Russell Buzby

When Canada ran the public consultation for its 2025-26 national AI strategy, 5,253 people took the trouble to write a submission. The government used AI to help summarise what they said, which is unremarkable now. On every measure agencies usually apply, the output held up. Then a researcher audited the process from the other end, tracing every participant forward into the summary rather than every claim back to a source. The result: across two policy topics, 16.9 per cent and 15.3 per cent of participants were effectively excluded from the summaries entirely, and the exclusions clustered. Participants expressing dissent or scepticism about AI were dropped at rates between 33 and 88 per cent (Mahajan, 2026).

The detail that should stop a policy audience mid-scroll is why nothing caught this. The study puts it precisely: a consultation summary can be factually grounded with every claim traceable to a real participant. It can be free of hallucinations and clearly reasoned. It can still leave 88 per cent of a dissenting cluster unrepresented. Every quality gate we currently apply to government AI, the explainability checks, the grounding checks, the hallucination detection, examines the output. This failure lives in the input relation, in who the summary is faithful to, and the paper argues that no existing evaluation framework addresses it because they measure output quality rather than input fidelity (Mahajan, 2026). Both of Canada’s official summaries actually performed worse on population coverage than a baseline built from randomly selected participants. Accurate and quotable, the summaries were still unfaithful to the people they claimed to compress, and the paper found that brevity, rhetorical register and distance from the semantic mainstream each independently predicted a submission being dropped. Disagreement has a writing style, in other words. The summarisation pipeline filtered the style.

Australia should treat this as directly relevant, because the machinery it describes is the machinery we are building. The government’s AI agenda was everywhere in early July. Michelle Grattan described AI’s opportunities and risks as front and centre for the Albanese government, with the National AI Plan of December 2025 setting a decentralised course: ministers and regulators carry AI within their own portfolios, an approach interpreted as light touch (Grattan, 2026). Her column casts Assistant Minister Andrew Charlton as a strong promoter of the technology who is equally minded about strict conditions and social licence. Kimberlee Weatherall’s assessment the next day gave the other half: the government has woken up to the risks, and now needs more ambition and more implementation capacity behind frameworks that remain thin (Weatherall, 2026). Her yardstick is funding, with an AI Safety Institute at A$29.4 million over four years against roughly A$460 million for its UK counterpart.

A decentralised, light-touch architecture is not inherently wrong. It is inherently dependent on implementation quality, because there is no central gate doing the checking. Every portfolio, every agency, every consultation team makes its own tooling choices, and summarising public input is exactly the kind of low-profile back-office use that spreads without ever rising to ministerial attention. That is how the representational failure the Canadian audit found would arrive here: not through a controversial flagship system, but through a productivity tool in a policy team, adopted in good faith, working precisely as specified.

The pattern generalises beyond consultations. A study of AI in parliaments published in July examined tools already embedded in parliamentary work, from research assistance to amendment analysis to the processing of public input, and found the accountability machinery has not followed them in. It names four gaps: epistemic opacity, institutional attribution, procedural oversight, and democratic-legal contestation, and its sharpest observation is that politically relevant choices are being displaced into infrastructure, settled inside systems before any formal decision point where they could be contested (Who governs parliamentary AI?, 2026). A summary that drops the dissenters is the perfect specimen: by the time a minister reads the consultation report, the filtering has already happened, invisibly, upstream of every accountable decision.

None of this argues for keeping AI away from public input, and the counter-evidence deserves its place. A thematic analysis of stakeholder submissions to the EU AI Act consultation found the public’s concerns aligned substantively with the final regulatory rationale, a case where consultation genuinely informed the instrument (Public concerns about emotion recognition systems, 2026). Consultation can work, and at the volumes governments now receive, AI assistance is how reading everything stays feasible at all. The question is whether the assistance is governed as if representation matters. Here two technical findings from the same July digests are encouraging, because they say the fix sits in deployment design rather than in waiting for better models. An institutional red-teaming study found that changing deployment rules alone, holding the underlying models constant, shifted simulated safety outcomes by as much as 58 percentage points (Institutional red-teaming, 2026). A runtime assurance architecture for agentic systems cut its measured autonomy-risk exposure by roughly a third for 95 milliseconds of latency (Runtime assurance, 2026). And the Canadian study itself ships an open-source audit tool that measures exactly the fidelity failure it exposes (Mahajan, 2026). The controls exist. They are procurement line items and evaluation clauses, not research problems.

For agencies, the operational takeaway fits in one question: when AI touches public input, who verifies representational fidelity? And where would the answer show up? An accuracy check will not do it, since the Canadian summaries passed accuracy. It requires auditing coverage, measuring which participants the summary is close to and which it has abandoned, and treating a dissent-shaped hole as a defect equal in severity to a hallucination. I wrote separately about the AI audit gap, where most executives cannot evidence the compliance claims their organisations make, and this is that gap in its democratic form. I write separately about the transparency illusion in Australian government AI disclosure, and the two failures compound: a process that silences dissent, described by a transparency statement that nobody affected can use, produces a system that is unaccountable twice over.

The reason this matters more in the AI era than it did for human summarisers is scale with consistency. A tired policy officer misreads submissions randomly. A pipeline drops them systematically, the same style of voice, every consultation, every agency, with no one deciding it. Consultation exists so that disagreement can reach decision-makers in a form that could change the decision. A summary that filters the disagreement leaves government with something worse than silence: its own assumptions, played back in the public’s voice, with a participation number attached.

References

Grattan, M. (2026, July 9). Grattan on Friday: AI’s opportunities and risks front and centre on Albanese government’s agenda. The Conversation. https://theconversation.com/grattan-on-friday-ais-opportunities-and-risks-front-and-centre-on-albanese-governments-agenda-286043

Institutional red-teaming: Deployment rules, not just models, causally shape multi-agent AI safety. (2026). arXiv:2607.07695. https://arxiv.org/abs/2607.07695

Mahajan, S. (2026). Participatory provenance as representational auditing for AI-mediated public consultation.arXiv:2604.20711. https://arxiv.org/abs/2604.20711

Public concerns about emotion recognition systems: A thematic analysis. (2026). Digital Society. https://doi.org/10.1007/s44206-026-00272-4

Runtime assurance for enterprise agentic AI systems: A policy-gated control model with quantitative autonomy-risk scoring. (2026). World Journal of Advanced Research and Reviews, 31(1), 512-522. https://doi.org/10.30574/wjarr.2026.31.1.1872

Weatherall, K. (2026, July 10). Australia’s government has woken up to the risks of AI. More ambition is needed. The Conversation. https://theconversation.com/australias-government-has-woken-up-to-the-risks-of-ai-more-ambition-is-needed-287059

Who governs parliamentary AI? Accountability gaps in the institutional use of AI. (2026). AI Law Politics, 2(1), 94-108. https://doi.org/10.5709/alp-02.01.2026-07