Canadian Politician Reads AI Prompt Instructions Aloud in Legislature

Bill Oliver, a Progressive Conservative member of the New Brunswick legislative assembly, read what appears to be an AI model's own formatting instruction aloud during a floor speech. The phrase "here's a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points" made it into the official record, a clear sign the LLM output was pasted in without any review. The clip went unnoticed at the time but spread on Reddit and Threads this week, drawing coverage from the CBC and the Toronto Star.

What happened

Detail	Fact
Who	Bill Oliver, Progressive Conservative MLA, New Brunswick, Canada
What he read aloud	“here’s a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points”
When it spread	Earlier this week, via Reddit and Threads
Media coverage	Canadian Broadcasting Corporation and the Toronto Star

Oliver’s speech included a substantive point about advocacy offices: “One of the dangers associated with creating advocacy offices is that citizens often develop expectations that exceed the powers actually granted to those offices.” That part reads normally enough. The problem came immediately after, when he read out the LLM’s own meta-instruction telling him how the rewritten text had been formatted, a sentence that no human speechwriter would ever include.

The remark went into the official legislative record without anyone in the chamber flagging it at the time. It took social media, specifically posts on Reddit and Threads, to surface the clip and push it into broader public attention.

Why it matters

This is not a story about AI being used in politics. That ship has sailed. Politicians, staffers, and lobbyists across every country already use LLMs to draft remarks, briefs, and correspondence. The story here is about the quality of review, or the absence of it.

When an LLM returns a rewritten passage, it typically includes a short explanation of what it changed and why. That explanatory line is meant for the person reading the response, not for anyone else. Pasting the entire output into a speech document, explanation included, and then reading it verbatim is a failure of basic proofreading.

The Toronto Star framed the incident as evidence of “a growing divide in our society: between the elites, who are only too happy to delegate their duties to the Borg; and the masses, who find this objectionable.” Whether or not you agree with that framing, the practical point stands: if you are going to use an AI tool to draft anything public, someone needs to read it before it goes live.

For businesses using AI to produce content, emails, or client-facing copy, the risk is the same. An unreviewed LLM output can include leftover instructions, contradictory suggestions, or placeholder text that makes it into the final product. We covered a related pattern in our look at how AI tools behave in unexpected ways when users push them beyond their intended workflows.

Our take

We use AI tools every day to draft, summarize, and restructure content. The productivity gain is real. But every output needs a human pass before it goes anywhere public. This is not a high bar.

The deeper problem is that some users treat AI output as a finished product rather than a first draft. An LLM will often include a framing sentence like “here is a revised version” or “below is a more conversational take.” Those are conversational scaffolding lines for the user, not content. If you are copying and pasting from a chat interface, you need to know where the AI’s commentary ends and the actual content begins.

For any business using AI to support AI-assisted content workflows, the fix is simple: build a review step into the process. Not a long one. A 30-second read-through would have caught this. The embarrassment here is not that Oliver used AI. It is that nobody checked.

What to do about it

1. Treat every LLM output as a draft, not a finished document. Always open it in an editor before sending or publishing.
2. Scan for meta-language: phrases like “here is,” “below is,” “I’ve revised,” or “this version” are signals that the AI’s framing text got included by accident.
3. Assign a named reviewer for any AI-drafted content that will be read aloud, published, or sent to clients. One person, one responsibility.
4. If you use a tool like ChatGPT or Claude in a chat interface, copy only the body of the response, not the full message thread including the AI’s explanatory preamble.

A 30-second proofread is cheaper than a national news cycle.

Source: Ars Technica · AI