HomeMy WebLinkAbout052526 email - Proposed (UPDATED) AI policy - Gemini says ___ALERT: BE CAUTIOUS This email originated outside the organization. Do not open attachments or click on links if you are not expecting them.
Commissioners,
Lest you think that ChatGPT’s analysis of your proposed (UPDATED) AI policy is flawed, I submitted it to Google Gemini.
Gemini’s response is below =====. Note that Gemini also caught the error in Section 7 (b) that caught my attention (see prior email – missing “not”)
Tom Thiersch
===============================
This policy offers an exceptionally detailed legal framework, but it contains critical logical contradictions, severe execution blocks, and document assembly errors that will disrupt
county operations.
________________________________
🛑 Fatal Logical Contradictions
* The Prohibited/Permitted Loophole: Section 7.a.i explicitly establishes Microsoft Co-Pilot as the only approved generative AI system (Policy pag... p. 11). However, Section 7.b immediately
declares that utilizing any AI system found on the approved list is strictly prohibited (Policy pag... p. 11). This typo effectively outlaws the county's only legal tool.
* The "Invisible" Blocked List: Section 7.b mentions a "list of AI systems... made available on the County's website" that are banned (Policy pag... p. 11). However, the Table of Contents
and Section 7.a.ii indicate that the website is only supposed to host the approved list (Policy pag... pp. 2, 11). The prohibited list has no clear owner or publishing home.
________________________________
🛠️ Operational & Technical Execution Flaws
* Draconian Data Restrictions: Section 7.f.iii strictly bans inputting any Personally Identifying Information (PII), explicitly including names, email addresses, and phone numbers (Policy
pag... p. 12). Because enterprise tools like Co-Pilot integrate directly with email and calendar workflows, this rule completely breaks basic office automation (e.g., summarizing an
email chain containing standard staff contact info).
* Vague "High-Risk" Review Escalation: Section 7.d requires "enhanced review" for high-risk AI, and Section 11.b.i tasks the Central Services Director with building a baseline evaluation
standard (Policy pag... pp. 11, 16). The policy fails to outline the concrete steps, metrics, or evaluation criteria required for a system to pass this enhanced review.
* Unrealistic Public Audit Expectation: Section 11.h.ii mandates that all technology vendors disclose their AI components and provide public auditing mechanisms via a standardized form
(Policy pag... p. 18). Major international software providers will not modify their core enterprise product structures or compliance disclosures to satisfy a single local county form,
effectively halting standard tech procurement.
________________________________
🎨 Layout, Usability, & Assembly Errors
⚠️ CRITICAL COMPLIANCE RISK
The document relies on legally binding appendices to define scope,
but fails to attach them properly. The text references Appendix D
for state statutes, but the Table of Contents ends at Appendix C.
* Broken Cross-References: Section 13.b states that disciplinary conduct standards are located in "Appendix C of to the Jefferson County Personnel Manual" (Policy pag... p. 20). This
double-preposition typo points to an external document, whereas the policy's own internal Appendix C covers an entirely different topic (records protected from disclosure) (Policy pag...
p. 2).
* Incomplete Templates: The formal running header across the entire policy leaves the mandatory Document Code No.: ________ line entirely blank (Policy pag... pp. 2-3). This compromises
formal legislative tracking.
* Sentence Fragmentation: Section 11.f cuts off abruptly in the middle of a thought at the bottom of the page, leaving a trailing, incomplete sentence: "...consistent with the County's
practice in communicating" (Policy pag... p. 17).
________________________________