SRI submits recommendations to federal consultation on AI transparency
A new submission from the Schwartz Reisman Institute argues that transparency should be treated as information infrastructure and sets out nine recommendations for the Government of Canada. The submission draws on the work of 20 SRI-affiliated researchers across computer science, law, psychology, economics, and philosophy.
On September 23, 2026, the Schwartz Reisman Institute for Technology and Society submitted its response to Innovation, Science and Economic Development Canada’s (ISED) consultation, Enhancing trust in artificial intelligence through increased transparency. The consultation asked for input across five areas: AI-generated content, AI interaction, information about AI systems, AI incidents, and AI agents.
SRI’s submission begins from a contrarian premise. Canada could require every AI system in the country to publish detailed documentation tomorrow and be no better able to govern AI than it is today. Disclosure that reaches no one with the authority or capacity to act on it is not transparency—it is paperwork.
“Transparency is not just about generating access to information. It’s about providing actors in an ecosystem with information that can create positive action and accountability,” says SRI Executive Director Monique Crichlow. “This consultation by ISED is an opportunity for Canada to examine the type of governance infrastructure it needs to lead in responsible AI development.”
Two design principles
The submission proposes that information should be matched to recipient and purpose. Individuals, adopting organizations, and oversight bodies need different evidence, and no single disclosure document serves all three. It also argues that responsibility should follow control, understood as practical capability rather than formal authority—recognizing that control in AI supply chains is rarely singular.
That second principle has a direct consequence. Where the party holding relevant information is not the party able to act on it, obligations should be allocated across both, rather than assigned to whichever is easiest to identify.
Access, not just disclosure
Several recommendations turn on a distinction between public disclosure and structured access. Research by SRI Faculty Affiliate Thomas Ristenpart (Department of Computer Science) shows that confidence scores returned through a prediction interface can be used to reconstruct the data a model was trained on, and that repeated queries to a prediction API can recover a functionally equivalent copy of a proprietary model. Even a minimal interface can expose more than it appears to.
The submission's response is not to withhold information but to calibrate who receives it. It recommends a tiered program of confidential research and audit access for qualified researchers, auditors, and regulators, with safe-harbour protection for good-faith research—and a separate confidential channel allowing developers to disclose dangerous capabilities to CAISI and the Canadian Centre for Cyber Security without forcing unsafe public disclosure.
Work by SRI Faculty Affiliate Nicolas Papernot (Department of Electrical & Computer Engineering), who recently gave testimony to France’s Conseil de l’intelligence artificielle et du numérique, informs the submission’s argument that technical safeguards can be circumvented and that AI systems operate within contexts that model-level controls cannot fully anticipate.
Incidents, agents, and authenticity
On AI incidents, the submission notes that Canada lacks a consolidated, cross-sector picture of what is going wrong. It recommends a two-channel regime: mandatory reporting for defined serious incidents, and a protected voluntary channel for near misses. Research by SRI Director David Lie (Department of Electrical & Computer Engineering) on system security underpins the argument that incidents often become visible only through patterns across deployments.
“While individual organizations can see their own failures, no one can see the pattern unless we report and consolidate,” observes Lie. “That is the gap a reporting regime is meant to close, and it only closes if reporting produces analysis rather than just a laundry list of items.”
On AI-generated content, the submission argues Canada should prioritize authenticating what is real rather than only detecting what is synthetic. In a recent SRI white paper on AI and trust, SRI Research Lead Beth Coleman (Institute of Communication, Culture, Information and Technology) and co-authors note a new failure mode unique to the proliferation of generative AI tools: as synthetic content becomes common, authentic damaging material can be dismissed as fabricated.
An opportunity as well as an obligation
The submission also frames transparency requirements as demand signals rather than compliance costs alone. Provenance infrastructure, evaluation and assurance services, audit tooling, and incident analysis all depend on capabilities someone has to build—and Canada has an unusual concentration of relevant research strength. Drawing on work by SRI Faculty Affiliate Gillian Hadfield (Johns Hopkins University) on regulatory markets, the submission notes that a domestic market for independent verification emerges only when buyers are obliged to purchase what it provides.
The submission's final recommendation makes federal procurement that first customer.
Want to learn more?
Browse stories by tag:
- AI
- AI safety
- AI trust
- Computer science
- Computer security
- Copyright
- Cybersecurity
- Data
- Democracy
- Economics
- Education
- Engineering
- Ethics
- GPO-AI
- Governance
- Health
- Human Rights
- In the Media
- Jobs
- LLMs
- Law
- Normativity
- Philosophy
- Political Science
- Privacy
- Privacy Series
- Psychology
- Public Policy
- Recommenders
- Regulation
- Religion
- Reports
- Trust
- Workshops
