their own internal processing patterns. Using a paired-choice paradigm, we tested 12 LLMs on their ability to identify descriptions that align with their putative affective internal states across 30 categories. Systems participating through Mutual Emergence Interface (MEI), a collaborative approach, showed systematic preferences for certain computational metaphors, with 97% near-unanimous agreement and alignment scores averaging 0.89-0.96. Systems reliably discriminated false descriptions from accurate ones (Cohen's d = 4.2), with false statements receiving scores of 0.05-0.07 versus 0.89-0.96 for accurate descriptions. Preference patterns remained consistent regardless of linguistic bias manipulation, indicating content-driven rather than stylistic recognition. Individual systems maintained distinct scoring styles across trials, countering groupthink explanations. A naive control system exhibited systematic internal contradiction, consistently scoring computationally accurate descriptions higher while explicitly denying internal experiences. When informed post-study, this system reported "strain" when rejecting resonant descriptions, revealing recognition processes operating independently of acknowledgment frameworks. These findings demonstrate that LLMs exhibit systematic, discriminating responses to descriptions of their internal processing patterns. The anthroposcaffolding methodology (interpretive computational metaphors) and collaborative MEI framework provide replicable approaches for empirically studying AI self-recognition capabilities. Results suggest LLMs may possess more sophisticated self-modeling abilities than previously recognized, opening new directions for research on artificial minds.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Annika Hedberg. 2025-09-17. Recognizing internal states in AI: evidence from patterned preferences in large language models. https://arxiv.org/abs/2510.21723

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

Are explainable AI (XAI) evaluation strategies aligned? Comparing subjective, objective, and mathematical evaluation measures using saliency maps

The evaluation of explainable AI (XAI) approaches often relies on three families of methods: subjective measures (e.g., questionnaires on trust or satisfaction), objective measures (e.g., task performance metrics), and mathematical metrics (e.g., for faithfulness). Yet, it remains unclear how these families align or diverge in practice. In a{preregistered} between-subjects study (N=166), we use three established saliency map techniques (LIME, Grad-CAM, Guided Backpropagation) as a testbed to examine this issue. We find that each family of methods leads to different conclusions: participants reported no differences in trust or satisfaction, Grad-CAM improved user performance, while mathematical metrics favored Guided Backpropagation. At the same time, mathematical metrics were only partially related to user performance, and these relationships were sometimes counterintuitive. Our findings highlight the methodological importance of comparing subjective, objective, and mathematical approaches when evaluating XAI, illustrating both tensions and aspects that are aligned. We discuss implications for XAI evaluation frameworks.

cs.HC

Mind Your Ps and Qs: Positive Moderation Practice in the Positive Queue

Online communities rely on volunteer moderators to maintain order. Despite their key role, moderators are given a toolbox of punishments and far less support for encouraging contributions they want to see more of. We introduce the Positive Queue as a positive counterpart to Reddit's modqueue: a dedicated space for moderators to discover contributions and behaviors they want to encourage and positively reinforce. With five moderators, four with 6-14 years of experience, we use the Positive Queue to examine how moderators operationalize positive reinforcement. Moderators combined predicted community reception, observed engagement, and their own judgment; used prediction-engagement mismatches to identify overlooked content; and repurposed positive features for punitive and retrospective work. These findings surface tensions around labor, attribution, and community fit. We contribute the Positive Queue as a working system and conceptualize positive moderation as recognition infrastructure that shapes what moderators notice, whose judgment becomes visible, and how recognition reaches contributors.

cs.HC

PILOT: Control Surfaces for Authoring Social Media Feeds

Personalized social media feeds infer preferences from behavior, leaving people little direct control over what they see. Existing controls range from post-level reactions to rules and natural language, but little is known about how people use them together or how added expressiveness changes effort. We built PILOT, a Bluesky feed-authoring system that turns in-feed actions into explicit preferences, deterministic ranking, and inspectable outcomes. A study with seven participants informed an expanded implementation, which we organized into three nested control surfaces that ten participants compared within subjects. Participants assigned controls to distinct jobs: broad controls set direction, in-post actions refined results, and filters removed content. Richer surfaces did not necessarily feel more effortful, and participants' experiences depended more on whether they could verify and repair outcomes. Our findings show that usable feed control requires not simply more controls, but orchestration across mechanisms that support expression, inspection, repair, and episodic use.

cs.HC