October 12, 2026 ·AI Assisted

Tailor this proposal to your city

Pick a city to fill in its name, library, provincial privacy law and cost figures. The link updates so you can share a version for that city.

Library Card AI: A Proposal for Public AI Access in [City]

This proposal asks [City] to pilot free, metered access to leading AI models through the public library card, delivered through a gateway run by [the public library] that can later carry sovereign Canadian models to every resident.

Cardholders: 100,000
Expected cost / yr: $120,000
Cost ceiling / yr: $300,000

The gap

Capable AI now sits behind a monthly subscription, and the residents who would gain most from it are the least likely to pay for one. Job seekers, newcomers, students, seniors and sole proprietors get the weaker free tier, usually in exchange for their data.

Canada is behind on this. KPMG research ranks Canada 44th of 47 countries in AI literacy and training (KPMG). The federal strategy released June 4, 2026, AI for All, names public libraries as AI literacy partners, but analysts note no funding is earmarked for that role (Cultural Policy Hub).

Libraries solved this problem once already. They bought books, then internet access, then databases and e-books, and shared them through one card. AI is the next shared resource, and [the public library] already has the three things it needs: an identity system, public trust on privacy, and staff who teach people to use information tools.

Phase 1: AI on the library card

Every [the public library] cardholder gets a free monthly allowance of AI use, signed in with a card number and PIN, from home or in a branch.

A library-run chat page, not a vendor account. Residents never hand an email, phone number or payment card to an AI company.
Paid-tier models from at least two vendors, such as Claude, ChatGPT and Gemini, bought under metered enterprise contracts.
A visible monthly allowance. Each card has a usage budget, $5 in the illustration below, that resets monthly. When it runs out, the resident drops to a cheaper model until the reset.
Staff support. Branches run drop-in help and short workshops on using AI well and checking its answers.

Metered contracts are the key design choice. Consumer subscriptions are sold per person, and Toronto Public Library reports that no licensing model exists for shared public use (Urban Libraries Council). Buying by usage means the library pays only for what residents use and can cap the total.

How it works

One sign-in, two phases of models.

Every request passes through a gateway the library controls. The gateway is the only part [City] has to own, and it is why models can be added or swapped without residents changing anything.

Phase 2: a conduit for sovereign AI

The same gateway becomes the public distribution channel for any model a Canadian government releases or certifies. Adding a model is a configuration change, and every cardholder in [City] has it the same day.

Governments get three things they would otherwise have to build:

Reach. A sign-in that residents already hold, with no new app or account system.
A trusted proxy. The library sits between the resident and the model operator, so neither a vendor nor a government department sees who asked what.
A place to enforce standards. Admission to the gateway can require Canadian hosting, both official languages, audit logging, and the Canada Trusted AI Certification the federal strategy proposes.

The federal direction supports this. AI for All commits to a public supercomputer, sovereign compute under Canadian law, Canadian foundation model research, and trusted AI agents for all post-secondary students (Bennett Jones). Libraries extend that last commitment to everyone who is not a student.

No government has yet announced a sovereign model for public use. Phase 2 is an option the design keeps open, and Phase 1 does not depend on it.

Precedents

Three precedents show the demand, the delivery model and the policy backing. None yet offers remote, per-card access.

Precedent	What it did	What it leaves open
Toronto Public Library AI Upskilling Initiative (April 2025 to December 2026)	Free generative AI tools at 40 sites, plus training. Funded by $2.8 million from Google.org. 95% of surveyed participants reported more confidence with AI.	In-branch only, funded by one vendor's charity, ends December 2026.
Public AI Libraries Project (US, 2025 to 2026)	A cloud AI kiosk on an existing library computer, with a private account per patron. Libraries pay $1,000 per kiosk per year. Six active partner libraries; partners include the Library of Congress.	In-branch kiosk only, no home access by card.
AI for All, Canada's national AI strategy (June 4, 2026)	Names public libraries as AI literacy partners and commits to sovereign compute and Canadian foundation models.	No library funding earmarked; no public channel for sovereign models.

This proposal adds what is missing: access from anywhere on the card, a budget per resident, several vendors, and a route for sovereign models.

Cost

Model cost is capped by design: annual cost = active users × monthly allowance × share of allowance used × 12. The library sets the allowance to fit the budget, and spending cannot exceed the ceiling.

The figures below are illustrative, not vendor quotes. Choose a city above, or replace them with the library's cardholder count and contract pricing.

Input	Illustrative value
Cardholders	100,000
Active AI users per month	5,000 (5%)
Allowance per user per month	$5 of metered usage
Share of allowance used	40%
Expected model cost per year	$120,000
Ceiling on model cost per year	$300,000

Costs outside the formula need quotes from the library: gateway hosting and maintenance, a privacy impact assessment, staff training, and a program coordinator. The gateway itself need not be built from scratch; the US Public AI Libraries Project publishes its kiosk software as open source.

Candidate funding: the library operating budget, provincial library grants, the federal National AI Literacy Initiative once funded, and vendor credits accepted only under a multi-vendor rule.

Privacy, safety and governance

The gateway applies library privacy practice to AI: the library knows a card was used, and no one keeps a record of what the resident asked.

Identity stays at the library. Requests reach the model without the card number or name attached.
No training on resident data. Contracts prohibit vendors from training on prompts and require minimal retention.
Counts, not content. The library stores usage totals per card for metering and does not store conversations by default.
Canadian data residency where a vendor offers it, and mandatory for sovereign models in Phase 2.
Age rules follow the library's existing internet-use policy, with guardian consent for children's cards.
Acceptable use relies on the vendors' built-in safety systems and the library's code of conduct.
Public reporting. The library publishes the model list, costs and usage statistics, and completes a privacy impact assessment before launch.

The library board owns the policy. In Ontario, library boards are subject to MFIPPA; other provinces have equivalent municipal privacy law, and legal review should confirm the fit in [City].

Pilot

A 12-month pilot in three stages tests demand and cost before any permanent commitment.

Stage	Months	What happens	Gate to continue
Prepare	1 to 3	Privacy impact assessment, two-vendor procurement, gateway set-up, staff training	Board approves the assessment and acceptable-use policy
Limited launch	4 to 9	Two or three branches plus home access for a capped first cohort of cardholders	Cost per active user within budget; no serious privacy incident
Open access	10 to 12	All cardholders, with the allowance sized to the budget	Evaluation report to [City] Council with a continue, change or stop recommendation

The evaluation reports six measures, with targets set by the library before launch:

Monthly active users, and the share who had no paid AI access before
Model cost per active user
Tasks residents report completing, such as job applications, forms and coursework
New library card sign-ups attributed to the service
Staff hours spent on support
Privacy or safety incidents

Objections

The strongest objection is that free AI tiers already exist, and the pilot is designed to test it.

Objection	Response
Free tiers already exist.	They are weaker, rate-limited, tied to a personal account, and often paid for with user data. If free tiers close the gap, the pilot's survey of prior access will show it and Council can stop.
Costs will run away.	Spending is capped per card and in total. The allowance shrinks before the budget grows.
Public money goes to foreign tech firms.	So do database and e-book licences. The gateway keeps the library free to switch vendors and to move to Canadian models as they arrive.
AI gives wrong answers.	It does, and residents are using it anyway. The library is the one provider that pairs access with teaching people to check it.
The sovereign model may never come.	Phase 1 stands alone. Phase 2 costs nothing until a model exists.
Cards will be shared or abused.	The loss is limited to one card's monthly allowance, and the existing code of conduct applies.

The ask

[City] Council and the [the public library] Board are asked to approve four steps:

Direct library and IT staff to scope the 12-month pilot and report back within 90 days with a budget.
Authorise a privacy impact assessment and a two-vendor procurement for metered AI access.
Set a pilot budget ceiling of $300,000 for model costs, with the per-card allowance sized to fit it.
Write to the provincial and federal ministers responsible for AI, offering [the public library] as a pilot channel for sovereign models and requesting funding under the National AI Literacy Initiative.

A shared gateway specification, adopted by libraries neighbouring [City], would turn one city's pilot into a national distribution layer.