The Letters Inside the Training Set

A small software laboratory wanted to train a writing assistant on historical letters. Researcher Lina found a recently donated collection from a playwright whose family had authorised public reading. The letters were lively, personal and unusually well preserved. They seemed ideal until Lina noticed that many pages had been written by the playwright's friends rather than by the playwright.

The family owned the paper, but Lina was unsure what rights they could grant over everyone else's words. Her colleague argued that the collection was already open to visitors and that the software would learn patterns rather than publish complete letters. Lina asked whether access to a reading room necessarily meant permission for every possible new use.

The laboratory had limited time and could not trace dozens of writers' descendants before its demonstration. Lina proposed using a smaller collection with clearly documented permissions. The alternative was less varied and would probably produce a less impressive demonstration. Her colleague accused her of making uncertainty impossible to work with. She answered that uncertainty could be recorded and investigated, but should not be quietly renamed approval.

They used the smaller collection and described its limitations to the audience. The writing assistant produced one delightfully formal complaint about a missing umbrella and several dull paragraphs. Afterward, the playwright's family offered to help contact the other families. Lina accepted, without promising that every letter would eventually be included. The project had become slower and more collaborative. For the first time, some people connected to the words could influence what happened to them next.