Back to reporting this week, after a week too full of fun conversations with real people to leave time for diary-writing! It was great to finally see the exhibit in Newcastle and hear local historians thoughts on the prospects of a national collection. The next day focusing on Generative AI in Leeds was a stimulating conversation, and has made me quite confident an article presenting the ethical problems alongside the historiographic potential of feeding historical sources to ChatGPT ought to be an output of the circulars work.

On the theme of article writing, I’ve started making my own list of potential journals to aim for with each investigation. If anyone has a similar resource of their own and would like to swap notes, please do!

Now, for some updates:

Personal researcher notes

Thanks to Jon and Arran who helped me improve the personal researcher notes feedback form. This was sent to six of the researchers on Monday, with two still researching.

The form exists to establish:

• If the researchers took notes differently to usual
• If they had specific research questions in mind
• If they think the descriptions are complete, accurate, and whether or not they would change them
• How they would expect to encounter AI generated descriptions within the archive
• If they feel more or less confident about automatically generated descriptions since starting work on this investigation

So far, five have replied, meaning I can share early results that:

• The majority find the descriptions to be accurate and complete enough for use in the archive – however, all but one person would like to change the description if given the chance (this, I believe, is a happy outcome of engaging passionate specialists)
• While one person would expect to be credited if these descriptions were in an archive, another “would not like to be associated with such an overconfident summary”
• 3/5 people would want AI generated descriptions to be tagged as such, 1/5 would like them to have a clear summary sentence and a link to an article explaining the process, and 1/5 would not expect AI descriptions to be shown any differently to usual
• 4/5 would feel confident trusting automatically generated descriptions in the future
• 4/5 feel more confident about automatically generated descriptions now than they did at the start of the investigation

Gender

I have imported Lucy’s baseline gender-bias model into a Jupyter notebook, and attempted to feed the BT catalogue into this. Before we can start seeing results, we need to clean BT’s catalogue to make it match Lucy’s original data input, and get confirmation from Lucy about which type of “prediction” to input into Jupyter to get a result.

We are, for the first time in a while, close to some actual results. Fingers crossed for more updates soon!