A Canadian politician delivered more than he intended in a recent legislative speech when he inadvertently read aloud a chatbot’s editing instruction, turning a routine address into a viral cautionary tale about the perils of unreviewed AI-generated text.
On a day that should have been business as usual in the New Brunswick Legislative Assembly, Progressive Conservative MLA Bill Oliver rose to speak about advocacy offices and statutory powers. But sandwiched between his policy points came a sentence that didn’t belong: “Here’s a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points.” The line had all the hallmarks of an AI writing assistant’s meta-commentary—explainer text meant for the person drafting, not for public delivery.
Video of the slip, first reported by Ars Technica and International Business Times Australia, spread quickly across Reddit and Threads. For everyone who has ever used a generative AI tool, the moment was painfully recognizable. It was the kind of wrapper text that ChatGPT, Copilot, and similar assistants generate around a suggested rewrite. And it opened a window into a workflow where the final review failed.
What Happened in New Brunswick?
The speech, delivered earlier this summer, focused on the pitfalls of creating advocacy offices with powers that may not match public expectations. Oliver’s substantive point was that such offices can lead citizens to expect more than what the legislation actually allows. But after that statement, he delivered what sounded exactly like an AI’s transition phrase: the kind that appears when you ask a chatbot to “make this flow better” or “turn these bullet points into a speech.”
Neither Oliver nor his office has publicly confirmed which tool, if any, was used. The available reporting does not prove the text came from a language model, nor does it establish who prepared the final document. As OpenAI has long advised, even asking a chatbot whether it wrote a particular passage is unreliable—models don’t have memory of generating specific content. Yet the wording is unusually damning because it explicitly references a “more natural, flowing version” and contrasts a legislative speech with “a series of short points.” That’s classic production-side commentary, not something a speechwriter would draft for an elected official to recite.
The chain of events likely went like this: a draft or set of bullet points was fed into an AI tool; the tool returned a revised passage along with that explanatory note; someone copying the output failed to remove the note; the document was printed or loaded onto a teleprompter; and the legislator read it aloud, apparently without noticing the foreign sentence. The error lay not with the AI, but with a breakdown in human quality control.
How AI Slips Into Your Work—and When It Goes Wrong
Oliver’s gaffe is memorable because it happened in a formal chamber, but similar artifacts have been cropping up in professional documents for years. The problem isn’t that people use AI to draft. It’s that they treat the output as finished without applying editorial scrutiny.
Common AI holdovers include phrases like “Here is a revised version,” “Certainly — below is…,” “You may want to adjust this based on your audience,” “This version is more professional and concise,” or “I hope this helps.” These can remain hidden inside a document if the person copying text skims instead of reads. And while such verbal leftovers are embarrassing, the more dangerous errors are factual: AI can fabricate quotes, misrepresent statistics, cite non-existent reports, or invent legal precedents.
The New Brunswick incident is a stark reminder that AI-generated text is a first draft at best. It may be fluent, but it isn’t guaranteed to be accurate, appropriate, or even relevant to the final setting. The only defense is a human reviewer who reads every word.
The Meta-Text Minefield: Common AI Tag-Alongs
If you use AI to assist with writing, get in the habit of searching for these red-flag phrases before you hit send or publish:
| Phrase | Typical Context |
|---|---|
| “Here is a more natural, flowing version…” | Style or tone adjustment |
| “Certainly — below is…” | Assent before delivering requested text |
| “You may want to adjust this based on your audience.” | Contextual caveat |
| “This version is more professional and concise.” | Comparison to previous draft |
| “I hope this helps.” | Friendly sign-off |
| “As an AI language model…” | Self-identification disclaimer |
Removing these is the absolute minimum. But the real work involves checking whether the surrounding content is factually sound, tonally appropriate, and faithful to your own voice and intentions.
How We Got Here: The Rise of AI Writing Assistants
Generative AI has moved from a novelty to a workplace staple in just a few years. Microsoft Copilot is now integrated into Word, Outlook, PowerPoint, and other Microsoft 365 apps, putting AI drafting assistance in front of hundreds of millions of Windows users. Google, Apple, and countless third-party tools offer similar capabilities. The promise is productivity: transform rough notes into polished prose in seconds.
But the speed is also the trap. When a coherent paragraph appears with a single click, the natural inclination is to trust it. This cognitive ease, combined with workplace pressure to produce more with less, can lead to what researchers call automation complacency—the tendency to accept machine outputs uncritically. The U.S. National Institute of Standards and Technology (NIST) has flagged this as a key risk, noting that AI-generated content can be “persuasive enough to pass a casual review” even when erroneous.
Research on workplace AI disclosure suggests that many employees use AI quietly, worried that admitting it will make them seem less capable or replaceable. This secrecy can backfire: when AI assistance is hidden, peer review and collaborative editing—natural safeguards—are skipped. The pressure to deliver perfect-seeming work fast can lead to just the kind of oversight Oliver’s speech represents.
For professionals in politics, law, medicine, journalism, and other high-stakes fields, the consequences of an unreviewed AI draft can range from personal embarrassment to legal liability. The New Brunswick speech is a relatively harmless example, but it’s a symptom of a broader cultural shift: we’re increasingly willing to outsource both writing and, crucially, the final check.
What to Do Now: A Practical Review Checklist for AI Text
Whether you’re a legislator, a business executive, or a Windows user drafting a sensitive email, the lesson is the same: AI can help, but you must own the output. Here’s a step-by-step workflow to ensure your AI-generated drafts don’t become the next viral cautionary tale:
- Use AI to create a first pass, not a final product. Think of it as an intern or a brainstorming partner. It gives you raw material, not a finished document.
- Read the entire text aloud. Oliver’s mistake might have been caught if he had rehearsed. Speech reveals awkward phrasing and errant meta-text that silent reading can miss.
- Search for placeholder phrases. Run a quick word search for “here is,” “certainly,” “as an AI,” “I hope,” etc., and delete any you find.
- Verify every fact. Cross-check names, dates, statistics, quotes, and references against original sources. Do not assume the AI got them right.
- Check tone and context. Is the language appropriate for the audience? Does it sound like you? Does it align with your organization’s values and policies?
- Remove confidential information. Ensure you didn’t inadvertently paste proprietary or personal data into a public AI tool. Review your prompt history and the output for sensitive details.
- Conduct a final human review. Have another person read the document if possible. A fresh pair of eyes can catch what you missed.
- Retain a final approved copy. This helps you track what was actually delivered, in case questions arise later.
This isn’t just for formal speeches. The same workflow applies to reports, client emails, presentations, and even internal memos. As AI becomes your co-author, treat its contributions as material to be scrutinized, not as a finished script.
For organizations, the incident reinforces the need for clear AI usage policies. Employees should know when AI is allowed, what tools can be used with what data, and who is responsible for reviewing outputs. Training should emphasize that generative AI is a drafting aid, not a source of truth.
The Bigger Picture: AI in Professional and Public Life
Bill Oliver’s slip may be funny, but the reaction reveals a public expectation that elected officials should know and own their words. Even though speechwriting has always involved staff, researchers, and talking points, the visible intrusion of an AI’s instructions made the process appear careless. And in an age where deepfakes and disinformation are real concerns, any hint that a politician’s words might not be their own erodes trust.
Looking ahead, we can expect more of these incidents as AI becomes more deeply embedded. But they don’t have to be disasters. They can be teachable moments that push individuals and institutions to adopt better habits. Features like AI watermarking or mandatory disclosure may help, but the ultimate solution is cultural: we must treat AI outputs with the same skeptical eye we’d apply to a hurried intern’s work.
For the millions of Windows users now finding Copilot suggestions in their Word documents and email drafts, the takeaway is straightforward: never let the tool be the final arbiter. The “more natural, flowing version” you get might sound good, but only you can judge whether it’s correct, appropriate, and truly yours.