Incident Overview
On June 9, 2026, during the 61st Legislature’s 2nd Session of the New Brunswick Legislative Assembly, Member of the Legislative Assembly (MLA) Bill Oliver delivered a speech containing unedited fragments of large language model (LLM) instructions. The incident was captured in the official recording of Daily Sitting No. 49, with the problematic passages occurring at 1:28:21 and 1:34:20. The speech, which began at 1:07:56 and concluded at 1:41:46, included direct readings of LLM-generated prompts such as ‘Here’s a more natural, flowing version of that section that reads like a legislative speech.’
The disclosure occurred unintentionally, as Oliver read aloud portions of the AI-assisted draft without removing the embedded instructions. The incident was later confirmed by the Legislative Assembly of New Brunswick and widely shared as an example of undisclosed AI use in political discourse. The recording is publicly accessible through the Assembly’s webcast archive.
The incident was documented under Incident ID 1607 in the AI Incident Database (AIID) on July 24, 2026, following its public circulation. The AIID categorized the event under the broader issue of unedited AI-generated content in legislative proceedings, noting its implications for transparency and public trust in political communication.
Possible Root Cause
The root cause of the incident appears to stem from the incomplete editing of AI-generated drafts prior to their use in official legislative speeches. Oliver acknowledged using an LLM to assist in drafting his remarks, but the final version retained visible traces of the model’s internal instructions.
The presence of unremoved LLM instructions indicates either a lack of familiarity with AI-assisted drafting tools among staff or legislators, or an oversight in the editing workflow. Such instructions often include placeholder text, reformulation suggestions, or structural guidance that are intended for internal use by the AI system or its human collaborators.
While AI can enhance efficiency in drafting, its use introduces new risks if proper safeguards—such as automated detection of residual prompts or mandatory human review—are not implemented.
Takeaways
The unintentional disclosure of LLM instructions not only compromised the credibility of the speech but also raised questions about the authenticity of the content itself. This incident serves as a cautionary example of how residual AI artifacts can undermine public trust in political processes.
- The public exposure of such oversights can erode confidence in the integrity of legislative proceedings, particularly when AI tools are involved. The incident became a widely circulated example of poor AI governance in political contexts, reinforcing perceptions of carelessness or deception.
- Misinformation and Fabricated Claims: The presence of LLM instructions in the speech introduced ambiguity about the authenticity of the content. While the instructions themselves were not false claims, their inclusion raised concerns about whether other portions of the speech might have been similarly altered or fabricated. This ambiguity can contribute to broader skepticism about AI-generated political messaging, even when the intent behind its use is legitimate.
- Had Oliver disclosed the use of AI-assisted drafting upfront, the unintentional inclusion of instructions might have been framed as a technical error rather than an act of concealment. Moving forward, legislative bodies should establish guidelines for the responsible use of AI, including mandatory reviews to remove residual prompts and transparent acknowledgment of AI assistance.
- Workflows and Human Oversight: The case demonstrates the necessity of robust editing workflows when incorporating AI-generated content. Automated tools or human reviewers should be tasked with identifying and removing non-speech elements, such as LLM instructions or placeholder text, before finalizing drafts for public use. This incident suggests that current practices may not adequately address the unique challenges posed by AI-assisted drafting.
- Broader Implications for AI in Politics: The incident reflects a growing trend of AI adoption in political communication, where the speed of integration often surpasses the development of accompanying safeguards. This case serves as a reminder that AI assistance must be balanced with human oversight to mitigate risks of miscommunication or misinformation.
Sources
Incident summary from the AI Incident Database, used under CC BY-SA 4.0.