I’m on my way back from Code4Lib, which has really become the go-to library technology conference (in this country). C4L 2025 was held at Princeton University. Some notes:

We build scaffolding: locally built and maintained support structures around the Big Systems we buy, to make them work the way we want. We may start out thinking the scaffolding is temporary, but if it works, it isn’t going away. The “we” in these statements means the people who are hands-on with those systems day in and day out, and who may have ideas very different from the systems implementors and administrators.

(Psst: people who reported good impressions of a new system and boosted morale six months after a migration often saw that taper off over the next year. Those with bad impressions and lower morale? Yeah, that sticks around. Just saying.)

Hey, guess what about those noble plans to change outdated or, um, problematic subject headings in your catalog, if you’re an Alma library and your discovery layer pulls in Community Zone records, plus HathiTrust records and records from who knows where else? Yeah.

Everyone is fighting bots. The [grudging] consensus is that we should not ban them outright, but only when/as they affect system performance. That just leads to the awkward question of How? Everyone is seeing a new generation of bots that A) seem to be AI-related, B) really do not follow the long-standing rules for good bots, and C) are using distributed networks of many IP addresses. Fortunately, there is a growing toolkit to tackle the problem.

Metadata is important and powerful (and emotional). And a lot of metadata still runs through MARC. If you thought that was going away anytime soon, take a look at your QWERTY keyboard.

Many of the power presentations this year were from people with a job title of “Metadata Librarian.”

By a rough estimate, “we used this cool AI tool” and “we evaluated this AI tool and did something else” were running neck and neck. The best approaches came down to, “we put a lot of effort into getting the cool AI tool to work as well as possible, and then had humans check its work.” The result: doing OCR on a 19th century diary with beautifully unreadable Spencerian script and getting back a very low 4% error rate.

Tiny mantras that suffused the conference:

• Be kind and communicate (this is the solution to a lot of problems).
• Be liberal in deleting old code.
• Find one source of truth. (A good idea in general, but specifically for metadata quality.)
• Big Tech is not your friend – whether “Big” in your context means AMZN and GOOG, or Clarivate.
• Information takes the shape of its container. (Or in other words, the relations in your relational database say a lot about, and perpetuate the values of, the people who set it up.)