AI Agents Need Transparency Measures to Reliably Interact With Canadian Commerce

Recently, the Canadian Shield Institute submitted a policy brief to the federal government’s consultation on AI transparency. 

It’s good that Ottawa is studying this, and we hope they’ll take swift action. Unless you’ve been living under a rock, you’ve probably heard something about AI agents going rogue in the past few months.

You can read our full policy brief here, but I’ll talk about our key policy recommendation below. 

But for the sake of context, let’s first look at one AI hack that’s notable precisely because it’s so mundane.

In April, an Australian man asked his AI agent to book him into a full pilates class. The agent discovered that the studio’s system didn’t check who was cancelling reservations, so it cancelled the booking of the person at the top of the waitlist and moved its user up. 

The agent did what was asked, and a stranger lost their spot. Most people would feel a twinge of guilt about exploiting a software bug to cut the line, but the AI agent just did it.

This incident is important for a couple reasons:

1. We only know about the breach because the user, a software developer, reported the flaw to the studio and the story later made the news. A less scrupulous or less technical user would never have noticed.

2. There are a staggering number of systems that might contain similar vulnerabilities — medical appointment queues, daycare sign-up portals, restaurant reservation platforms. We use tools like this every day, and it’ll be a total mess if all of it gets turned over to agents who will exploit software with mechanical efficiency.

The immediate risk from agents is not a rogue superintelligence, but millions of small, untraceable harms to people who never knew an agent was involved. When something goes wrong, no one can say who acted, on whose authority, or who was hurt. If OpenAI needed a week to find its own agents in its own logs, a pilates studio has no chance. So, what’s the solution?

In our policy brief to the federal government, we argue that agents need ID and proof of authority to act on a user’s behalf. 

Any time an agent is clicking around the internet, the bot should identify itself as a bot, and it should have a traceable, pseudonymous ID number. The ID should only be traceable back to the AI company and the end user as part of a legal process, but if bots are out there hacking or making a mess on the internet, we should be able to identify who’s ultimately responsible for them.

Designing agent identity and authorization records now costs far less than retrofitting them later.

If Canada doesn’t write these rules, someone else will. Visa, Mastercard and Google are already drafting standards that serve their own networks. Canada’s rule of law and neutrality make it the right place to build the independent version. 

One good place to start would be implementing these rules with payments.

Payment regulations are federal jurisdiction, and we already have sophisticated regulators in place. The government is also involved in a big update to the underlying rails that govern payment infrastructure.

By focusing on payments regulations, we avoid biting off too much in one go.

With other technologies like social media, targeted advertising and algorithmic video, we’ve waited until the technology was more mature before thinking about how to regulate. 

It turns out, when we take that approach, Canada is usually forced to accept somebody else’s rules, because we’re late to the game. 

Now’s the time to write our own rules.