19473_Authority_Oct-opt

municipalauthorities.org │ 1 7 g uardrails , n ot g ates AI G OVERNANCE FOR P ENNSYLVANIA ' S M UNICIPAL A UTHORITIES By Ben Johnson, Freya Systems should have a named human who checked it. Then cover the outside. This is the piece most policies miss. June 4, 2026, AWWA and WaterISAC advised water and wastewater systems to safeguard sensitive operational information, after a run of public records requests from AI service providers. One asked for a full year of SCADA logs across treatment, distribution and wastewater management. Requests like that arrive lawfully. The cost of asking has collapsed, and the cost of answering has not. [www.waterisac. org/tlpclear-advisory-safeguarding- sensitive-operational-information] Decide in advance what you will tell your constituents about how the authority uses AI, and decide how you will handle records requests for operational data, including which exemptions you intend to rely on. Making those calls before a request lands protects your staff's time and keeps you inside the Right-to-Know rules instead of improvising against a five-day clock. A word on AI notetakers AI notetakers deserve separate consideration, because the instinct is to ban them and I believe that is the wrong call. The silver tsunami is here. According to the recent 2026 AWWA State of the Water Industry Report, roughly 20% of the water workforce is within five years of retirement eligibility, and what leaves with them is judgment, not headcount. Notetakers are one of the more effective answers to that: what a veteran operator explains during a shift handover is real knowledge transfer, captured. Three rules make them safe. [www.awwa.org/state-of- the-water-industry/] Be clear which meetings they may be used in, with executive session as the obvious line. Require that their use is declared to everyone present, every time. And remember that your meetings are public and your records fall under the Right-to-Know Law, so a transcript on a vendor's server is a record you are responsible for. Guardrails that encourage Guardrails are the other half, and their job is to encourage people rather than only restrain them. Decide what gets verified and how. For low-risk work, reviewing afterward is fine. For anything touching operations, money, or the public record, a permission model works better. The AI proposes, a person approves, and only then does anything happen. If that sounds familiar, it should. It is Section 302.1208 applied to a different kind of system. And the cost of skipping it is already visible. The Office of Open Records warned last July that filings now arrive citing cases that do not exist. Nobody sets out to send one. It happens when nothing sits between the tool and the document that goes out the door. [Wagenseller, Liz. "Artificial Intelligence and the Right-to-Know Law: Be Aware of the Risks." Open Records in Pennsylvania — official blog of the Pennsylvania Office of Open Records, 17 July 2025. https://openrecordspennsylvania. com/2025/07/17/artificial- intelligence-and-the-right-to-know- law-be-aware-of-the-risks/ ] Invest in fluency. People use tools badly when nobody has shown them how to use them well. Anthropic publishes a free course built around the 4D framework: Delegation, deciding what to hand over in the first place; Description, explaining the task properly; Discernment, judging whether the output is any good; and The greater risk is not that someone at your authority uses AI badly. It is doing nothing while it happens anyway. ŽŶƟŶƵĞĚ ŽŶ ƉĂŐĞ ϱϭ͘