Blog
Filings, personal data and artificial intelligence.
A few pieces, written when we have something to say: what actually leaves an office when a filing ends up in a chat window, where hand work goes wrong, and which words to use when somebody asks what happened to the data.
Using AI responsibly
An AI policy for a three-person firm
One page, six sections and a list of permitted tools. Twenty-page policies do not get read, and an unread rule is not a measure.
Using AI responsibly
When the model invents a judgment
The Court of Florence dealt with a filing citing judgments that did not exist. What it decided, and what it teaches about review.
Using AI responsibly
Human review is not a checkbox at the bottom of the screen
“Checked by a professional” is worth as much as the conditions that check happens in. What makes a review real rather than formal.
Using AI responsibly
What a model remembers about what you gave it
Training, request retention and conversation history are three different things with three different answers. Where to look for each.
Using AI responsibly
Local or cloud: the question that comes before the others
Where the program runs decides who sees the documents, how many contracts you need and what goes in the record. The other questions come after.
Using AI responsibly
What stays in a file after you take the names out
Tracked changes, comments, document properties, the text layer under a scan. Five things that survive an anonymisation done by reading the page.
Using AI responsibly
Redact or substitute: two different jobs on the same filing
A black box and a placeholder solve different problems. When you need each, and why the word “anonymisation” blurs them together.
Try it on a case you have already closed.
Three days with every feature. The filing stays on your computer, and at the end of the trial you decide.