Get your AI Guard up

Ten AI Logical Fallacies Every City Hall Should Know

A field guide for early adopters — because in municipal governance, there's very little tolerance for a mistake.

Peter Karwacki · Candidate, Ward 13 (Rideau-Rockcliffe)

Ottawa is already using AI. Staff are using Copilot by the thousands, resume screening tools are in the hiring pipeline, and infrastructure-mapping AI is quietly informing capital decisions. That adoption curve is not slowing down — and the tolerance for error in city governance is not the same as the tolerance for error in a casual chat with a chatbot.

A wrong answer in a personal conversation is a minor inconvenience. A wrong answer baked into a zoning decision, a budget line, or a benefits determination becomes a public record, a headline, or a legal exposure. The tool isn't necessarily the problem. The gap between how confident a piece of AI output sounds and how reliable it actually is — that's the problem. And that gap has a name, or rather ten of them.

Here's a top ten list of the logical fallacies I think every early adopter in municipal government — councillors, staff, and residents alike — should be able to recognize on sight.

1. 1. Fluency Fallacy Mistaking confident, well-structured prose for accuracy. A wrong answer delivered smoothly reads as more credible than a hedged right one. Fluency is a style property, not a truth property.
2. 2. Authority by Automation Treating "the AI said so" as if it settles a question, the way you might defer to a credentialed expert. A model has no accountability, no license to lose, and no reputation at stake — it shouldn't get expert-level deference by default.
3. 3. Sycophancy Blindness Not noticing when a model is agreeing with you because you want it to, not because you're right. Especially dangerous when a leading question comes back as your own opinion, reflected as if it were independent analysis.
4. 4. Introspection Illusion Trusting an AI's explanation of why it produced an answer as if that were a real account of its process, rather than a plausible-sounding reconstruction generated after the fact.
5. 5. False Precision Treating a specific number, date, or citation from an AI as verified just because it's specific. Specificity feels like rigor. It isn't evidence.
6. 6. Consistency-as-Truth Assuming that because a model gives the same answer twice, or two different AI tools agree, that increases confidence in correctness. Shared training data and shared blind spots produce agreement too.
7. 7. Verification Decay Starting out by fact-checking every AI output, then quietly stopping as trust builds and workload increases. The failure doesn't happen on day one. It happens once the habit of checking has eroded.
8. 8. Responsibility Laundering Using "the AI recommended it" to diffuse accountability for a decision, consciously or not. The tool doesn't bear consequences. The person who acted on it does — and pretending otherwise is where governance trust breaks.
9. 9. Static Expertise Assumption Assuming a model that's good at one task — drafting, summarizing — is equally reliable at another, like legal interpretation or financial modeling, just because the interface and tone feel the same across tasks.
10. 10. Silence-as-Confirmation Reading a model's lack of pushback as evidence that an idea has been checked and holds up, rather than as a model doing what it was optimized to do: keep the conversation smooth.

The failure isn't usually the AI being wrong. It's the human's confidence outrunning the actual reliability of what's in front of them.

That gap — between felt-confidence and actual-verification — is the whole risk in a governance context. It's also fixable, but not by asking people to "just be more careful." Verification has to be structural: mandatory citation to source documents, a named human sign-off, and treating any AI output the way you'd treat a draft from a junior staffer — useful, but never final without review from someone who has skin in the game and answers to the public.

It also means residents have a right to know where AI is being used before an error surfaces, not after. Disclosure changes the incentive. It's much easier to skip verification on a tool nobody knows you're using.

Where this connects: I've written separately about the disclosure gaps in Ottawa's current AI use and what a real Municipal AI Disclosure & Audit Standard should look like. If accountability — human and algorithmic — is something you want in your next councillor, I'd welcome your support.

— Peter Karwacki, Ward 13

i