A Canadian legislator read an AI prompt out loud in a floor speech, and the clip is doing the rounds this week for all the wrong reasons. According to Ars Technica, Bill Oliver, a Progressive Conservative MLA in the New Brunswick legislature, said "Public confidence in the office of an advocate matters," and then kept reading: "Here's a more natural, flowing version of that section that reads like a legislative speech rather than a series of short points."
That second sentence was never meant to be spoken. It's the chatbot talking to the human. He read it anyway, reportedly without blinking.
I'm not writing this to laugh at a politician. I'm writing it because the exact same mistake is one keystroke away for every student, freelancer, and small-team builder in Sri Lanka who now drafts with an LLM. The interesting question isn't "was AI used." It's "who was supposed to read this before it went out, and why didn't they."
🔍 What actually went wrong (it wasn't the AI)
The model did its job. It rewrote a bullet-point section into flowing prose and prefixed the result with a one-line note explaining what it did. That prefix is standard chatbot behaviour. The failure was entirely on the human side of the handoff:
- Someone pasted the model's entire reply into the speech, including the meta-commentary.
- Nobody read the final text end to end before it was delivered.
- The speaker read words off a page he hadn't internalised.
Key takeaway: The tool didn't fail. The review step failed. An LLM will hand you a draft with scaffolding still attached — the "here's a version that…" line, a stray "Let me know if you'd like…", a placeholder in square brackets. Deleting that scaffolding is your job, not the model's.
Every one of those three steps is something you control. None of them requires trusting the AI more or less. They require reading your own output before you ship it.
🛠️ The AI-slop tells you should grep for before you hit send
LLM output has a recognisable residue. If you draft with a chatbot, learn to spot these before the text leaves your machine. Here's the common giveaways and the fix:
| Tell | What it looks like | Fix |
|---|---|---|
| Meta-commentary | "Here's a more natural version of…" | Delete the whole framing line |
| Sign-off bait | "Let me know if you'd like me to expand" | Delete; it's talking to you |
| Placeholders |
[Insert company name], [date]
|
Search for [ and fill or cut |
| Hedge stacking | "It's important to note that…" x5 | Cut the throat-clearing |
| Filler adjectives | overuse of "seamless", "cutting-edge" | Replace with specifics or delete |
The reliable move is mechanical: after the model gives you text, run a find for [, for "here's", for "version of", and for "let me know". Thirty seconds. It would have caught the New Brunswick speech instantly.
If you write in Sinhala or Tamil and translate through a model, the residue often survives translation too, so the same check applies to both languages.
💡 Use the model as a first-drafter, not a final-drafter
The correct mental model is that the LLM is a fast junior who writes an okay first draft and never checks their work. You would never forward a junior's draft to a client untouched. Same rule here.
- Draft with it. Bullet points to prose, rough notes to a structured email — this is where models genuinely save you an hour.
- Never publish it raw. Read every sentence aloud in your head. If a line sounds like a chatbot, it is one.
- Own the facts. Models invent citations, numbers, and quotes. If you can't verify a stat yourself, cut it. This applies to code comments and commit messages too.
I lean on AI to reshape messy paragraphs constantly, and I still read the whole thing back before it goes anywhere near a reader. If you want a second machine opinion on whether your text reads as obviously AI-generated, our AI content detector comparison lines up how the popular detectors score the same passage — useful for a gut check, though no detector is a substitute for reading it yourself. And if the problem is the opposite — a wall of chatbot text you need to shrink — the AI text summarizer will do that in the browser without an account.
🌐 Why this matters more for a Sri Lankan builder on a free tier
If you're a UCSC student submitting coursework, a freelancer sending a Colombo client a proposal, or a two-person startup writing your first landing page, you don't have a comms department to catch a slip like this. You are the review step. There is no editor between your draft and the world.
The higher the stakes and the smaller the team, the more the final read-through is non-negotiable. A legislature has Hansard transcripts, opposition MLAs, and a national press watching. Your client just quietly decides you're careless and doesn't reply.
The good news: the fix costs nothing and needs no tooling. A free-tier chatbot plus your own eyes beats a paid one used carelessly, every time. The skill that's scarce now isn't prompting. It's the discipline to read what came back.
🚀 What this means for you
The New Brunswick clip is funny for a day and forgotten by next week. The pattern behind it is going to keep happening as more people draft with AI and skip the read-back. Don't be the case study.
Three habits, starting today:
- Treat every model reply as a draft with packaging still on it. Unwrap it before you use it.
-
Grep your output for
[, "here's", "version of", "let me know" before anything ships. - Read the final text end to end, out loud if you can. If a sentence sounds like a robot wrote it, rewrite it in your own voice.
The tools are good enough that the only remaining variable is you. That MLA had a working AI assistant and a broken review habit. Keep the assistant. Fix the habit.
Top comments (0)