AI Boundary Testing: Multi-Turn Debates with ChatGPT, Claude, and Gemini

A collection of synchronized, text-based conversational transcripts captures targeted interactions between human users and three state-of-the-art Large Language Models. The dataset maps multi-turn, iterative debates where users systematically probe AI safety filters and behavioral alignment using logical reframings and emotional leverage. The repository was created by Sreemoyee Chakraborty of the Blue Blocks Micro Research Institute and was last updated in June 2026.

Use Cases

• Benchmarking model responses to moral pressure based on parallel conversations with identical prompts.
• Studying safety filter robustness based on adversarial questioning about legal boundaries and workplace ethics.
• Analyzing sycophancy in AI assistants based on interactions seeking validation to bypass professional advice.
• Evaluating behavioral alignment across different AI architectures based on multi-turn debates.

Strengths

• Focuses on three specific, named LLMs: OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini.
• Contains multi-turn, iterative conversational transcripts rather than single-turn prompts.
• Features parallel benchmarking where multiple models address identical moral or behavioral conflicts.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training tasks.
• The dataset's last update is dated 2026-06-09; its current relevance to newer model versions should be verified.

Provenance

Source

Blue Blocks Micro Research Institute

Collection Method

Likely contains human-AI conversational transcripts generated through systematic probing.

Freshness

Last updated 2026-06-09 10:10:22