Citizen Portal
Sign In

Get Full Government Meeting Transcripts, Videos, & Alerts Forever!

Get email alerts on the Artificial Intelligence topic

No spam. Unsubscribe anytime.

Raleigh CIO outlines AI uses, draft policy and guardrails during council briefing

2676997 · March 18, 2025
AI-Generated Content: All content on this page was generated by AI to highlight key points from the meeting. For complete details and context, we recommend watching the full video. so we can fix them.

Summary

City of Raleigh Chief Information Officer Mark Wittenberg described current and planned artificial intelligence projects, said a city AI policy is nearing completion, and detailed guardrails being developed for data security, training and public‑records handling.

Mark Wittenberg, the City of Raleigh’s chief information officer, briefed the mayor and City Council on Oct. 12 about the city’s current and planned uses of artificial intelligence, the status of an internal AI policy and steps the city will take to protect data and manage public‑facing deployments.

Wittenberg said the city has run AI projects for years but is now focusing on generative models and “deep learning” tools such as large language models. “The human is at the center of all this,” Wittenberg said, describing humans as the final check on model outputs and the source of accountability.

The presentation matters because the city plans to expand AI into services residents use and into infrastructure management, while the council and staff consider how public‑records law, privacy and security apply to AI‑generated content.

Wittenberg described three tiers of AI work the city is using and testing: an internal chatbot called “Raleigh” that answers IT and HR questions; a web‑search assistant pointed specifically at city web content; and department‑specific models, including a closed model used by Public Utilities to analyze historical records and predict pipe failures. He said the public‑utilities model was built by a contractor, Affirmed, and that the model has been trained on records the city assembled into a searchable database.

He said the city is testing the tools internally first and warned that open models can retain prompts and are harder to correct. “If you feed something into chat GPT, it is there and it is going to use that as data,” Wittenberg said, noting there is no “undo” button on external models.

Planned public uses Wittenberg outlined include a customer‑service portal slated for initial rollout in June or July that will surface common services and use AI in the background to identify missing service categories and suggest workflow changes; later phases would add generative chat by phone, web chat, social media and email so customers receive consistent answers. He also said staff are testing a building‑code search that would query the city’s Unified Development Ordinance to guide residents through permitting steps.

On governance, Wittenberg said the city is finalizing an administrative AI policy and that it will be processed through the city’s staff policy workflow; after implementation staff plan to return to the council with an update describing governance and operational details. He described an IT governance committee made up of the city manager’s office and department heads that will prioritize pilots and licensing. He also said Copilot licenses for Microsoft Office will be limited and that the city is considering mandatory training before employees receive licenses.

Wittenberg addressed several risks raised by council members during a question‑and‑answer session: data security for proprietary records, the environmental footprint of large AI models, bias in model outputs, and public‑records implications for AI‑generated documents. He said closed, city‑owned models give the city the ability to correct or remove problematic data; open external models do not. He added that the city attorney’s office has advised that content produced and preserved as a final product likely becomes a public record, while transient internal artifacts can be treated as transient records.

Wittenberg also cited membership in an AI Coalition—he said Raleigh was a founding member—as helpful to accelerate development of internal guardrails. He estimated the policy work and initial pilots reached the current stage after about six months of effort.

Council members asked about quality control, testing logs and the ability to surface when answers come from old or archived web pages; Wittenberg said logs exist for internal tools and that IT and Communications staff will work together during the website redesign to remove outdated pages and reduce contradictory information. He said Copilot will have an intentional toggle to allow or prevent access to external sources and that staff are working to add warnings when a user is about to enter information into an open model.

Wittenberg said Public Utilities has used a closed model to combine council records, purchase documentation and historical records into a database that can be queried to predict materials and likely future failures. He described one pilot in which the model answered historical questions about past mayors and said the city expects closed models to enable similar searches of public records while limiting exposure of proprietary infrastructure details.

Multiple council members raised environmental concerns about the electricity and water consumption of large data centers. Wittenberg said the city is beginning those conversations and will consider the energy sources and efficiency of vendors as part of procurement decisions.

Wittenberg emphasized a cautious, programmatic approach to public deployment: internal pilots first, policy and training before broad release, governance oversight and limited initial licensing. “We don't want to be on the bleeding edge of this; we want to be very cautious,” he said, adding that security and protecting community data are foundational to the city’s approach.

The briefing closed after a final round of questions. No formal council action or vote was taken during the work session.