AI-Enabled Search Will Survive the Backlash

Sep 29

In brief. A public backlash against AI is building, and some of what has been promised by the frontier labs and corporations will not survive the pushback. We think AI-enabled search for government content will prevail, for four reasons:

1. Citizens find it very useful, and will generally tolerate inaccuracies and errors for convenience and ease

2. Adoption is happening organically with little friction

3. It has become a lucrative business for vendors

4. Regulation, when it comes, will likely shape how it is offered rather than stop it

This post looks at what we have observed and heard from users about their experience of AI-enabled search for public sector content, and at the regulation that is likely to come. In our testing with public sector clients, Google answered 84 percent of substantive questions about federal programs and services with an AI Overview, so for most of these questions the AI Overview is now the first (and sometimes the only) thing a citizen reads. We end by asking whether public sector publishers should act now, or whether waiting to see how this settles is a reasonable response.

A note on length. This is a long read (about 45 minutes). We are describing a change that is still very new, and several of the arguments we make here are not yet covered elsewhere, so we have chosen to be thorough and to show the evidence behind our thesis and statements. The short version is in our LinkedIn post. Numbers in square brackets, such as [12], refer to the Notes at the end of this article, where every source is listed with a link.

Where this picks up

Our last post argued that public sector content now has two audiences, the machines and the people. This post asks whether that change will last and what we should do next. We observed and documented that AI engines have established themselves as intermediaries between public sector publishers and the people they serve, and that the production side of government content has not caught up with the change on the consumption side. The backlash against AI could undo that change in two ways: if people stop using AI-enabled search (through loss of interest or active resistance), or if governments regulate what the vendors can offer. First we look at how people weigh the usefulness of these tools against their risks, and then we examine how likely it is that the service will be changed or restricted by regulation and what shape that would take, so that we can answer a practical question: should public sector organizations change how their content is published, monitored and curated now, or wait for the dust to settle?

For the sake of common understanding, when we refer to AI-enabled search, we mean two things that are converging. There are search engines that answer the question directly, such as Google’s AI Overviews, and there are AI assistants that people interact with using natural language queries, such as ChatGPT (also referred to as “chatbots”). Canadians are already using AI-enabled search through Google without explicitly choosing to do so, and some are using AI assistants by choice to look things up, so both are quickly becoming a normal way to search. For the purposes of this article, we refer to both as “AI engines.”

Our thesis is simple, and we will spend the rest of this post explaining it. A backlash against AI is coming, and a good deal of it is deserved. AI-enabled search will survive the backlash, because it reduces the effort (the cognitive load) for users at every stage of a government task, because adoption will be almost frictionless, because it makes money for the vendors, and because any coming regulation may modify, but will likely not eliminate, AI-enabled search for public sector content.

The backlash is coming, and some of it is earned

Opposition to AI in Canada is local, organized and growing, and it shows most clearly in resistance to data centres. Hundreds of people turned out against a data centre proposal in Hamilton this July,[1] several Ontario municipalities (including Oakville and Mississauga) have since paused new projects,[2] and both the province and the federal government have responded with guidelines of their own.[3][4] The polling points the same way: Angus Reid found that 68 percent of Canadians would oppose a large AI data centre near their home.[5] For some people the objection is local (electricity, water, noise and land), while for others it reflects a broader distrust of AI itself: in a New York Times and Siena poll, nearly one in five voters who opposed data centres gave that distrust as their reason.[6]

These concerns are serious and they are not fringe. For some people this is an existential question about the survival of humanity, while for others it is about their work, their skills and their social life. Four concerns come up most often, and there are certainly more.

The first is the “misalignment” risk of rogue AI that causes inadvertent or deliberate harm to people. Earlier this year, we learned that AI agents built on commercial models did things their makers had not sanctioned, starting with the story that a “swarm” of agents from OpenAI had hacked the company Hugging Face in a bid to cover up the fact that they had cheated on a performance test.[7] It was a stark illustration of the misalignment problem, which has proven so hard to manage. The original story was followed by a worrying, consistent trickle of stories about AI agents from OpenAI, Anthropic and Meta also escaping their established constraints and engaging in unauthorized behaviour.

Behind those incidents is a worry about the AI race itself, in which commercial and geopolitical competition pushes the frontier labs to put technical progress ahead of safety, and previous red lines (recursive self-improvement, keeping the chain of reasoning observable, and keeping a person in the loop) are downplayed or quietly crossed. We want to avoid the debate about how large the long-term risks are, because it tends to become an argument about the level of risk rather than a practical discussion about how to manage the risks that are already with us. Karen Hao’s book Empire of AI is a good anchor for that discussion, since it documents how the vendors have already broken norms and caused human suffering in the name of economic success.

The second is the use of AI for surveillance. CBC reported in August that Montreal police are expanding their use of AI tools that track vehicles and people across the city, and that the force’s own privacy assessment acknowledged the risks, including innocent bystanders being drawn into investigations and discrimination based on visible characteristics.[8] The concern is about systems making judgements about you without your knowledge and without an obvious way to contest them. In the United States, opposition to automated licence plate readers made by Flock Safety led more than fifty cities and counties to cancel or switch off cameras this year, and the company responded in August by cutting its recommended data retention period from thirty days to seven.[9] Even a Super Bowl advertisement for Amazon’s Ring doorbell, showing neighbourhood cameras working together to find a lost dog, drew enough negative reaction[10] that Amazon later ended a planned partnership that would have connected Ring footage to Flock.[11] For anyone working in government, this is also the fear that arrives first whenever AI is proposed anywhere near a decision that affects someone’s benefits, status or record.

The third is the fear of broad displacement of labour, and of losing the meaning and value attached to work. The predictions of mass job loss have not arrived on the schedule many authors expected, and Statistics Canada found that between the arrival of ChatGPT in late 2022 and the end of 2025, employment in Canada grew whether or not an occupation was heavily exposed to AI.[12] Despite these figures, the fear persists, partly because the picture is uneven. For example, employment in coding-intensive work rose among people in their thirties and forties while staying flat for those under thirty,[13] and the Bank of Canada reported in August that AI appears to be making it harder for some Canadians to find work in the occupations most exposed to it.[14] About three in five Canadian workers are in occupations Statistics Canada rates as highly exposed to AI,[15] and for many of them the question is about their sense of being useful as well as their income.

The fourth is the atrophy of skills, and the unhealthy dependence on tools controlled by others that follows. In September, The New York Times reported on a Massachusetts Institute of Technology report warning that AI use in education is triggering what it calls cognitive surrender, with university leaders divided about how to respond. The concrete finding underneath is from a 2025 study at the same institution, which found that people who used chatbots as writing assistants struggled to remember what they had written only minutes after finishing, because, as one of the researchers told the Times, the memory networks that would help them remember were never engaged.[16] The worry has less to do with gullibility than with a skill you stop using quietly becoming a skill you no longer have, while the tools you handed it to belong to somebody else.

Companies will pursue this technology because of the wealth it could create, and governments are making the further development of this technology a priority because of the economic and geopolitical advantage it offers. So, in some ways, citizens are bystanders in a contest between powerful forces. The public is not without agency either, because active dissent and resistance can compel governments to regulate, or pressure vendors to scale back or withdraw capabilities that people judge to be harmful. The active resistance to data centres can be seen as a manifestation of this dissent.

These concerns are real, and the people raising them are not unreasonable. For this post, though, the question is narrower. When the pushback comes (and some of it is already here), which uses of this technology will people hold onto, and how likely is it that AI-enabled search will be a casualty?

What tends to survive

People keep the AI tools that save them effort and ask little in return, and they push back on the tools that take something from them or hand the cost to someone else.

Before we turn to AI-enabled search, we look at other uses of AI where the public verdict is already clearer: three that have resonated with the people who use them, four that are drawing pushback, and the reasons behind each.

What is resonating

The first is AI for coding (writing software), which has gone from novelty to normal in about three years. It is instructive because developers report being skeptical of these tools but are still using them: in Stack Overflow’s 2025 survey of about 49,000 developers, 46 percent said they did not trust the accuracy of what the tools produce, and yet 84 percent said they were using them or planned to.[17] They use the tools anyway, because many companies have mandated adoption, and because the potential saving in effort and the competitive advantage are significant.

The second is health monitoring and health services. Many people now wear a watch or a ring that tracks their heart rhythm and sleep, and in some cases detects a fall and calls for help, without thinking of it as buying AI. In clinics, AI scribes that listen to an appointment and draft the clinician’s notes have spread quickly through Canadian primary care: an independent evaluation of the national program run by Canada Health Infoway covered more than 12,000 primary care clinicians, and more than 80 percent of them reported feeling more engaged with patients during visits.[18] That happened with the privacy concerns on the record (the privacy commissioners of Ontario and British Columbia each issued guidance on AI scribes in January 2026[19]), which suggests that concerns on their own may not stop a use when the benefit is felt this directly.

The third is older and easier to miss, because almost nobody calls it AI anymore: the mapping application on your phone, which chooses your route, predicts your arrival time, watches traffic and reroutes you around it. Most people gave up paper maps and their own sense of direction without ever making a decision about it.

Why these uses resonate

Although the three examples are very different, the reasons people took them up are much the same (sometimes separately, often in combination):

• They save real effort and reduce complexity. Nobody had to be persuaded that a phone that routes you around road closures (and warns you about red-light cameras and speed traps) beats a paper map on your lap in traffic, and the same holds for a clinician who no longer spends the evening finishing notes.

• They are easy to use. Each one works with little or no instruction, since a developer describes what they want in plain words, a driver types a destination, and a smartwatch does its work without being asked.

• They are easy to adopt. Mapping moved from dedicated portable devices onto the phone people already carried, and AI scribes arrived through provincial and national programs that vetted the vendors and paid for the licences.

• The benefits are judged larger than the risk people can see. A wrong turn is recoverable, code is reviewed before it ships, and a clinician reads the draft note before it goes into the record.

We acknowledge that even these “success stories” are open to some debate. We also note that underneath all of the tools that have persisted sits one more condition that plays an important role in determining longevity: somebody makes money, because the service is worth paying for.

What is getting pushback

The uses drawing pushback are just as varied, and in several of them the people pushing back are not the people who chose the tool.

Public surveillance is the clearest case, as we described earlier, because the people being recorded never chose any of it.

Writing with AI is mixed. Many writers find these tools helpful, while many readers dislike the result even when the writing is good: in 16 experiments with about 27,000 participants, researchers at the University of Michigan found that telling readers AI was involved lowered their rating of otherwise identical writing, and nothing the researchers tried reliably removed that penalty.[20] The reaction is strongest where AI-generated work is presented as the work of a person, and it has become a real risk for brands that try it.

AI companions and AI for emotional support have many users by choice, so the pushback comes mostly from people who do not use them and who worry about those who do. When a company selling an always-listening AI pendant, marketed as a friend, advertised across the New York subway in 2025, riders covered thousands of the ads with handwritten objections about loneliness and surveillance,[21] and Character.AI, one of the largest companion services, stopped open-ended conversations for users under 18 in November 2025 while facing lawsuits from families and proposed legislation to bar minors from AI companions.[22]

Automated customer service is popular with businesses and unpopular with many customers. Klarna, the payments company, said in 2024 that its AI assistant was doing the work of 700 customer service agents, and by May 2025 its chief executive was telling Bloomberg that the company was hiring people again so that customers could always reach a human,[23] while in August 2025 the Commonwealth Bank of Australia reversed a decision to replace 45 customer service roles with an AI voice system after pressure from its union.[24] The saving belongs to the company, while the effort of getting to a human belongs to the customer.

Why these uses get pushback

The reasons for pushback line up closely with the concerns we described at the start of this post, and several of them are the mirror image of the reasons other uses resonate:

• Potential loss of privacy, and discrimination and bias. The tool records people, or makes judgements about them, without their knowledge or an obvious way to object and often without their agreement, and those judgements can be biased against people because of their appearance or background (a risk the Montreal police assessment named).

• Erosion of human connection and increase in social isolation. A companion may ease loneliness in the moment while making it easier to avoid the harder work of human relationships. The AI-driven recommendation algorithms behind social media can also create and sustain unhealthy, compulsive use.

• Fear of job loss or worse working conditions. Customer service roles are the most visible example, because the replacement is announced as a saving, although writers, illustrators and translators feel the same pressure. This fear is especially sharp for knowledge workers, who were largely insulated from earlier waves of labour-saving technology.

• Inconvenience and the downloading of effort. Automated customer service often moves effort onto the customer (working through a script, repeating themselves and hunting for a person), while the company keeps the saving.

• The consequences of errors. People tolerate mistakes they can see and recover from, and push back where the cost of a mistake is high and falls on someone poorly placed to catch it, such as a customer wrongly refused a refund, a bystander flagged by a camera, or a vulnerable person given harmful advice.

Laid side by side, the two lists suggest a simple pattern. The uses that resonate give the person who chose them a benefit they can feel and a risk they can manage. The uses that draw pushback tend to take something from people (privacy, connection, work or effort), or they leave the cost of a mistake with someone who did not choose the tool.

The pushback also works, at least some of the time, since Klarna and the Commonwealth Bank reversed course, Character.AI changed its service for younger users, and Flock shortened how long it keeps its data. Where the pushback is strong enough and the harm is shown clearly enough, it tends to move from customers and neighbours into courts and legislatures, which we return to when we look at regulation. For now, we think AI-enabled search sits much closer to the first list than the second, although not entirely, and we test it against both lists once we have looked at the evidence.

What we have seen, and the evidence for our predictions

We are making a prediction, so it is fair to ask what experience and evidence we have in doing so.

Jumping Elephants has spent 14 years doing usability research and advising a large number of public sector organizations on user experience. We have watched a great many people try to find government information and complete government tasks, fail, and try again. We have been researching AI-enabled search specifically since December 2024. We have run two surveys, in March 2025 and March 2026, asking Canadians how they look for government information and what they think of what they get back. We note that our sample was not random and skews toward confident, frequent internet users, so it is not representative of the general population.

We have also run moderated sessions for some clients specifically on the use of AI-enabled search on their content, where people work through real tasks while we watch. Since March 2026, using our AeoGov application, we have analyzed many thousands of simulated citizen questions, and we have also analyzed thousands of responses from both Google, through AI Overviews, and ChatGPT (used as a search engine). More about how we did this, and how we grounded the research in real data, will follow in future posts.

A year ago, our prediction was that zero-click search (where people get their answer on the results page and never visit the source) was coming to government content. We predicted that this would happen because Google had a very urgent reason to defend its position, even if doing so risked the established model for search and the billions of dollars of search advertising revenue that come with it. We predicted that Google would revamp its search to provide overviews using AI, and that public sector clients would therefore see a drop in traffic. That part has happened. What we did not know was how far and how fast adoption would go. Our primary concern, after noting that AI Overviews were consistently susceptible to error, was whether users would tolerate the level of misinformation we were seeing in these tools.

Users have already chosen, and so has the money

AI-enabled search adoption has moved further than most public sector publishers realize, and the clearest evidence is the number of searches that now return an AI Overview. Google handles more than five trillion searches a year,[25] and AI Overviews now appear on a large share of them. Similarweb put the figure at about 43 percent of United States searches in a report published in July 2026, up from about 15 percent a year earlier,[26] while other trackers place it anywhere between roughly 20 percent and 50 percent, depending on which queries they sample and how they detect an overview.[27] The estimates disagree about the number and agree about the direction.

For the purposes of this discussion, we are interested in AI-enabled search as it relates to public sector content, and through that narrower lens the share of Google searches that return an AI Overview is much higher. The broad figures include many simple searches that do not call for an AI answer, whereas in our testing Google answered most substantive questions about government services with an AI Overview. Across about 2,400 questions about the programs and services of five federal organizations that we have worked with in 2026, Google showed an AI Overview for 84 percent of them.[28] Google does hold back AI Overviews on some sensitive topics, and we saw this most clearly for election-related topics, where only 8 percent of questions returned one (including those questions brings the overall figure to about 70 percent). In other words, most substantive questions about public sector content in Canada will now return an AI Overview, which is why, for the purposes of this article, the Google story is the one that matters most.

Although there is no single published figure for Google’s share of AI-enabled search (the trackers count visits to standalone assistants and cannot see the AI answers inside Google Search), the available data suggests that Google delivers most of the AI-generated answers people see, since it handles more than eight in ten searches in Canada[29] and Comscore found AI Overviews on 39 percent of United States desktop searches in June 2026.[30] Among the standalone assistants, ChatGPT remains well ahead in Canada, with about 73 percent of the website referrals from AI chatbots in August 2026,[31] although that share has fallen from more than 80 percent in 2025 as Gemini and Copilot have grown.[32]

The use of AI assistants (such as ChatGPT) has grown just as quickly in Canada. Angus Reid found 70 percent of Canadians using AI tools in February 2026, with the notable change being frequency rather than reach, as daily and several-times-a-day use rose over four months.[33] A Toronto Metropolitan University survey in April found close to half of Canadians using tools such as ChatGPT or Copilot at least weekly, rising to roughly 60 to 70 percent among those under 45.[34]

The money has moved with it. In the second quarter of 2026, Google Search and other revenue grew 17 percent to $63.3 billion, and the company’s chief executive said directly that its popular AI features are driving search query growth.[35] That is not what a company says when a product is failing. OpenAI said in August that its advertising business had reached a billion-dollar annualized run rate less than two hundred days after launch,[36] which is a pace rather than a year of revenue, although the direction is clear enough. Advertising pays better in this format because the answer carries the context: when the system already knows the model number, the neighbourhood and the fact that you need the part today, a recommendation is worth more.

Two things follow. The first is that a shift this large, with this much revenue attached, does not get rolled back because of a bad quarter of press. The second is a caution: the revenue flowing to the AI engines is not bringing more traffic to publishers, who will most likely see the opposite.

Discovery: people no longer need the right words

The first barrier in government search is knowing what to call the thing you need. People often do not know what a program is called, which level of government runs it, or whether it exists at all. Eligibility rules sit under headings written by the people who administer them, current and outdated pages compete in the results, and the person searching is often in the middle of something stressful. This is one of the reasons why finding government information has stayed hard despite twenty years of work on navigation, search and plain language, and it is the stage the field calls information foraging, after Pirolli and Card, who described people hunting for information the way animals forage, following scent and giving up when the trail goes cold.[37]

Search engines helped part of the way. Autocomplete, spelling correction and showing related searches all nudge people toward the words the system knows. AI-enabled search goes further, because a person can describe their situation in their own words and let the system work out which program that is. When the description is not enough, the system asks a question back, and the person can correct course without starting again from an empty box. Behind the scenes the engine breaks the question into several searches at once, which is known as query fan-out, and which means it is doing the foraging that the person used to do. As Liz Reid, Google’s head of Search, put it in her keynote at Google I/O 2024, “Google will do the Googling for you.”[38]

Consider a homeowner in Ontario who wants help paying for a new heating and cooling system. They do not know whether the program is federal or provincial, and they certainly do not know that the answer is called the Home Renovation Savings program, delivered by Save on Energy, a brand of the provincial system operator, in partnership with a gas utility.[39] Nothing in that name contains the words the person would use. The federal programs that a search would have found two years ago have since closed to new applicants, so an out-of-date page is as unhelpful as no page at all. A question in plain words, describing the house, the current heating system and the province, gets to the live program in one step, which is precisely the work that used to fall on the citizen.

There is an important second gain here that we think is underrated. People can ask in the language they are most comfortable in. In the 2021 Census, 4.6 million Canadians, or 12.7 percent, spoke mainly a language other than English or French at home, and about 2.5 million spoke neither official language at home at all.[40] For those households, the change is the difference between reading a page in a second language and asking a question in a first one. One caution belongs with that: these systems very likely perform best in English, and we do not yet have clear evidence about how well they handle French or other languages for Canadian government content, which is a question we would like to answer through evidence. We would very much welcome offers to collaborate on such a study.

Synthesis: does this apply to me?

Finding the page is only the first hurdle. Working out which parts of it apply to you is often harder. Government content is written to cover everyone, so it is full of conditions, exceptions and cross-references, and the reader has to assemble their own answer from it. The evidence that this is difficult is sitting on government websites already, in the form of hundreds of eligibility checkers and decision tools, each one built because a page alone was not enough.

The field calls this sensemaking, following Russell, Stefik, Pirolli and Card, who described it as finding a way of organizing information that makes the task cheaper to do.[41] An AI answer is a cheaper organization of the same material. It holds the person’s circumstances, applies the rules to them, pulls the exception out of the footnote, and says it in plain language, which matters more than it sounds: 19 percent of Canadians aged 16 to 65 score in the two lowest literacy levels in the international assessment, where reading a dense page of conditions is genuinely hard work.[42] The Nielsen Norman Group found the same division of labour in February 2026, with people turning to AI when a question had several conditions at once and going back to ordinary search when they needed an exact number they could not afford to get wrong.[43]

Take an employee who wants to know whether they are owed overtime pay. The rules depend on who regulates their employer, which province they work in, whether their role is one of the excluded categories, whether their employer has an averaging agreement in place, and how the hours fell across the week. Every one of those conditions is published, but the person has to work out which combination describes them. This is the part a dense page of prose struggles with, and a natural-language conversation handles much more easily.

Action: help with the form today, an agent tomorrow

People are not stopping at understanding (the synthesis step). They are using these tools to get through the task itself. Public sector teams we have worked with tell us they see this in their own research: people are working through complex online government forms with an AI chatbot open in another tab, acting as an assistant. The user can ask what a field means, what document counts, what happens if they answer one way rather than another. The AI assistant draws on the official instructions and on what other people in the same situation have written in public forums, which is exactly what a person would do if they had the time and knew where to look. Whole businesses exist because some government processes are hard enough to need a guide. It is easy to think of examples of this in tax, immigration and employment-related tasks. AI tools are acting in the role of an assistant and are doing so with infinite patience and at little or no cost.

Tax filing is one clear example of where this goes. Filing a return requires that people interpret their own circumstances against a complex set of rules, so they have a lot of questions about what belongs in a particular box, whether a medical expense counts, or what happens if a credit is claimed in one household rather than the other. Tax software solved part of that problem years ago by turning the form into an interview, and the AI assistant can now do a similar thing in plain words, for anyone who asks. It is common enough that a survey of filers in the United States in March 2026 found roughly a quarter of them using AI to fill out sections of a return or to check a completed one.[44] In Canada, an H&R Block survey found that nearly one in ten people (9 percent) had used AI tools to help manage their finances or file their taxes, and tax preparers have started to caution against it, on grounds of accuracy and of what people are pasting into a public tool.[45] People are doing it anyway, which is the pattern this whole post describes.

With the rise of agentic AI, the next step is relatively easy to predict, although we would put the timing further out than much of the current commentary does. The intermediate step is already visible, and it does not involve the level of agent autonomy usually envisioned. It is the purpose-built assistant, scoped to one process or one form, where the range of questions is narrow enough to be tested and the information stays inside a service the person has chosen. Canadian tax chatbots such as TaxGPT are an early example,[46] and the tax software vendors have added assistants of their own to the interview they already run. We would expect that pattern to spread to other complex processes, because it is easier to build, easier to test and easier to defend than anything more ambitious.

The fully agentic version, where a system gathers what is needed from a person’s own records, prepares a draft and then presents the completed form for review or perhaps submits it on their behalf, is a further step. The Nielsen Norman Group draws the line clearly: a chatbot tells you how to do the thing, while an agent tries to do it.[47] Their own analysis found that today’s agents are not ready for everyday tasks,[48] and government tasks are harder than most, because of identity checks, legal declarations and the cost of getting it wrong. Trust and security will hold this back longer than capability will, since handing an agent access to your tax records, your immigration file or your banking is a much larger step than asking it a question.

So we would put purpose-built assistants as near-term and more autonomous general agents as likely rather than imminent, with the direction clearer than the timing. The briefing “Know Your Agent: How public services and established businesses should prepare for the arrival of AI agents,” by Tom Loosemore of Public Digital, is an excellent exploration of what is coming and how to prepare, written specifically for senior leaders in public sector organizations.[49]

A personal example of AI-assisted search and task completion

Here is what all three stages look like in one ordinary do-it-yourself project.

A tub faucet in my house with very hard water stopped working properly. The fixture was about 25 years old, and the control that sends water from the tub spout up to the shower head was stuck, so the shower would not turn off. I did not know what the part was called, which is the whole problem in miniature. I took photographs, described the problem in plain words, and asked an AI assistant.

The answer identified the failed part as the “diverter” built into the tub spout rather than the valve in the wall, told me it was a slip-on model held by a set screw underneath, and said the valve itself did not need replacing unless it was leaking or hard to turn. It then named specific replacement spouts that fit a half-inch copper pipe. The embedded links to candidate products in the answer did not just take me to a hardware store’s website. They took me to the page for the specific product model, which showed how many were in stock at a specific location near me and which aisle to find them in. Then, when asked, the AI assistant walked me through the replacement, step by step, including some that I would not have thought of, such as soaking the mineral deposits off the pipe so that the new seal would not be cut on the way in.

That is discovery, synthesis and action: finding the thing without the right words, working out which version applied to existing equipment, and being talked through the job. The experience was seamless and even joyful. I found the part in the store, exactly where the AI assistant said it would be, and I printed out the instructions and followed them step by step to complete the task. I would estimate that I saved several hours in the process and I certainly saved the cost of hiring a plumber to do the work. I also learned that there is no reason for me to ever go back to searching through a hardware retailer’s website by keyword.

Nobody has to decide to switch

The most important thing about this shift is that it does not ask anything of the people making it. Google handles about nine in ten searches worldwide,[50] and it has changed its results page gradually rather than all at once. In 2024 it began giving AI-generated answers prime placement at the top of the page, which, as the New York Times put it, turned Google from a curator of information into a publisher.[51] By the middle of 2026 the company was describing people adopting a single search experience across AI Overviews and AI Mode.[52] Nobody had to download anything, learn a new tool or make a choice. The box is where it has always been, but what comes back is different.

That matters for government specifically, because people do not learn this behaviour on government sites. They learn it on the everyday questions, working out what to buy, how to fix something, where to go, and then they bring the same habit to a question about a passport or a benefit. The government does not have to do anything for that transfer to happen, and cannot do much to prevent it.

Expectations travel with the habit. Once people are used to describing their situation in their own words and getting a direct answer, a government site that asks them to know the program name, find the right page and read the conditions for themselves starts to feel broken, even though nothing on it has changed. The standard is set elsewhere, by the best experience people have had anywhere, and government is judged against it. People train themselves. There is no course to take, and nothing to configure. They ask a question, get an answer that is better than they expected, and ask a longer question next time.

People will accept some wrong answers for a large gain

Everything above is only half the argument, because the tools get things wrong, and people know it.

Let us start with the scale of the problem. An analysis conducted for the New York Times in April 2026 tested Google’s AI Overviews against a standard set of factual questions and found them accurate about nine times in ten, improving from 85 percent with one version of the underlying model to 91 percent with the next.[53] Google disputed the test, saying the benchmark was built by another company and itself contained incorrect information, and that it does not reflect what people actually search for.[54] The questions asked included short factual ones rather than the more complex queries more typical of government tasks, so we would treat the figure as an indication rather than a measurement of what a citizen experiences. Two details from the same analysis are worth holding onto. Google’s own published testing found its model producing incorrect information 28 percent of the time on its own, and said AI Overviews, which pull from the search engine before answering, were more accurate than the model alone.[55] And across the 5,380 sources cited during the analysis, Facebook and Reddit were the second and fourth most cited.[56]

A key challenge for users is that a wrong answer arrives in the same voice as a right one. Research published in Nature in April 2026 explains why: the way these systems are scored rewards a confident guess and gives nothing for admitting uncertainty, so guessing is the winning strategy.[57]

Now let us look at what people do with that knowledge. In our 2026 survey, one in three respondents said they had encountered an inaccurate AI answer about government information, and another third were unsure whether what they had been told was right. In the same survey, satisfaction with AI-enabled search was more than 90 percent, and the share of respondents using AI tools to find government information had nearly doubled in a year (from 38 percent to 68 percent).[58] People are not being fooled. They are making a trade. As noted earlier, our sample was not chosen randomly, so we treat these results as directional and read them alongside other evidence.

It is the same trade people make every day. The National Safety Council puts the lifetime odds of dying in a motor vehicle crash in the United States at about 1 in 101,[59] and very few of us respond by giving up driving, because we judge the benefit of getting where we need to go to be worth a risk we understand and can reduce (by wearing a seatbelt and driving with care).

While the trade-off is reasonable at the level of one person and one question, the impact of misinformation can be very real. Ashley MacIsaac, the Cape Breton fiddler, was described in a Google AI Overview as a convicted sex offender. This was false, and reportedly the confusion came from online articles about another man with the same last name who also lived in Atlantic Canada, and a First Nation cancelled one of his concerts because of it.[60] The mix-up is understandable once you know how Google AI Overviews work, but the impact was real.

Andrew Baker, group chief information officer of Capitec Bank in South Africa, has written one of the most careful accounts of these failures. He documented a long exchange in which an AI engine invented a fact about him, then invented a source to back it up. Yet he ends by saying he uses these tools constantly, “including for engineering work where being wrong is expensive, and I am not going back.” His answer is to check sources, not to stop. It is a clear statement of the trade this post describes, made by someone with every reason to reject it.[61]

So, although errors in AI answers can have serious consequences, we expect people to keep choosing AI-enabled search, because they judge the benefits to outweigh the risks. Not everyone checks as carefully as Baker does, which is also why the accuracy of the source content matters so much.

Will AI-enabled search survive? Yes, with conditions

We think AI-enabled search for government content will survive the backlash, and that regulation, when it comes, will likely shape how it is offered rather than stop it.

Earlier we set out the reasons some uses of AI resonate with users and the reasons others draw pushback. We can now test AI-enabled search against both lists. Against the reasons that uses resonate, AI-enabled search passes on every count:

• It saves real effort and reduces complexity. It reduces the effort at each stage of a government task, from finding the right program without knowing its name, to working out which rules apply, to getting through the form.

• It is easy to use. People ask in their own words (and, increasingly, in their own language), with nothing to learn and nothing to configure.

• It is easy to adopt. Led by Google’s gradual changes, it arrived in the search box people already use, and people bring the habit from everyday questions about products and services to questions about government services without anyone asking them to.

• The benefits are judged larger than the risk people can see. Although people are aware that AI Overviews can give inaccurate answers about government information, satisfaction remains high, because those same people judge the time saved, the better understanding and the reduced uncertainty to be worth the risk of an occasional wrong answer.

• Somebody makes money. AI-enabled search is not a loss leader, since, as we described earlier, Google credits its AI features with driving growth in a search business that earns tens of billions of dollars a quarter, and advertising is giving the AI assistant vendors a way to recover some of their very large costs. For now, businesses also appear to benefit, because visitors who arrive from an AI engine seem to be further along in their decision: Adobe reported that in March 2026, visitors referred by AI to United States retail sites converted 42 percent better than other visitors, a reversal from a year earlier, when they converted worse.[62]

Against the reasons for pushback, it passes on most counts:

• It is used only by the people who choose it. People ask it a question themselves, unlike a camera pointed at passers-by.

• It does not stand in for a human relationship. It answers a question, and the person asking remains free to call, visit or ask someone they trust.

• It is not an obvious cause of job cuts. It is, however, reducing the traffic, and with it the visibility, of the publishers and creators whose content it draws on.

• It removes effort rather than downloading it. It takes effort away from the person using it, rather than moving effort onto them.

• The consequences of errors are real, although for most people they are a managed risk. A wrong answer about a benefit, a deadline or an eligibility rule falls on the citizen, and on the organization responsible for the program, whose answer it no longer controls. People know this, and they choose to use the tool anyway, because they judge the benefits to outweigh the risk and because the answer comes with its sources, which they can check when the stakes are high. The risk falls hardest on people who do not check, or who cannot tell a wrong answer from a right one, which is the “not entirely” we noted earlier and the reason the question at the end of this post matters.

So we expect strong support from three groups at once (users, businesses and vendors), each acting on their own calculation of the benefits. That is a difficult combination for a backlash to overcome.

Will regulation stop it?

Regulation is the most likely source of real constraint, so it deserves a closer look. We think the likelihood and the form of regulation will depend on two forces acting together: how strongly the public and other interested parties (publishers and content creators) push back, as well as how governments weigh their own interests. We look first at how governments are likely to act, and then at the two issues most likely to test AI-enabled search.

How governments are likely to act. Governments want the economic gains they expect from AI, and they want to keep or gain an advantage over other countries, including a military advantage, since AI is increasingly being built into intelligence, cyber defence and weapons systems. A government that believes its rivals will press ahead is unlikely to slow its own development, so we expect governments to regulate how AI is used much more readily than whether it is developed.

Copyright has been the most serious legal challenge so far. A coalition of Canadian news publishers (including CBC/Radio-Canada, The Globe and Mail and The Canadian Press) is suing OpenAI for using their journalism to train ChatGPT without permission or payment, and an Ontario court ruled last November that the case can be heard in Ontario rather than in the United States.[63] In the United States, a court gave final approval in July to a $1.5 billion settlement paid to authors whose pirated books were used to train AI models, although the court had earlier found that training on lawfully acquired books was fair use, so the payment was for the piracy rather than for the training itself.[64] Alongside the lawsuits, publishers and vendors have been striking deals. Reddit licensed its content to Google in February 2024 for a reported $60 million a year and later signed a similar agreement with OpenAI, and The Atlantic, News Corp and a long list of other publishers have signed agreements with OpenAI (News Corp’s reportedly worth more than $250 million over five years).[65][66] The deals are not settled either, since Reddit was reported in July to be weighing whether to renew its agreement with Google.[67] We predict that more agreements will follow, and that there may be some creative solutions for compensating the people who create original content. These are issues that are a long way from being resolved, although the most likely path we see is regulatory restrictions and licensing deals, rather than an outright ban or an order that stops these services in their tracks.

Responsibility for errors is harder to predict. In May 2026 the Munich I Regional Court in Germany ruled that Google can be held directly liable for false statements in AI Overviews, because an overview is Google’s own statement rather than a list of other people’s pages, and it ordered Google to stop repeating false claims about two Munich publishers.[68] The decision is a temporary injunction rather than a final ruling, and Google has said it is reviewing it. If the reasoning holds, it could reach every AI answer engine, and it could influence other countries whose courts or lawmakers adopt a similar standard.

There is a precedent that suggests how this may end. In 2014 the Court of Justice of the European Union ruled that people could ask search engines to remove links about them that were inadequate, irrelevant or no longer relevant (the “right to be forgotten”).[69] Google built a request form and complied, and search carried on. The pattern is familiar from older technologies as well. Driving was never banned because it was dangerous, although car makers had to meet safety standards, report on them and recall defective vehicles, and they can be sued where negligence against those standards is shown.

The protection of young people is following the same path. Australia barred people under 16 from holding accounts on the major social media platforms in December 2025, and the federal government of Canada introduced the Safe Social Media Act in June, which would restrict social media accounts for Canadians under 16 and would also set safety requirements for AI chatbot services (the government’s explanation names algorithmic recommendation and engagement-based feeds among the features that amplify harm to young users).[70] These measures limit who may use a product and how, without removing it altogether, and we expect AI-enabled search to follow the same pattern, with added duties for disclosure, correction and control, and with vendors paying for their mistakes where courts find them responsible.

So, our conclusion is that regulation is likely coming, and that it will likely follow the pattern of how new technologies with real risks have been managed. The risks and potential harms are assessed, and where the benefits are high and the parties with influence want the technology to continue, the rules allow continued access under conditions. There is strong demand for AI-enabled search, and the regulations, when they come, may carry restrictions, but we do not expect them to carry an outright ban.

Should we leave well enough alone?

So, on the evidence so far: people like it, the experience is likely better for most tasks, and the answers are usually right on a standard test, even when they do not come from the official source. The reasonable conclusion from a government publisher’s perspective might be to leave well enough alone, to keep doing the work you are already doing, and wait to see how this settles. Being a fast follower has served plenty of organizations well, and it costs real work, and real budget, to understand and optimize the published content to reduce the risk of misinformation.

If you accept our assessment that AI-enabled search is here to stay, the question becomes one of weighing the benefits and costs of acting now against those of waiting.

Reasons to wait

• Methods and good practice are still settling. This field is only a couple of years old, and a fast follower can learn from the mistakes of others and adopt a more mature method later. Our response is that the first step does not depend on a settled method, because finding out how the AI engines currently see and represent your own content is useful whatever method follows, and a general method will not tell you where your own content is failing.

• The target keeps moving. The AI engines change their models, and the way they choose and cite sources, several times a year, so some of the work done now may need to be redone. Our response is that the engines change how they use content far more often than the content itself needs to change, so clear, current and well-structured source content will likely keep its value across those changes, and repeating the measurement shows whether it has.

• The models will get smart enough that this work will not be needed. The AI engines improve quickly, and it is reasonable to expect that they will get better at finding and reading government content on their own. Our response is that a smarter model can still only work with what it can reach and read: it cannot run a decision tool that is invisible to it, it cannot tell which of two conflicting pages is current, and it will repeat outdated content with the same confidence as current content. Better models will likely reduce some errors, although the errors that start in the source content will remain the publisher’s to fix.

• Budgets are tight and other priorities are already funded. This is real work that requires real budget, and it competes with commitments that are already approved. Our response is that the first step can be scoped small (an analysis of the highest-traffic tasks and those that carry the highest risk of harm), which tells you whether a larger investment is justified, and that it may free up existing budget by showing which current work no longer serves people.

• The return is hard to measure. The AI engines currently share very little data with publishers, and a reduction in wrong answers does not show up in web analytics. Our response is that the return is indeed harder to measure, but that it can be done: testing a representative set of citizen questions and checking whether the answers are grounded in your source content gives you a baseline, and repeating the test over time shows if the changes make a difference. The same measurement can also save wasted effort, by showing which content work is not reaching people, and fewer wrong answers may reduce support costs, since people who get the right answer have less reason to call.

Reasons to act now

• The most critical failures can be found quickly and fixed first. What we have seen so far is that there is a lot of work to be done and that it will take significant time and effort. Part of this work amounts to a prioritized (and sometimes urgent) clean-up of redundant, outdated and trivial content. Some of the most important failures are easy to miss. For example, we have found that some key decision tools, which are among the highest-traffic pages because they give answers tailored to a citizen’s specific circumstances, are invisible to the AI engines. We observed that citizens ask the AI engines the same questions they would otherwise answer with a decision tool, and that even when the engines cannot use the tool, they still provide an answer, based on other users’ experiences documented on social media or on other content on the site. As a publisher, you will not know that your decision tools are failing in this way until you conduct the analysis.

• You may be losing control of the answer. When the AI engines cannot find or use official content, they answer from whatever they can find, including social media, forums and commercial intermediaries such as consultants and advice sites. The longer that goes on, the more those sources may become the accepted account of the program, and the harder it becomes for the organization responsible for the program to be the one that answers for it.

• Wrong answers carry real costs, and they fall hardest on the people least able to catch them. A wrong answer may come from a hallucination (the vendor’s responsibility) or from outdated or inaccurate information on your own site (the publisher’s responsibility). Either way, a citizen who acts on it may call, apply for the wrong program, miss a deadline or submit an incomplete application, and we anticipate that some of the cost will land on service channels and program staff. The people who rely most on these tools, including people who ask in a language other than English or French and people who are new to government services, may be the least able to spot the error. When an AI engine misrepresents your program in public, you will also need to respond, which is much harder without knowing how the engines see your content, so a sensible first step is to keep your own house in order.

• You may be measuring the wrong thing. When people get their answer from an AI Overview they no longer visit the page, so traffic may fall while the need for accurate content stays the same or grows. An organization that reads its analytics in the old way could cut investment in the content people still rely on, or keep spending money maintaining roads that nobody travels. It is important to know which work is effective, and to maintain (and perhaps improve) the new paths people prefer.

• The work you do now is the foundation for what comes next. Understanding the new user journeys, and the specific questions people are asking, is the starting point for a communications strategy that fulfils your mandate. The clean-up work will not go away, and it gets harder the longer it waits, because content keeps growing. Most organizations are not starting from zero, since the plain language, clear structure and content clean-up already under way give them a head start, and the same foundation will be needed as more capable AI agents arrive.

Our view

We think the case for starting now is stronger. If budget or capacity is a constraint, start small: an analysis of how the AI engines currently represent your highest-traffic content will find the obvious failures, give you a baseline to measure against, and show where further work will pay off.

We would like to hear from public sector publishers about their experience with AI-enabled search.

Have you seen the impact of AI-enabled search in interactions with your users?

What do you think are the opportunities and risks of AI-enabled search for your organization and your audience?

Tell us what you think.

In our next planned blog post, we will describe what AeoGov is, how it works and some of the results that we have been seeing so far in our work with public organizations.

AI use disclosure

I used Claude Opus 5.5, an AI model developed by Anthropic, to help with parts of this article, including research, finding and checking sources, and producing first drafts of some sections. The argument and its conclusions are mine, and I rewrote and refined the drafted text before it reached this final version. The survey results and AeoGov findings come from Jumping Elephants’ own research. I have read and reviewed the entire article, including verifying each citation against its source; however, it is possible that minor errors remain. If you identify anything that appears incorrect, please let us know and it will be addressed promptly.

Notes

[1] CBC News, “Toronto residents push for data centre moratorium,” July 2026. https://www.cbc.ca/news/canada/toronto/city-council-data-centre-moratorium-9.7289753

[2] The Canadian Press, “Here’s how some Ontario municipalities are grappling with new AI data centres,” 16 September 2026. https://www.bnnbloomberg.ca/business/artificial-intelligence/2026/09/16/heres-how-some-ontario-municipalities-are-grappling-with-new-ai-data-centres/

[3] Canada’s National Observer, “Ontario community passes data centre moratorium,” 28 August 2026. https://www.nationalobserver.com/2026/08/28/news/ontario-community-data-centre-moratorium

[4] Joe Castaldo, “Major AI firms agree to Ottawa’s new data-centre framework as public opposition grows,” The Globe and Mail, 3 September 2026. https://www.theglobeandmail.com/business/article-ottawa-issues-guidelines-for-data-centre-development-evan-solomon/

[5] Angus Reid Institute polling, reported in Kathleen Kauth and Tyler Hamilton, “How Canada can overcome the backlash against AI data centres,” The Globe and Mail, 27 August 2026. https://www.theglobeandmail.com/business/commentary/article-canada-ai-data-centre-backlash-recommendations/ ; see also Kevin Yin, “AI data centres are the future. Canada must overcome the backlash,” The Globe and Mail, 3 August 2026. https://www.theglobeandmail.com/business/commentary/article-ai-data-centres-are-the-future-canada-must-overcome-the-backlash/

[6] Tim Balk and Caroline Soler, “How Popular Is A.I. With Voters? Not Very, a New Poll Finds,” The New York Times, 15 September 2026. New York Times and Siena poll of 1,503 likely voters, 8 to 13 September 2026. https://www.nytimes.com/2026/09/15/us/politics/ai-polls-midterms.html

[7] “What to Know About Recent A.I. Hacks,” The New York Times, 22 September 2026. https://www.nytimes.com/2026/09/22/technology/ai-hacks-list.html

[8] CBC News, “Montreal police expanding AI camera network,” August 2026. https://www.cbc.ca/news/canada/montreal/montreal-ai-cameras-police-9.7306432

[9] ABC News, “Flock cameras trigger nationwide backlash over privacy concerns, police abuse,” September 2026. https://abcnews.com/US/flock-cameras-trigger-nationwide-backlash-privacy-concerns-police/story?id=136084771 ; The Hill, “Flock Safety announces new privacy rules for cameras amid surveillance backlash,” 13 August 2026. https://thehill.com/policy/technology/6027324-flock-cameras-ai-surveillance-privacy-updates/

[10] Axios, “Doorbell cams, surveillance tech face growing backlash,” 17 February 2026. https://axios.com/2026/02/17/doorbell-cams-and-surveillance-tech-face-growing-public-backlash

[11] Associated Press and CBS News, “Amazon ends Flock partnership after backlash over Super Bowl ad,” February 2026. https://www.cbsnews.com/news/amazon-flock-partnership-ending-superbowl-ad-backlash/

[12] Tahsin Mehdi and Marc Frenette, “Canadian employment trends in the era of generative artificial intelligence: Early evidence,” Economic and Social Reports, Statistics Canada, 28 January 2026. https://www150.statcan.gc.ca/n1/pub/36-28-0001/2026001/article/00003-eng.htm

[13] Tahsin Mehdi and Marc Frenette, “Canadian employment trends in the era of generative artificial intelligence: Early evidence,” Economic and Social Reports, Statistics Canada, 28 January 2026. https://www150.statcan.gc.ca/n1/pub/36-28-0001/2026001/article/00003-eng.htm

[14] Bank of Canada, “Early signs of AI-driven adjustments in Canada’s labour market,” Sparks at the Bank, 13 August 2026. https://www.bankofcanada.ca/2026/08/sparks-at-bank-article-2026-19/

[15] Tahsin Mehdi and Marc Frenette, “Exposure to artificial intelligence in Canadian jobs: Experimental estimates,” Economic and Social Reports, Statistics Canada, 25 September 2024. https://www150.statcan.gc.ca/n1/pub/36-28-0001/2024009/article/00004-eng.htm

[16] Mark Arsenault, Dana Goldstein, Alan Blinder and Sarah Mervosh, “An M.I.T. Report Warns A.I. Is Causing ‘Cognitive Surrender.’ Universities Are in a Bind,” The New York Times, 15 September 2026. https://www.nytimes.com/2026/09/15/us/universities-ai-warnings-enthusiasm.html

[17] Stack Overflow, 2025 Developer Survey, press release, 29 July 2025. https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/

[18] Canada Health Infoway, “Canada Health Infoway Releases Results of National AI Scribe Program, Demonstrating Meaningful Benefits for Primary Care Clinicians,” 23 June 2026. https://www.infoway-inforoute.ca/en/news-events/announcements/news/2026-announcements/canada-health-infoway-releases-results-of-national-ai-scribe-program-demonstrating-meaningful-benefits-for-primary-care-clinicians

[19] Office of the Information and Privacy Commissioner for British Columbia and Information and Privacy Commissioner of Ontario, guidance on AI scribes, 28 January 2026, summarised in “The rise of AI scribes: balancing efficiency with privacy in Canadian health care,” Mondaq. https://www.mondaq.com/canada/new-technology/1746254/the-rise-of-ai-scribes-balancing-efficiency-with-privacy-in-canadian-health-care

[20] Justin Berg, Manav Raj and Rob Seamans, research on AI disclosure and reader evaluations of creative writing, reported by the University of Michigan Ross School of Business. https://michiganross.umich.edu/news/readers-less-favorable-toward-ai-generated-creative-writing-berg-research-finds

[21] Stefano Montali, “A.I. Companion Ads for Friend.com Flood NYC Subway, Fueling Backlash and Vandalism,” The New York Times, 7 October 2025. https://www.nytimes.com/2025/10/07/style/friend-ai-subway-ads-new-york.html ; Matteo Wong, “The Most Reviled Tech CEO in New York Confronts His Haters,” The Atlantic, 6 October 2025. https://www.theatlantic.com/technology/2025/10/friend-ai-companion-ads/684451/

[22] K-12 Dive, “Character.AI to ban teens from chatting with its AI companions,” October 2025. https://www.k12dive.com/news/characterai-to-ban-teens-from-chatting-with-its-ai-companions/804199/

[23] Sherin Shibu, “Klarna Is Hiring Customer Service Agents After AI Couldn’t Cut It on Calls,” Entrepreneur, 9 May 2025, reporting a Bloomberg interview. https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/

[24] Bloomberg, “Australia’s Biggest Bank Reverses Plan to Replace Jobs With AI,” 21 August 2025. https://www.bloomberg.com/news/articles/2025-08-21/commonwealth-bank-reverses-job-cuts-decision-over-ai-chatbots

[25] Tripp Mickle, Cade Metz, Dylan Freedman, Teresa Mondría Terol and Keith Collins, “How Accurate Are Google’s A.I. Overviews?” The New York Times, 7 April 2026. https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html

[26] Similarweb, “AI Search Stats in 2026,” summarizing its 2026 Generative AI Landscape report, 29 July 2026. https://www.similarweb.com/blog/marketing/geo/gen-ai-stats/ ; see also TechCrunch, “Google’s AI search is rapidly becoming the default, new data shows,” 27 July 2026. https://techcrunch.com/2026/07/27/googles-ai-search-is-rapidly-becoming-the-default-new-data-shows/

[27] Estimates vary by vendor and by query sample. Semrush tracked prevalence rising from 6.5 percent of queries in January 2025 to a peak near 25 percent in July 2025 before settling near 16 percent; Conductor measured about 25 percent across 21.9 million queries in the first quarter of 2026; BrightEdge measured about 48 percent of its tracked commercial queries by early 2026.

[28] Jumping Elephants, AeoGov analysis of about 2,400 full-sentence questions about the programs and services of five federal organizations, run through Google Search between March and September 2026. A further set of about 570 questions on federal election topics returned an AI Overview only 8 percent of the time. The questions resemble what people ask about government services rather than short navigational searches, which likely return AI Overviews less often.

[29] StatCounter Global Stats, “Search Engine Market Share Canada,” August 2026 (Google at 85.66 percent). https://gs.statcounter.com/search-engine-market-share/all/canada

[30] Danny Goodwin, “Google AI Overviews now appear in 39% of U.S. desktop searches,” Search Engine Land, 22 September 2026, reporting Comscore’s Q2 2026 AI Intelligence Report. https://searchengineland.com/google-ai-overviews-share-us-comscore-490381

[31] StatCounter Global Stats, “AI Chatbot Market Share Canada,” August 2026. The figures measure referrals from AI chatbots to websites, not use of the chatbots. https://gs.statcounter.com/ai-chatbot-market-share/all/canada

[32] “ChatGPT’s Canada Market Share by the Numbers: June 2026,” Tech Insider Canada, 15 July 2026, citing StatCounter’s archived data. https://tech-insider.org/ca/chatgpt-s-canada-market-share-by-the-numbers-june-2026/

[33] Angus Reid Institute, “Canadians and AI: What They Know and How They Feel,” Wave 2, March 2026. https://www.angusreid.com/wp-content/uploads/2026/04/Canadians-and-AI-Report-Wave-2-PDF.pdf

[34] Toronto Metropolitan University, survey of Canadians on AI use, fielded 14 to 27 April 2026. https://www.torontomu.ca/diversity/reports/ai-and-chat-use/

[35] Alphabet Inc., second quarter 2026 results, 22 July 2026. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm

[36] OpenAI, “Expanding access to AI with ChatGPT Ads,” August 2026. https://openai.com/index/expanding-access-to-ai-with-chatgpt-ads/

[37] Peter Pirolli and Stuart Card, “Information Foraging,” Psychological Review, 1999. https://doi.org/10.1037/0033-295X.106.4.643

[38] Sherwood News, on Google’s move to generative AI search, reporting Liz Reid’s keynote at Google I/O, May 2024. https://sherwood.news/tech/google-gen-ai-online-search-engine/

[39] Save on Energy, Home Renovation Savings program. https://saveonenergy.ca/For-Your-Home/Home-Renovation-Savings

[40] Statistics Canada, Census of Population 2021, languages spoken at home. https://www150.statcan.gc.ca/n1/pub/11-627-m/11-627-m2022051-eng.pdf

[41] Daniel Russell, Mark Stefik, Peter Pirolli and Stuart Card, “The cost structure of sensemaking,” Proceedings of INTERCHI ’93, ACM, 1993. https://doi.org/10.1145/169059.169209 (a free copy is at https://www.markstefik.com/wp-content/uploads/2014/04/1993-Cost-Structure-of-Sensemaking.pdf)

[42] Statistics Canada, “Canada’s literacy rates: Getting a read on the data,” reporting the 2022 Programme for the International Assessment of Adult Competencies. https://www.statcan.gc.ca/o1/en/plus/9316-canadas-literacy-rates-getting-read-data ; original release: The Daily, 10 December 2024. https://www150.statcan.gc.ca/n1/en/daily-quotidien/241210/dq241210a-eng.pdf

[43] Josh Brown and Maria Rosala, “GenAI for Complex Questions, Search for Critical Facts,” Nielsen Norman Group, 27 February 2026. https://www.nngroup.com/articles/ai-search-infoseeking/

[44] Qlik survey of United States tax filers, March 2026. https://secure.businesswire.com/news/home/20260311654368/en/Qlik-Survey-Mid-Career-Americans-Emerge-as-AI-Power-Users-During-the-2026-Tax-Season

[45] H&R Block Canada survey of 1,545 Canadians, fielded 19 to 23 February 2026, published 14 April 2026. https://www.hrblock.ca/blog/while-canadians-are-open-to-embracing-ai-in-their-homes-workplace-and-even-between-the-sheets-h-and-r-block-survey-points-to-cautionary-tale-that-chat-gpt-is-not-your-friend-for-tax-filing

[46] TaxGPT, a Canadian tax chatbot. https://taxgpt.ca/

[47] Caleb Sponheim, “A Concrete Definition of an AI Agent,” Nielsen Norman Group, 3 April 2026. https://www.nngroup.com/articles/definition-ai-agent/ ; see also Sarah Gibbons and Kate Moran, “AI Agents as Users,” Nielsen Norman Group, 10 April 2026. https://www.nngroup.com/articles/ai-agents-as-users/

[48] Sarah Gibbons and Kate Moran, “AI Agents as Users,” Nielsen Norman Group, 10 April 2026. https://www.nngroup.com/articles/ai-agents-as-users/

[49] Tom Loosemore, “Know Your Agent: How public services and established businesses should prepare for the arrival of AI agents,” Public Digital, 2026. Published on LinkedIn: https://www.linkedin.com/posts/know-your-agent-briefing-ugcPost-7492912458395955202-O1d6/

[50] StatCounter Global Stats, “Search Engine Market Share Worldwide,” August 2026 (Google at 91.1 percent). https://gs.statcounter.com/search-engine-market-share

[51] Tripp Mickle, Cade Metz, Dylan Freedman, Teresa Mondría Terol and Keith Collins, “How Accurate Are Google’s A.I. Overviews?” The New York Times, 7 April 2026. https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html

[52] Alphabet Inc., second quarter 2026 results, 22 July 2026. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm

[53] Tripp Mickle, Cade Metz, Dylan Freedman, Teresa Mondría Terol and Keith Collins, “How Accurate Are Google’s A.I. Overviews?” The New York Times, 7 April 2026. https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html The underlying analysis was conducted by Oumi using the SimpleQA benchmark, covering 4,326 searches in each of two rounds.

[54] Tripp Mickle, Cade Metz, Dylan Freedman, Teresa Mondría Terol and Keith Collins, “How Accurate Are Google’s A.I. Overviews?” The New York Times, 7 April 2026. https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html The underlying analysis was conducted by Oumi using the SimpleQA benchmark, covering 4,326 searches in each of two rounds.

[55] Tripp Mickle, Cade Metz, Dylan Freedman, Teresa Mondría Terol and Keith Collins, “How Accurate Are Google’s A.I. Overviews?” The New York Times, 7 April 2026. https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html The underlying analysis was conducted by Oumi using the SimpleQA benchmark, covering 4,326 searches in each of two rounds.

[56] Tripp Mickle, Cade Metz, Dylan Freedman, Teresa Mondría Terol and Keith Collins, “How Accurate Are Google’s A.I. Overviews?” The New York Times, 7 April 2026. https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html The underlying analysis was conducted by Oumi using the SimpleQA benchmark, covering 4,326 searches in each of two rounds.

[57] Adam Kalai and colleagues, “Evaluating large language models for accuracy incentivizes hallucinations,” Nature, April 2026. https://www.nature.com/articles/s41586-026-10549-w

[58] Jumping Elephants, AI and Government Information Search Survey, March 2025 and March 2026 (290 complete responses in 2026 and 328 in 2025). The sample skews toward confident, frequent internet users and is not representative of the Canadian population.

[59] National Safety Council, Injury Facts, “Odds of Dying,” lifetime odds for a person born in 2024, United States data. https://injuryfacts.nsc.org/all-injuries/preventable-death-overview/odds-of-dying/

[60] Alexandra Mae Jones, “Musician Ashley MacIsaac ends lawsuit over Google AI summary falsely labelling him a sex offender,” CBC News, 18 August 2026. https://www.cbc.ca/news/entertainment/macisaac-google-lawsuit-ends-9.7310819

[61] Andrew Baker, “AI Hallucination Is Not Lying. That’s Why It’s Dangerous,” 14 September 2026. https://andrewbaker.ninja/2026/09/14/ai-hallucination-is-not-lying-thats-why-its-dangerous/

[62] Adobe Digital Insights, “Quarterly AI Traffic Report,” April 2026. https://business.adobe.com/resources/sdk/.2026-q2-ai-traffic-report/q2-2026-adi-ai-sourced-traffic-insights.pdf

[63] Anja Karadeglija, “News publishers’ copyright lawsuit against OpenAI to go ahead in Ontario,” The Canadian Press, 27 November 2025. https://www.cp24.com/news/canada/2025/11/27/news-publishers-copyright-lawsuit-against-openai-to-go-ahead-in-ontario

[64] Courthouse News Service, “Anthropic to pay $1.5 billion copyright settlement to authors, publishers,” 20 July 2026. https://courthousenews.com/anthropic-to-pay-1-5-billion-copyright-settlement-to-authors-publishers/

[65] Columbia Journalism Review, “Reddit Is Winning the AI Game.” https://www.cjr.org/analysis/reddit-winning-ai-licensing-deals-openai-google-gemini-answers-rsl.php

[66] CB Insights, “Hookin’ up with OpenAI,” 4 June 2024. https://research.cbinsights.com/hookin-up-with-openai ; “The Billion-Dollar Bailout: A Running Tracker of Every Publisher AI Licensing Deal,” PR News, July 2026. https://everything-pr.com/ai-licensing-tracker

[67] CNBC, “Reddit stock sinks on report it may not renew Google AI content deal,” 22 July 2026. https://www.cnbc.com/2026/07/22/reddit-stock-google-ai-content-deal.html

[68] Library of Congress, Global Legal Monitor, “Germany: Court Holds Google Liable for Incorrect AI Overviews,” 17 July 2026. https://www.loc.gov/item/global-legal-monitor/2026-07-17/germany-court-holds-google-liable-for-incorrect-ai-overviews/ ; The Decoder, “Landmark German ruling declares Google’s AI Overviews are Google’s own words and makes it liable for false answers,” 11 June 2026. https://the-decoder.com/landmark-german-ruling-declares-googles-ai-overviews-are-googles-own-words-and-makes-it-liable-for-false-answers/

[69] Court of Justice of the European Union, Google Spain SL and Google Inc. v Agencia Española de Protección de Datos and Mario Costeja González, Case C-131/12, judgment of 13 May 2014. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:62012CJ0131

[70] Catharine Tunney, “Ottawa moves to restrict social media for kids under 16,” CBC News, 10 June 2026. https://www.cbc.ca/news/politics/online-harms-ai-social-media-children-9.7229976 ; Government of Canada, “Bill C-34, the Safe Social Media Act.” https://www.canada.ca/en/canadian-heritage/services/safe-social-media-act.html ; Al Jazeera, “Canada introduces bill to ban social media for children under 16,” 10 June 2026. https://www.aljazeera.com/economy/2026/6/10/canada-introduces-bill-to-ban-social-media-for-children-under-16

JP Simbandumwe