Schwartz Reisman Urges Tiered AI Transparency Infrastructure for Canada
The Schwartz Reisman Institute for Technology and Society at the University of Toronto submitted nine recommendations to Innovation, Science and Economic Development Canada on September 23, 2026, answering Ottawa’s Enhancing trust in artificial intelligence through increased transparency consultation. Drawing on twenty affiliated researchers, SRI argued that disclosure without actors who can act is paperwork, and that Canada needs tiered research and audit access, confidential channels for dangerous capabilities, mandatory serious-incident reporting plus a voluntary near-miss track, and authenticity infrastructure that proves what is real rather than only hunting synthetics.
Filed under Policy and dated October 1, 2026, this AI4Canada briefing treats the SRI package as Canadian AI-governance infrastructure news distinct from yesterday’s Bell–Cisco sovereign stack MOU. Executive director Monique Crichlow framed transparency as information that creates accountability; director David Lie stressed consolidating failure patterns across deployments; and the submission’s final recommendation makes federal procurement the first customer for provenance, assurance and audit tooling—an angle that sits beside unfinished federal rulemaking after the Artificial Intelligence and Data Act died.
Why it matters: Canadian agencies already buy models whose logs, prompts and agent tools leave thin accountability trails. Structured transparency can close gaps—but only if access tiers, safe-harbour research rules and named incident owners travel with every disclosure mandate.
What it means in practice
Canadian privacy, procurement and counsel leads should inventory which AI systems would need tiered audit access under SRI’s design; demand named owners for serious-incident definitions and near-miss channels; assign an owner for CAISI and Canadian Centre for Cyber Security confidential capability notices; run time-boxed reviews of prediction-API leakage risks flagged by faculty affiliates; and prefer contracts that keep humans on irreversible model promotions. Place the submission beside Ottawa’s transparency consultation close and Solomon’s G7-G20 pragmatic-rules push.
Caveats come first. A research institute submission is not draft legislation; ISED may pick only fragments; and authenticity systems can lag deepfake volume. AI4Canada therefore presents the recommendations as directional policy context until published government responses appear.
What to watch next: whether ISED cites SRI’s two-channel incident regime; procurement language that buys assurance services; and any bill text that separates public disclosure from confidential research access. Readers can continue on the AI4Canada homepage, or browse the Newsroom for additional briefings.
Bottom line: treat this update as orientation, not instruction. Canadian AI transparency debate is moving from slogans toward information infrastructure and remains early. Organizations that benefit most will demand actionable access, keep humans on promotion gates, and refuse to confuse consultation replies with finished law.