AI and Boston.gov

Create

Effort to create these AI components (Chat, Search etc) is minimal.  The setup can be performed using the Google console without the need for programming.  Some specialist skills need to be developed to tune the components (especially chat), but these skills are learned in the console, and changes to the components are made in the console.

There is some effort required to embed the resulting components into the Drupal website.  This can be done using existing City of Boston developed Drupal components - but there would be value in creating custom components to host the widgets.  That would mean that no specialist skills would be needed in Drupal either, it would be a case of dropping the component snippets from the Google console into a Drupal config page.  The time and effort to design the component is unknown, but (regardless of the design) the time to code a custom Drupal component to host Chat or Search site-wide and/or page-specific would be less than a week.

Maintain

These resources require maintenance. Maintenance is at 2 levels:

1. Refinement Both search and chat would need to be monitored and periodically analysed and the component models tweaked.  This work would be done in the Google console and would typically not require any changes in Boston.gov. The amount of work required is not much, a few hours a month, but the more time and effort applied, the better the UX,
2. Updates This is a rapidly moving area, and changes to the underlying services offered by Google will be constantly changing and need to be monitored by engineering and architecture focused staff.  This may include changes to the way the app is configured in the Google Console, but also may require changes to the way the components are embedded in the site - or even the existence of the components as new services and products are created and the AI ecosystems mature.

City of Boston risks.

Gen-AI models scan data and return answers to questions based on the scanned data. Generic Gen-AI models (Bard, ChatGPT etc) scan data available (mainly) from public websites.  Trained models (such as the ones we will use on boston.gov) are given the data we wish them to consider.  Because PDF files on our website are public, generic AI models will potentially scan these documents as well as the text on web-pages.

Certain data contained on Boston.gov (web-pages and pdf/doc/xls files) is time sensitive, in that newer information renders older information somewhat wrong (e.g. a new version of a policy) or flat out incorrect (e.g. a policy which is later removed). 

Given that boston.gov (as the digital presence for City of Boston) is an authoritative source, we have a duty to try to ensure deprecated information is marked as such - particularly for our own AI agents.  At the moment, we rely on the capability of the AI to realize that new information replaces old, and that will probably be OK in the main, however we should put in place some strategies so that we can highlight deprecated or lower-weight information.

1. Regular reviews of PDFs and their deprecated versions which are available on-site.  Our current Asset Library is not really up to this task, and we may want to consider upgrading the document and image library tools on boston.gov.
2. For public-notices, events and some other content types, we should consider removing these pages after preset times, or having them moved to archive folders and be marked as archived in the body copy.
3. Ensure that each page has a last updated date clearly marked on it, so that the AI has some hints on recency of the information.
4. Other ideas ....