New Brunswick MLA Reads Raw LLM Text Into Hansard Record

What Happened on the Floor of the New Brunswick Legislature

Here is what can and cannot be established. According to a widely shared social-media post, a member of New Brunswick's legislative assembly appears to have used a large language model (LLM) to help draft a floor speech and then delivered it without adequately reviewing it — reading AI-generated scaffolding or placeholder text aloud into the chamber. Academic Robert Komaniecki flagged the moment on Bluesky, describing the politician as "this utter muppet of a conservative Canadian politician in New Brunswick." That post is the origin of the story as it circulated in tech and political circles.

It is important to be precise about what is verified and what is not. As of this writing, the politician has not been independently identified, the exact phrases said to have been read aloud have not been independently confirmed, and there is no independently verified Hansard citation establishing that the text entered the official record. Komaniecki's post is the substance of what spread; the underlying claim has not, at the time of publication, been corroborated by the legislature's transcript, by named journalists, or by the assembly itself. Readers should treat the specifics as an as-yet-unverified allegation from a single source rather than an established fact. We have not been able to confirm the party affiliation of any specific member, and we do not name or attribute negligence to any identifiable individual.

Why cover it at all, then? Because the category of failure the post describes is real, recurring, and worth examining regardless of whether this particular instance holds up. Floor speeches in a Canadian provincial legislature go into the official Hansard record — the verbatim transcript of legislative proceedings, published, archived, and citable long after the fact. If an LLM artifact — a dangling instruction like [insert supporting statistic here], a meta-comment like Note: you may want to adjust this for your audience, or an unfilled template placeholder — were read aloud in such a setting, it would plausibly become part of that permanent record. Hansard has no undo button, and that permanence is precisely what makes the scenario worth taking seriously.

If the incident is as described, it would carry a particular sting in any partisan context, because it names a failure of professional diligence rather than ideology. But the more durable lesson here does not depend on the identity or party of the person involved: it is about how AI-assisted drafting fails when the human review step is skipped, in a venue where the output becomes permanent.

The Anatomy of an LLM Artifact: What Gets Left Behind

To understand how a scenario like this could occur, it helps to understand how politicians and staffers typically use LLMs. The workflow is rarely "ask the AI to write a speech and read it verbatim." More commonly, a staffer or the member prompts a model — ChatGPT, Claude, Gemini, or another — with something like: "Draft a five-minute floor speech opposing Bill X on the grounds of fiscal responsibility." The model produces a full draft. But depending on how they are prompted and which model is used, LLMs frequently insert:

• Bracketed instructions to the human, e.g., [Add your personal anecdote about a constituent here]
• Placeholder citations, e.g., [Source: Statistics Canada, year TBD]
• Structural meta-commentary, e.g., You may wish to soften this paragraph if the audience includes members from rural ridings
• Formatting headers designed to organize the draft for editing, not to be spoken aloud
• Hedging or transitional language that reads as characteristically AI-generated rather than as human political rhetoric — phrases like "it is important to note that" stacked closely together, or the characteristic LLM em-dash sandwich

Any one of these, read aloud, is immediately recognizable to anyone modestly familiar with how LLMs structure their output. The tell isn't just the content — it's the register. AI-generated scaffolding sounds like instructions about a speech rather than the speech itself. When a speaker abruptly shifts from political rhetoric into what sounds like a project manager's to-do list, listeners tend to notice that something has gone wrong.

This Is Not the First Time: LLM Artifacts in Public Documents

Whatever the truth of the New Brunswick allegation, it points to a documented and growing category of failure. Since LLMs became widely accessible in late 2022, artifact leakage into public-facing documents has been reported across multiple professional contexts.

The best-documented example is legal. In the 2023 US federal case Mata v. Avianca (S.D.N.Y.), attorneys submitted a brief containing case citations that ChatGPT had fabricated — plausible-sounding precedents that did not exist. The lawyers had not verified them, the citations entered the court record, and the court sanctioned the attorneys. That case is on the public record and is frequently cited as a cautionary tale.

Beyond that, observers have repeatedly noted a broader pattern of tell-tale AI phrasing surfacing in published text — for example, boilerplate disclaimers or interface language appearing where it plainly should not. AIBites has not independently verified specific instances of such phrasing in official municipal minutes or in peer-reviewed papers, and we present those as widely circulated anecdotes rather than confirmed cases. What is not in dispute is the underlying mechanism: users trusting model output without reading it closely enough to catch what the model left behind.

The through-line across these situations is consistent: a user trusted the model's output without reviewing it carefully enough to catch the artifact, and — where it happened — the artifact entered a context (a court record, official minutes, a published paper, a legislative Hansard) that assigns it a permanence and authority it was never designed to carry.

Why "Utter Muppet" Resonated as More Than a Punchline

Komaniecki's characterization — "this utter muppet of a conservative Canadian politician" — is blunt, but it resonated because it names a claimed failure of professional responsibility rather than a merely technical slip. Using AI to assist with drafting is, at this point, largely normalized across law, medicine, journalism, and politics. The alleged failure here isn't in using the tool. It is in not reading the output before speaking in a legislative chamber — a task requiring no technical sophistication whatsoever, only the basic habit of reading one's own speech before delivering it.

If a politician reads LLM scaffolding into the Hansard record, they have not merely embarrassed themselves — they have demonstrated, on the permanent public record, that they did not write, review, or meaningfully engage with their own legislative speech.

That distinction matters for democratic accountability. A floor speech is not a tweet. It is a formal act of legislative participation — the mechanism by which elected representatives place their views, arguments, and reasoning into the official record of governance. When that record contains an LLM's internal stage directions, it raises a legitimate question: whose speech was this, and did the member delivering it understand what they were saying?

This is also why the framing spread online rather than being dismissed as partisan point-scoring. The critique, as it circulated, was less ideological than professional — about basic competence in the use of a tool that has become ubiquitous. Wikipedia's community has moved to restrict AI-generated text in article writing and to expand deletion pathways for unreviewed AI content, reflecting a related anxiety: that AI-generated text inserted into authoritative contexts without adequate human review corrupts those contexts in ways that are structurally difficult to undo. (Readers should note that Wikipedia's rules are community-enforced guidelines and speedy-deletion criteria rather than a single blanket statutory "ban.")

Politicians and AI Speechwriting: A Growing and Largely Unregulated Practice

Even setting aside the specific allegation, AI-assisted political drafting is not isolated. Across democratic legislatures, politicians and their staff have been incorporating LLMs into the drafting process for speeches, press releases, committee questions, and constituent correspondence at an accelerating pace since late 2022. The tools are free or cheap, fast, and — when properly reviewed — genuinely useful for producing first drafts. The recurring problem is the "properly reviewed" part.

Legislative bodies have generally been slow to establish formal policies governing AI use. Most Canadian provincial assemblies, like their federal counterpart in Ottawa, have no widely publicized standing-order provisions specifically addressing AI-drafted content in speeches or official submissions. In practice, the main safeguard against an incident like the one alleged is the individual member's own diligence.

Context	AI Use Status	Formal Policy?	Risk if Unreviewed
Legislative floor speech	Increasingly common	Rare (most jurisdictions)	May enter permanent Hansard record
Press releases / communications	Widespread	Rarely	Reputational damage, factual errors
Constituent correspondence	Common	Rare	Misleading or irrelevant responses to voters
Committee question preparation	Growing	Rare	Embarrassment when challenged on facts
Court filings (US, Mata v. Avianca, 2023)	Occurred; condemned	Emerging judicial guidance only	Hallucinated citations; attorneys sanctioned
Wikipedia articles	Restricted	Yes — community guidelines / deletion criteria	Misinformation in public knowledge base

The table illustrates a consistent pattern: the higher the stakes and the greater the permanence of the context, the less formal governance tends to exist around AI use within it. A volunteer encyclopaedia project has, in many respects, moved faster to implement AI-content rules than most elected legislatures. In the US, some individual federal judges have issued standing orders requiring disclosure of AI use in filings, but those orders are piecemeal and unevenly applied. Comparable, uniform rules for Canadian provincial legislatures do not appear to be in place.

What This Means for Developers Building AI Writing Tools

For the developers and technically fluent readers who make up much of this publication's audience, the scenario raises a practical design question: whose responsibility is it to prevent an LLM artifact from escaping into a high-stakes public context?

The current default answer — "the user's" — is arguably insufficient when the user is a politician or staffer with limited familiarity with how LLMs structure their output, working under time pressure, and treating the AI's draft as a near-final document. Several design interventions are worth building and evaluating:

• Artifact detection and flagging: A post-generation pass — in the model's output layer or in the application wrapper — that identifies bracketed instructions, hedging meta-language, unfilled placeholder patterns, and transitional boilerplate, then surfaces them prominently to the user before the content can be exported or copied. Simple pattern-matching for common LLM formatting signatures can catch many of these.
• Contextual risk warnings: When a user's prompt explicitly signals a high-stakes public context — "floor speech," "press statement," "official testimony," "Hansard" — the interface could apply a higher-friction review gate. A bare "Have you read this fully?" confirmation is weak; a structured checklist of artifact types to check for would be more effective.
• Explicit draft-versus-deliverable state separation: Distinguishing clearly in the UI between a "working draft" state and a "ready to deliver" state, requiring active user confirmation that artifacts have been identified and resolved before the document transitions. This mirrors the "track changes accepted" workflow in document editing and is not technically novel.
• Transparency and attribution prompting: Nudging users to disclose AI assistance in formal public contexts — both for transparency and to keep accountability where it belongs, with the human who chose to use the tool.
• Output register auditing: Flagging when a generated document contains abrupt register shifts — from formal political rhetoric to instructional or meta-commentary language — which is often the precise signature of an artifact that was never meant to be read aloud.

None of these interventions are technically complex. What they require is product teams deciding that preventing user embarrassment — or, more seriously, democratic harm — is worth the mild friction they introduce into the generation flow. As AI systems are pushed into increasingly autonomous, high-consequence workflows, the case for robust guardrails at the human-facing output layer only grows stronger.

The Broader Accountability Gap in AI-Assisted Politics

Stories like this one sit at the intersection of two trends: declining trust in political institutions, and rising uncertainty about the authenticity of AI-mediated communication. Together they can produce something more corrosive than either would alone.

If voters come to believe that a politician's speech was drafted by an LLM — and that the politician did not read it before delivering it — the reaction tends not to stop at cynicism about one member. It seeds a broader suspicion: how many other speeches, committee appearances, and policy announcements are AI-generated drafts the member never genuinely engaged with? That suspicion, once planted, is difficult to rebut. It is worth stressing again that, in the New Brunswick case specifically, the underlying facts remain unverified — which is itself a reason to be careful about how such a story is amplified.

The Bilingual Dimension: A Plausible Elevated Risk in New Brunswick

There is a specifically Canadian dimension worth flagging as analysis rather than established fact. New Brunswick is Canada's only officially bilingual province, with a legislature that operates in both English and French and a substantial Francophone (largely Acadian) population — a large minority of residents report French as a first language. Political speech in that context is not merely translated: it operates within two rhetorical traditions and two sets of cultural and regional expectations about how a politician sounds.

LLMs can produce fluent-sounding but contextually shallow output when navigating bilingual or culturally specific political environments, particularly when prompted primarily in one language. It is reasonable to hypothesize that the risk of an AI-drafted speech in New Brunswick's legislature failing on dimensions beyond visible artifact leakage — getting the register wrong for Francophone listeners, flattening culturally specific nuance, or missing the temperature of a bilingual chamber — is higher than in a monolingual setting. This is our analysis of a plausible risk, not a documented finding about any specific speech.

Key Takeaways

• A social-media post by academic Robert Komaniecki alleged that a member of New Brunswick's legislative assembly read LLM-generated text — including AI scaffolding never meant to be spoken — aloud during a floor speech. His description of the politician as "this utter muppet of a conservative Canadian politician in New Brunswick" is how the story spread.
• The specifics are unverified: the politician's identity, the exact text, the party affiliation, and even confirmation that the event entered Hansard have not been independently corroborated at the time of publication. Treat the incident as an unverified allegation from a single source.
• Hansard, the verbatim legislative transcript, is durable and hard to correct; if an LLM artifact were read into it, it would plausibly become part of the permanent record.
• The scenario belongs to a real, documented category of failure — most clearly the 2023 US case Mata v. Avianca, in which lawyers were sanctioned over ChatGPT-fabricated citations that entered the court record.
• Most Canadian provincial legislatures appear to lack a formal, publicized policy governing AI use in drafted speeches, leaving individual members as the main safeguard.
• For developers, the episode reads as a design brief: LLM interfaces do not yet do enough to stop artifacts from escaping into public, permanent, high-stakes contexts, and several practical interventions are available.
• New Brunswick's bilingual environment plausibly raises the risk profile beyond visible artifacts, as LLMs can flatten cultural and linguistic nuance — presented here as analysis, not a documented finding.
• The broader risk is to trust: if a politician is shown not to read their own speeches, voters gain a legitimate basis to ask whose reasoning is actually entering the legislative record.

What Comes Next

If the allegation holds up, the immediate next step — beyond political embarrassment — would likely be a quiet internal tightening of AI review practices within the affected office, and possibly an uncomfortable conversation in provincial assemblies about whether standing orders should address AI-drafted content at all. Whether that conversation produces actual policy is another matter. Legislatures move slowly, and the incentive to formally regulate something that many members across parties are quietly using themselves is structurally limited.

There are reference points elsewhere, though they should be read carefully rather than treated as uniform bans. The UK Parliament and various public bodies have published non-binding guidance on generative AI use by staff. Some US state legislatures have begun discussing disclosure requirements for AI-assisted legislative work. In the EU, institutions operating in the shadow of the AI Act have examined internal AI-use practices. Canada's provincial assemblies have, to date, largely stayed out of this conversation. A high-profile incident — verified or not — has a way of making that conversation harder to avoid.

For developers and AI product teams, the pressure to build better human-review checkpoints into high-stakes output flows will only intensify as episodes like this accumulate. And they will accumulate — because the tools are fast, the temptation to skip review is real and deeply human, and the floor of a legislature is an unforgiving place to discover you have an AI problem.