What happened
On March 30, 2026, Trip Venturella released Mr. Chatterbox, a language model trained entirely on out-of-copyright Victorian-era text from the British Library — 28,035 books published between 1837 and 1899, totalling roughly 2.93 billion tokens. The model has 340 million parameters, roughly the size of GPT-2-Medium.
Simon Willison tried it and was honest: "it's pretty terrible. Talking with it feels more like chatting with a Markov chain." He'd been waiting for a model like this for years. He was also visibly underwhelmed by the result.
Both reactions are correct. And the thing this project actually demonstrates isn't visible in the output quality.
Why it's weak — and why that's a known variable
The Chinchilla paper (2022) established an empirical rule: train with roughly 20 tokens per parameter to get compute-optimal performance. For a 340M parameter model that's approximately 7 billion tokens. Mr. Chatterbox used 2.93 billion — less than half.
This isn't a fundamental failure of the provenance-verified approach. It's an engineering constraint: the British Library corpus is finite, and you can't retroactively publish more Victorian novels. But the design pattern that made this corpus usable isn't tied to the British Library specifically. It applies to any institution that holds a large, coherent collection and has the legal standing to authorize it for model training. National archives, academic publishers with long backlists, open-access repositories, standards bodies, government document collections — all of these are potential institutional authorization partners.
The gap between the corpus Venturella had access to and the corpus you'd need to train a useful modern model is a real gap. It's not a fundamental one. It's an institutional partnership and procurement problem.
The attribution wall and the bypass
A question came up in Willison's Bluesky thread: could you train on a single open source license and have the outputs inherit that license? Willison's answer was direct — most open source licenses include attribution requirements, and you'd have to credit every one of the millions of contributors whose code appears in your training corpus. That's not a compliance problem you can solve with a better model card. It's a scale collapse.
Mr. Chatterbox bypasses this entirely. The British Library authorized the corpus at the collection level. One institution, one decision, one clear chain of title. There are no individual contributors to track down or credit. The institution that holds the collection is the contracting party, and the individual authors — deceased for over a century — are not.
This is the design pattern that works at scale: institutional corpus authorization, not license inheritance from individual contributors. The challenge is building institutional partnerships and creating standardized frameworks that let institutions grant training rights without bespoke legal negotiation each time.
What verified provenance actually requires
The AI training data debate often presents a false binary: either scrape the web and accept legal uncertainty, or negotiate with individual rights holders one at a time and accept that it doesn't scale. Mr. Chatterbox is a working example of a third path:
- Out-of-copyright corpora with institutional curation — British Library, Project Gutenberg, Internet Archive, government archives
- Institutional license grants from entities that hold and control large, coherent collections and can speak for them legally
- Domain-specific institutional partnerships with universities, standards bodies, medical publishers, legal databases
None of these options produces the same scale or diversity as the open web. They produce something the open web can't: a complete, auditable provenance chain from data origin to model output.
That's increasingly what enterprise procurement, regulated-industry deployment, and government contracts are going to require. The EU AI Act's transparency provisions include training data documentation requirements for high-risk systems. US federal procurement is moving toward auditable AI supply chains. Regulated industries — healthcare, finance, legal — are already asking questions that "trained on internet-scale data" doesn't answer.
The gap that still needs closing
The current state: verified provenance is achievable at small scale with out-of-copyright text. The resulting models are weak by capability standards. The British Library corpus alone isn't enough.
What would make this path serious at the scale markets actually need:
- Multi-institutional partnerships that aggregate verified collections to hit Chinchilla-optimal token counts for useful model sizes
- Standardized corpus authorization frameworks so institutions can grant training rights without bespoke legal negotiation
- Machine-readable, auditable training data provenance in model cards — not just "trained on public data" or "licensed content"
Mr. Chatterbox doesn't close any of these gaps. It demonstrates the proof of concept and makes concrete what the engineering and institutional work actually is. That's a real contribution from a 2GB weekend project built by one person on out-of-copyright books.
The model is weak. The lesson is not.
Mr. Chatterbox was released by Trip Venturella on March 30, 2026. Simon Willison's writeup and llm-mrchatterbox plugin are at simonwillison.net. The model runs locally via llm and on HuggingFace Spaces.