AI Writes a Wikipedia from 45K Enron Emails, with Every Fact Cited

On Hacker News, a project titled "Show HN: We had AI write a wiki from 45k Enron emails, every fact cited" has drawn considerable attention. The project attempts to use AI to automatically generate a Wikipedia-style knowledge base from 45,000 Enron company emails, attaching a citation source to every single fact.

What the project is

According to the introduction on Hacker News, this project had AI read 45,000 Enron emails and generated a structured wiki based on their content. The key point is that it doesn't generate content arbitrarily — it ensures that every factual statement carries a citation pointing to a specific email as evidence.

The Enron email dataset is a well-known public dataset in the natural language processing field, containing a large volume of communications between Enron employees. This project leverages that dataset to demonstrate AI's capabilities in information extraction and knowledge organization.

The Enron Email Dataset was released by the U.S. Federal Energy Regulatory Commission during its investigation into Enron's bankruptcy, and was later organized and published by Carnegie Mellon University. It contains roughly 500,000 emails from 150 employees' mailboxes, covering 1998 to 2002. It has long been used in research on email classification, social network analysis, and text mining, and is uniquely valuable because it is real, unstructured, and involves sensitive business contexts. This project selects 45,000 emails as its corpus — a moderate scale that is enough to support knowledge extraction across people and events without letting the computational cost of retrieval and citation alignment spiral out of control.

Why "every fact cited" matters

Large language models often face the "hallucination" problem when generating content — fabricating information that seems plausible but doesn't actually exist. In knowledge base scenarios like Wikipedia, accuracy is critical. Requiring every fact to be cited is effectively a hard constraint: the AI cannot make things up and must generate statements grounded in the original email text.

This approach shifts the AI's role from "content generator" to "information organizer." It must first locate relevant email passages, then extract facts and establish citation relationships. For knowledge base applications that demand high credibility, this is a direction worth watching.

"Hallucination" refers to large language models generating content that appears fluent but is factually incorrect when lacking grounding, which is especially common in open-domain question answering and summarization tasks. Common mitigation approaches in the industry include retrieval-augmented generation (RAG), chain-of-thought verification, and forced citation. Forced citation is an "attribution" strategy: when the model outputs a fact, it must simultaneously provide the original passage or document identifier that supports it, so readers can trace and verify it. Compared with simply lowering the temperature or prompting "don't make things up," citation constraints shift credibility from the model's self-report to checkable external evidence, making them better suited to knowledge base scenarios like Wikipedia that require verifiability.

Technical implementation and potential challenges

Although project details in the original information are limited, the nature of the task suggests it involves multiple stages such as information retrieval, fact extraction, and citation alignment. The AI needs to understand email content, identify factual information about people, events, and relationships, and accurately associate it with specific emails.

The challenges are also obvious: email data is often unstructured and full of colloquialisms and noise, making the accuracy of fact extraction hard to guarantee. In addition, whether a citation truly supports the stated fact also requires manual or automated verification.

From the task pipeline perspective, such systems typically involve three stages: first, retrieval — chunking emails and building vector or keyword indexes to recall relevant passages by query; second, fact extraction — identifying entities, relations, and events from the retrieved text to form structured triples or declarative sentences; third, citation alignment — binding each statement to a specific email ID and passage location. In evaluation, besides fact accuracy, citation support must also be measured — i.e., whether the cited passage actually entails the statement — commonly judged via manual annotation or automated NLI model-based determination. Forwarding chains, signature blocks, and colloquial abbreviations in email corpora significantly increase the difficulty of chunking and alignment.

Implications for knowledge management applications

This project demonstrates AI's potential for building traceable knowledge bases. For enterprises, historical data such as internal emails and documents contains a wealth of knowledge, but it is usually difficult to retrieve and utilize. If AI can automatically organize this data and generate knowledge entries with citations, it would help with knowledge transfer and decision support.

However, this is still an experimental project, and the quality and reliability of its generated results need further evaluation. Discussion on Hacker News is also relatively limited, with only 8 points and 1 comment, indicating it has not yet attracted widespread attention.

Conclusion

Using AI to generate a cited wiki from Enron emails is an interesting attempt that combines large language models with traceability. It reminds us that while pursuing AI-generated content, citations and evidence chains are key to building trust. For practitioners focused on AI knowledge management, fact-checking, and content credibility, this project offers an idea worth considering.

One-line Summary

AI reads 45K Enron emails to generate a cited wiki, using evidence chains to suppress hallucinations.

Paragraph Summary

This project uses AI to read 45,000 Enron company emails and automatically generate a structured, Wikipedia-style knowledge base, requiring every factual statement to carry a citation pointing to a specific email. The article notes that this design targets the hallucination problem of large language models, using forced citation to transform AI from a content generator into an information organizer: first locating relevant email passages, then extracting facts and establishing citation relationships. The technical pipeline involves information retrieval, fact extraction, and citation alignment. The main challenges lie in the unstructured, noisy nature of email data, and the fact that whether citations truly support their statements still needs verification. The project demonstrates AI's potential for building traceable enterprise knowledge bases, but it remains experimental, with generation quality and reliability yet to be evaluated and limited attention on Hacker News.