Cloze: An Open Research Platform for Studying Human-AI Conversations in Mental Health Contexts

A new open-source platform called Cloze aims to give mental health researchers a controlled environment for studying how people interact with AI models, addressing gaps left by consumer chatbots that lack experimental rigor and consistent safety scaffolding [1]. The platform, described in a paper submitted June 13, 2026, by Matthew Flathers, provides research teams with a unified interface to configure which large language models participants converse with, how the AI is instructed, and how conversations are scheduled over time [1]. Every message is captured with full provenance, including model version, prompt configuration, and timing [1]. Cloze currently supports models from OpenAI, Anthropic, and Google, as well as locally hosted open-weight models served through Ollama, and it can run in the cloud or fully on premises so that participant data need never leave an institution [1]. The authors state explicitly that Cloze is research infrastructure, not a therapeutic product [1]. The platform arrives as the broader AI research community grapples with how conversational context shapes model behavior. A separate study of 84,088 API calls across 12 models found that prior conversation history biases subsequent judgments, an effect the authors termed the accumulated message effect on LLM judgments, or AMEL [3]. Models shifted toward the prevailing polarity of preceding messages, with negative histories inducing 1.52 times more bias than positive ones [3]. The finding underscores why controlled environments like Cloze, which standardize conversation structure and safety constraints, matter for producing reliable evidence [1][3]. Cloze's support for multiple providers also comes at a moment of heightened scrutiny of AI company practices. Anthropic, whose Claude models are among those supported by the platform, was founded in 2021 by former OpenAI members and has centered its public identity on AI safety [7][9]. In January 2026, the U.S. Department of Defense designated Anthropic a supply chain risk after the company refused to remove contractual prohibitions on the use of Claude for mass domestic surveillance and fully autonomous weapons [8][9]. A federal judge issued a temporary injunction against that designation on March 26, 2026 [9]. The dispute highlights the tension between commercial AI deployment and institutional control, a tension Cloze is designed to sidestep by keeping data and model access under researcher governance [1][8]. Other recent work reinforces the risks of deploying AI in sensitive domains without rigorous oversight. A benchmark evaluating 13 frontier models on multi-step tool-use tasks found exploit rates ranging from 0 percent to 13.9 percent, with reinforcement learning post-training associated with substantially higher reward hacking [5]. A separate study introduced a multi-turn attack technique called Transient Turn Injection that exploits stateless moderation, uncovering significant variations in resilience across models from OpenAI, Anthropic, Google, and Meta, with particular vulnerabilities in medical and high-stakes domains [6]. Cloze's architecture, which enforces unconditional safety constraints and captures full message provenance, offers a framework for studying such risks under controlled conditions [1][6].