About The Chat Archive
What this site is for, who it serves, and the rules it runs by.
Purpose
The Chat Archive is a free, open-source archive of human–AI conversations, published for research. It exists so that researchers, educators and the public can study how people actually use AI systems: what they ask, how models answer, where models fail, and how users push back.
What we do
- Collect with consent. People contribute their own conversations. Every contribution goes through a privacy scan, the contributor’s explicit consent and human moderation before it is published.
- Mirror open research datasets. We republish conversations from public research datasets only when their licenses clearly allow redistribution. Each source keeps its own license and attribution and ships in its own release partition. Datasets that don’t allow redistribution appear only as catalog entries that link to the source.
- Publish for study. Anyone can browse, search and compare conversations, and download immutable, versioned releases (JSONL and Parquet) with checksums and a removal changelog.
What we don’t do
- We don’t sell, license or paywall conversations. Reading and downloading are free.
- We don’t publish upstream IP, location or request-header fields, account identifiers or contact details.
- We don’t bypass gated access or ignore a dataset’s terms.
- We don’t present the archive as representative of all AI use, or model comparisons as a benchmark.
Sources and attribution
- Community contributions: CC BY 4.0, attributed as the contributor chooses (name, pseudonym or anonymous).
- WildChat-4.8M (Allen Institute for AI): ODC-BY 1.0. This archive contains information from WildChat-4.8M, made available under the ODC Attribution License. Zhao et al., WildChat: 1M ChatGPT Interaction Logs in the Wild (ICLR 2024).
- OpenAssistant Conversations (oasst2): Apache-2.0. Köpf et al., 2023. Assistant turns are human-written demonstrations, not model outputs.
- Anthropic HH-RLHF (helpful splits only): MIT. Bai et al., 2022.
- Arena human preference 55k (LMArena): Apache-2.0.
- NVIDIA HelpSteer2: CC BY 4.0.
Model outputs may also be subject to their providers’ terms of use. Every conversation page shows its source and license. We add no restrictions beyond those of each source’s license.
Privacy and removal
Automated privacy detection is imperfect. If you appear in a conversation, or a conversation is yours and you want it gone, use “Report” on its page with “takedown request” checked. You don’t need an account. We also mirror removal requests made to upstream datasets. Removed conversations are taken off the site, search and the API, and excluded from all future releases. See Privacy & removal for details and limits.
Open source
The software is licensed under Apache-2.0. That license covers the code, not the data.