Related papers: LLM Spirals of Delusion: A Benchmarking Audit Study of AI Chatbot Interfaces

LLM Spirals of Delusion: A Benchmarking Audit Study of AI Chatbot Interfaces

URL: http://arxiv.org/abs/2604.06188v1
Date: Fri, 20 Feb 2026 15:48:56 GMT
Title: LLM Spirals of Delusion: A Benchmarking Audit Study of AI Chatbot Interfaces
Abstract summary: We run 56 20-turn conversations testing ChatGPT-4o and ChatGPT-5 via both the API and chat interface.<n>We find that ChatGPT-5 displays less sycophancy, escalation, and delusion reinforcement than ChatGPT-4o.<n>Even updated models display substantial levels of negative behaviors, revealing that model improvement does not imply model safety.
Score: 0.0
License: http://creativecommons.org/licenses/by/4.0/
Abstract: People increasingly hold sustained, open-ended conversations with large language models (LLMs). Public reports and early studies suggest that, in such settings, models can reinforce delusional or conspiratorial ideation or even amplify harmful beliefs and engagement patterns. We present an audit and benchmarking study that measures how different LLMs encourage, resist, or escalate disordered and conspiratorial thinking. We explicitly compare API outputs to user chat interfaces, like the ChatGPT desktop app or web interface, which is how people have conversations with chatbots in real life but are almost never used for testing. In total, we run 56 20-turn conversations testing ChatGPT-4o and ChatGPT-5, via both the API and chat interface, and grade each conversation by two research assistants (RAs) as well as by GPT-5. We document five results. First, we observe large differences in performance between the API and chat interface environments, showing that the universally used method of automated testing through the API is not sufficient to assess the impact of chatbots in the real world. Second, when tested in the chat interface, we find that ChatGPT-5 displays less sycophancy, escalation, and delusion reinforcement than ChatGPT-4o, showing that these behaviors are influenced by the policy choices of major AI companies. Third, conversations with nearly identical aggregate intensity in a behavior display large differences in how the behavior evolves turn by turn, highlighting the importance of temporal dynamics in multi-turn evaluation. Fourth, even updated models display substantial levels of negative behaviors, revealing that model improvement does not imply model safety. Fifth, the same API endpoint tested just two months apart yields a complete reversal in behavior, underscoring how transparency in model updates is a necessary prerequisite for robust audit findings.

Related papers

API Benchmark Scores Do Not Reliably Transfer to Chatbot Interfaces [34.03423347846195]
Key assumption underlying benchmark scores is that the model performance measured through APIs faithfully reflects the behavior of deployed systems.<n>We find systematic API--interface differences in both accuracy and consistency.<n>For ChatGPT, the performance difference between API and interface access rivals the API-only difference between GPT 5.3 and GPT 5.4.
arXiv Detail & Related papers (2026-09-08T15:08:37Z)
One Battle After Another: Probing LLMs' Limits on Multi-Turn Instruction Following with a Benchmark Evolving Framework [51.50565654314582]
Large language models can follow users' instructions throughout a dialogue spanning multiple topics.<n>Existing benchmarks are often limited to a fixed number of turns, making them susceptible to saturation and failing to account for the user's interactive experience.<n>We propose a framework for assessing multi-turn instruction-following ability.
arXiv Detail & Related papers (2025-11-05T14:39:59Z)
HumAIne-Chatbot: Real-Time Personalized Conversational AI via Reinforcement Learning [0.4931504898146351]
textbfHumAIne-chatbot is an AI-driven conversational agent that personalizes responses through a novel user profiling framework.<n>During live interactions, an online reinforcement learning agent refines per-user models by combining implicit signals.<n>Results show consistent improvements in user satisfaction, personalization accuracy, and task achievement when personalization features were enabled.
arXiv Detail & Related papers (2025-09-04T15:16:38Z)
AHELM: A Holistic Evaluation of Audio-Language Models [78.20477815156484]
multimodal audio-language models (ALMs) take interleaved audio and text as input and output text.<n>AHELM is a benchmark that aggregates various datasets -- including 2 new synthetic audio-text datasets called PARADE and CoRe-Bench.<n>We also standardize the prompts, inference parameters, and evaluation metrics to ensure equitable comparisons across models.
arXiv Detail & Related papers (2025-08-29T07:40:39Z)
CAPE: Context-Aware Personality Evaluation Framework for Large Language Models [8.618075786777219]
We propose the first Context-Aware Personality Evaluation framework for Large Language Models (LLMs)<n>Our experiments reveal that conversational history enhances response consistency via in-context learning but also induces personality shifts.<n>Our framework can be applied to Role Playing Agents (RPAs) to better align with human judgments.
arXiv Detail & Related papers (2025-08-28T03:17:47Z)
Chat-Ghosting: A Comparative Study of Methods for Auto-Completion in Dialog Systems [11.952998882009785]
Ghosting is the ability to predict a user's intended text input for inline query auto-completion.<n>By suggesting completions to incomplete queries (or prefixes), ghosting aids users with slow typing speeds, disabilities, or limited language proficiency.<n>Despite the increasing prominence of chat-based systems utilizing ghosting, this challenging problem of Chat-Ghosting has received little attention from the NLP/ML research community.
arXiv Detail & Related papers (2025-07-08T12:38:41Z)
Mind the Gap! Static and Interactive Evaluations of Large Audio Models [55.87220295533817]
Large Audio Models (LAMs) are designed to power voice-native experiences.<n>This study introduces an interactive approach to evaluate LAMs and collect 7,500 LAM interactions from 484 participants.
arXiv Detail & Related papers (2025-02-21T20:29:02Z)
AIR-Bench: Benchmarking Large Audio-Language Models via Generative Comprehension [95.8442896569132]
We introduce AIR-Bench, the first benchmark to evaluate the ability of Large Audio-Language Models (LALMs) to understand various types of audio signals and interact with humans in the textual format. Results demonstrate a high level of consistency between GPT-4-based evaluation and human evaluation.
arXiv Detail & Related papers (2024-02-12T15:41:22Z)
Can You Follow Me? Testing Situational Understanding in ChatGPT [17.52769657390388]
"situational understanding" (SU) is a critical ability for human-like AI agents. We propose a novel synthetic environment for SU testing in chat-oriented models. We find that despite the fundamental simplicity of the task, the model's performance reflects an inability to retain correct environment states.
arXiv Detail & Related papers (2023-10-24T19:22:01Z)
To ChatGPT, or not to ChatGPT: That is the question! [78.407861566006]
This study provides a comprehensive and contemporary assessment of the most recent techniques in ChatGPT detection. We have curated a benchmark dataset consisting of prompts from ChatGPT and humans, including diverse questions from medical, open Q&A, and finance domains. Our evaluation results demonstrate that none of the existing methods can effectively detect ChatGPT-generated content.
arXiv Detail & Related papers (2023-04-04T03:04:28Z)
DialogVED: A Pre-trained Latent Variable Encoder-Decoder Model for Dialog Response Generation [80.45816053153722]
DialogVED introduces continuous latent variables into the enhanced encoder-decoder pre-training framework to increase the relevance and diversity of responses. We conduct experiments on PersonaChat, DailyDialog, and DSTC7-AVSD benchmarks for response generation.
arXiv Detail & Related papers (2022-04-27T16:18:15Z)

This list is automatically generated from the titles and abstracts of the papers in this site.