The AI Divide: A New Fault Line Opening Before We Finished Closing the Old One

Earlier this week I had the privilege of moderating a panel at Digital Access Day, the Canadian Internet Society's annual gathering of thoughtleaders working on the issues shaping our networks and our lives. This year's theme was Access, Autonomy, and the Open Internet, and my session connected affordability and access to the infrastructure choices behind digital sovereignty. The premise was simple enough: when governments and institutions invest in sovereign cloud or domestic networks, who actually benefits? Do those investments translate into more secure and affordable connectivity for Canadians, or do they just shift who gets paid?

Joining me on stage were three people who have spent years thinking harder about these questions than almost anyone in the country:

• Shelley Robinson, Executive Director, National Capital FreeNet
• Joel Templeman, Executive Director, Internet Society, Manitoba Chapter
• Dr. Michael Geist, Canada Research Chair in Internet and E-commerce Law, University of Ottawa

It was a genuinely great conversation. I walked in expecting a familiar debate about wholesale rates and spectrum auctions. I walked out thinking about two divides instead of one.

The divide we still haven't closed

For those not familiar, the "digital divide" is the gap between Canadians who have affordable, reliable internet access and those who don't. We have been talking about it since the late 1990s. That is more than a quarter century of promises, programs, and pilot projects, and we still have not closed it.

Shelley put it in terms that are hard to shake. Some Canadians, she reminded us, are making a monthly choice between internet access and food or medicine. That is not a rhetorical flourish. That is a lived reality for people trying to file for benefits, do homework, book medical appointments, and apply for jobs, all of which now assume a connection you may not be able to afford.

Joel followed with an anecdote from Winnipeg that landed even harder. There is public housing in the city wired for DSL and cable at the building level. The infrastructure is literally in the walls. But if a tenant earned enough income to afford an ISP's retail plan, they would no longer qualify for the housing. The very people the building was designed to serve are structurally locked out of the connectivity it was built to deliver. If you want a perfect illustration of how policy can fail quietly in the gap between two programs that never talk to each other, there it is.

Dr. Geist took the conversation into privacy and digital sovereignty, and this is where I started taking notes faster. He walked us through the two-pronged U.S. trade strategy currently being used to lock in American control over global data flows, a set of moves with direct consequences for what "digital sovereignty" can even mean for a country like Canada. His most memorable point, at least for me, was on de-anonymization at scale. He called back to the work of Professor Latanya Sweeney, who took anonymized hospital visit records, cross-referenced them with voting records in the state of Washington, and was able to match those records to specific individuals 43 per cent of the time. That was a human researcher with public data. Now picture a capable AI model doing the same job, across every "anonymized" dataset a government or hospital has ever released. And if we redact the records aggressively enough that an AI cannot reconstruct them, are they still useful for the research they were released to enable in the first place? That is the knot privacy policy has to untie, and we are not close.

The Connecting Families problem

This is where the conversation got concrete. ISED runs the Connecting Families Initiative, which on paper is exactly the kind of program you would want: deeply discounted home internet for low-income families. In practice it has a design flaw that excludes a whole class of providers.

The program caps the subsidized plan price at $20 per month. For a large incumbent with its own last-mile infrastructure, that math works. For a smaller provider like National Capital FreeNet, which has to pay wholesale access fees to reach a customer's home, $20 is often less than the wholesale cost itself. You cannot participate in the program without losing money on every subscriber. So community-focused ISPs, the ones whose entire mission is affordable access, are shut out of the program designed to deliver affordable access.

The panel floated a better model: turn the subsidy into a portable credit. The customer qualifies, the customer picks any participating ISP, the ISP provides service and redeems the credit from ISED. It puts the discount in the hands of the household instead of the provider, it lets smaller and community ISPs compete, and it aligns the market signal with the policy goal. We talked briefly about nationalizing the internet in Canada as an alternative, and there was rough consensus that we are too far down the market path for that to be realistic. A well-designed subsidy or credit system is a much more plausible route to universal access.

So after twenty-five years, our best current program cannot be used by the providers closest to the problem, and the fix is a relatively modest redesign that has not happened yet.

But wait, there's more: the AI divide

Here is the part of the conversation I did not see coming, and the part I cannot stop thinking about.

We have spent a quarter century trying and failing to close the digital divide. In that same window, AI has gone from science fiction to the infrastructure underneath your email, your hiring decisions, your loan application, and your kid's homework. GPT-3 landed in 2020. That is six years ago. ChatGPT is barely three. And in that tiny sliver of time, a second divide has opened up underneath the first.

Call it the AI divide. It is not just who has access to ChatGPT. It is who can afford the tools, who has the digital literacy to use them well, who works in industries being remade by them, who is being screened by an AI model they cannot see or challenge, and who has the capital and data to build with them.

Some numbers to anchor this. The World Economic Forum's 2023 Future of Jobs Report projects 69 million new jobs created and 83 million displaced between 2023 and 2027, a net loss concentrated in routine and lower-skilled roles. IBM's research points to AI spending exceeding USD $550 billion in 2024, with a projected AI talent gap of around 50 per cent. A separate survey found that 44 per cent of companies using or planning to use AI anticipated layoffs in 2024. The productivity gains from AI accrue disproportionately to high-skilled workers and capital owners, which means wage inequality is not a side effect of this transition, it is the default outcome unless we actively design against it.

Layer on the educational piece. Access to AI-related education is already uneven. Basic digital literacy, which we still have not universalized, is the floor you need before AI literacy is even possible. If you cannot get online affordably, you cannot learn to use the tools that are about to decide whether you get the job, the loan, or the benefit.

What the first divide should have taught us

So what can we do about this? The first honest answer is: look at what we got wrong on the digital divide, because the AI divide is running the same playbook at ten times the speed.

Lesson one: access without affordability is not access. Wiring every apartment building in Canada does not matter if the plans cost more than the household can pay. The same logic applies to AI tools. "Free tier" access to a model that gates the actually useful capabilities behind a subscription is the AI version of building fibre past the house and charging $90 a month to light it up.

Lesson two: programs designed around incumbents exclude the providers closest to the problem. Connecting Families excludes National Capital FreeNet because the price cap is set below wholesale cost. If we design AI access programs around the three or four hyperscalers, we will exclude the open-source projects, public interest labs, and small Canadian firms that could actually serve underserved communities. A portable-credit model for AI access (think subsidized API credits or compute allowances that follow the user, not the vendor) would avoid repeating that mistake.

Lesson three: information bubbles compound. Recommendation systems were already narrowing what people saw online. AI-driven personalization is going to deepen those bubbles and make them harder to notice. That has direct consequences for democratic participation at the exact moment governments are leaning more heavily on AI in service delivery and policy-making. Canadians on the wrong side of the AI divide will not just be economically disadvantaged, they will be increasingly excluded from the civic conversation.

Lesson four: speed matters. The digital divide gave us twenty-five years to get our act together and we still have not. The AI divide will not give us that long. The baseline level of disruption is already higher, and the rate of change is faster. A policy response that takes a decade to design, consult on, and pilot will be obsolete before it ships.

So what can we do about this?

A few concrete starting points, drawing from the panel and from the wider work happening internationally.

First, fix Connecting Families. Convert the subsidy into a portable credit that any qualified ISP can redeem. This unlocks participation by community networks and smaller CLECs, and it pressures large incumbents to compete on service rather than coast on a captive subsidized customer base.

Second, treat affordable connectivity as the prerequisite for AI policy, not a separate file. There is no AI inclusion strategy that works if a meaningful share of Canadians still cannot afford a reliable home connection. The two files need to be run together, with shared targets.

Third, invest in open-source and public AI. Mozilla's Public AI project is one model for this: AI tools, datasets, and infrastructure built in the public interest rather than extracted from it. A Canadian version, or meaningful Canadian participation in the international effort, would give our researchers, non-profits, and smaller firms a foundation they do not have to rent from a US hyperscaler.

Fourth, build real AI literacy into the education system, starting now and starting with adults as well as students. Retraining at scale is the only credible answer to the 83-million-jobs-displaced number, and it has to be accessible to the people most at risk, not only the people already on the inside.

Fifth, support "AI for Good" work that uses these tools to close gaps rather than widen them, and be deliberate about funding projects with measurable distributional impact. Groups like UNICEF and research outfits covered in Wired and UNICEF's own digital impact work have been sounding this alarm for a while. We should be listening.

The bottom line

The bottom line is this. We spent twenty-five years not finishing the digital divide, and a second divide opened up underneath us while we were still arguing about wholesale rates. The people forced to choose between internet access and medicine are the same people who will be screened out of jobs by AI they cannot see, scored by models they cannot audit, and excluded from services that increasingly assume both connectivity and fluency.

We cannot afford another quarter century of good intentions and bad program design. Fix Connecting Families. Fund public AI. Make affordability and literacy a single policy file. And call out the programs that look like progress on paper but quietly exclude the providers and communities they were supposed to serve.

The building is wired. The question is who we let turn on the lights.