Procedure and House Affairs Committee on Oct. 22nd, 2026

[Michael Cooper, Conservative, St. Albert—Sturgeon River, AB]

I take it from what I'm hearing that, overall, it has been a success. Is that fair?

[Luke Hussey]

We've been looking at it for a couple of years. To be frank, when it started, I wasn't very impressed by the tool, and the experience wasn't great. It has really improved in recent months. Because it uses access to different models, it's a much better experience than it was. It's much more equatable to your OpenAI ChatGPT equivalents than it used to be. It meets our needs for the sort of work we would expect people to want to do.

We need to consider niche examples and opportunities wherein we might want to provide other tools securely.

[Sarah Davies]

If I can come in on that one, I think PDS and our digital colleagues have been very proportionate.... What I would describe as the full-fat Copilot is expensive, and I think it's very unlikely that we will see a rollout to all users. Really, what we're looking for and what Luke has been working on, is evidence that it's demonstrating value. We're aware of how much it costs now. It might cost even more in the future, so we want to make those kinds of sensible decisions. Yes, there are uses, but it's not going to be the full-fat version for everyone.

[Michael Cooper, Conservative, St. Albert—Sturgeon River, AB]

How expensive is it?

[Sarah Davies]

Luke, how much is it at the moment for us?

[Luke Hussey]

It's about 220 pounds per user for the year. If we gave that to all 11,000 people in the parliamentary community here in Westminster, that would be 2.3 million pounds extra per year, so we just want to be thoughtful about that.

[Michael Cooper, Conservative, St. Albert—Sturgeon River, AB]

It is expensive.

I have 30 seconds. On this same general topic, Ms. Davies, you referenced the use and the need for members to be able to access publicly available AI tools. You referenced the model context protocol. Could you elaborate on that?

[Sarah Davies]

Yes. Model context protocol is open-source. It is originally from Anthropic. It is basically a tool that seems to enable us to effectively interrogate parliamentary information and bring up results based only on that parliamentary information. That's why we think it's interesting and promising.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

We now have Madame Vandenbeld for five minutes, please.

[Anita Vandenbeld, Liberal, Ottawa West—Nepean, ON]

Thank you very much, Mr. Chair.

It's always very interesting to have other parliaments come and talk to us.

I think it was Ms. Davies who mentioned that you have guidelines. You have a governance structure. You have the Procedure Committee. Has there been a study, similar to what we're doing in public, by your Procedure Committee?

I would ask the same question of Mr. Stoddart.

[Sarah Davies]

Our Procedure Committee at this end, in the House of Commons, has touched on some of these issues. It has looked recently in particular at the impact of AI technology on written parliamentary questions, among other things. That's a really interesting report, but it hasn't done a specific inquiry into AI.

We have a departmental scrutiny committee, which covers the government department. That's very active in this area. The Procedure Committee has looked very specifically at written parliamentary questions and not, so far, at the general issue.

[Anita Vandenbeld, Liberal, Ottawa West—Nepean, ON]

Would the Procedure Committee be involved in approving any guidelines or have input into those guidelines? I'm wondering who decides on the guidelines. Is it members or is it the administration?

[Sarah Davies]

In terms of guidance—and this is an area in which we could progress quite early—our Speaker set up a steering group on AI in parliaments about two years ago. I'm a member of that. The Deputy Speaker chairs it. It has a small number of members involved who have, really helpfully, very different views and experience with AI adoption. It is that body, working with PDS, that produced the first version of this guidance, in a Q and A format. That's now on about its fourth or fifth edition, and then we have wider governance around that.

[Anita Vandenbeld, Liberal, Ottawa West—Nepean, ON]

Thank you. That's very interesting.

Mr. Stoddart, go ahead.

[Alistair Stoddart]

The standards committee, which is the equivalent to your committee, conducted an inquiry into committee effectiveness at the end of the last session, although I had a quick look and artificial intelligence wasn't really a big part of that remit. Nothing's been looked at with that, but there is another standards committee for this session.

[Anita Vandenbeld, Liberal, Ottawa West—Nepean, ON]

Thank you. That's very interesting and helpful.

We deal with a lot of things that are under parliamentary privilege, such as in camera testimony, our own MP emails and draft reports.

For instance, what if somebody were to put something that was said in camera into an AI? Have you looked at how privilege intersects with the use of AI and, potentially, information that shouldn't be in the public domain getting into the public domain?

[Sarah Davies]

Yes, and this has been a really interesting aspect of looking at Copilot and the possible consequences of more widespread data management there. There are two angles to this.

First, one of the key issues we initially looked at in the guidance for members and for members' staff was awareness. They have legal responsibilities in our system for the data they hold. One of the key things we did very early on was to make it clear to them that this kind of personal, confidential and sensitive data should not be shared under any circumstances with these generative AI tools. That was one of the front-and-centre aspects of the initial versions of the guidance on that, and that was probably about 12 to 18 months ago.

Second, more recently, what we've seen as part of Luke's pilot is that Copilot sometimes surfaces information if that information has been incorrectly labelled and logged within a system where you might not have expected it to be. You have to look at the underlying permissions, the underlying structure and your information management, because otherwise it can find things you would not expect it to find and possibly even share them with others.

In parallel with all the work on evaluation with Copilot, there has been all this work: We've had to go through all our permissions and do a bit of proper old housekeeping of how we're storing, sharing and labelling information to minimize that privilege risk. Luke will know that there are colleagues here who've been extremely concerned, particularly from a privilege perspective, about the information that might be surfaced if we don't do that evaluation correctly.

[Anita Vandenbeld, Liberal, Ottawa West—Nepean, ON]

Thank you very much.

I'm out of time, unfortunately. We could ask you questions all day. Thank you.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

Ms. Normandin, you have the floor for three minutes.

[Christine Normandin, Bloc, Saint-Jean, QC]

Thank you very much.

I have a question for you, Ms. Davies.

Many witnesses, including you, have spoken about the importance of having AI with a human-centred approach. You use Copilot. I would like to know the extent to which the tool truly enables you to adopt this human-centred approach, and whether Copilot's architecture is open enough to allow humans to conduct checks at various stages.

I'd also like you to tell us about the trade-off between the benefits of using AI and the human labour costs involved in ensuring that it's working properly.

[Sarah Davies]

Thank you very much for the question.

My starting point partly has reference to the previous question as well. The nature of the confidential data that we hold means that part of the challenge here is that because our email data and everything in our out boxes is, as it were, as safe as we think it is in Microsoft, if we are going to use a system and experiment with a product, it needs to be Copilot. I can't enter the same information into another external tool, because it would be less safe.

We have that starting point. As Luke has already mentioned, I need people within my teams here to be experimenting safely with these forms of technology to see how they can be used in this parliamentary environment. That's why we've been having this rollout that we can show.

As you mentioned, there are some inherent weaknesses in the way these models work about really knowing what they surface, what they don't surface and what they miss. The thing I've learned over time is if we can have a core group of staff at all levels and in different parts of the organization testing with things such as Copilot, which is our secure opportunity in terms of parliamentary data, then we start to develop an informed cost-benefit analysis in terms of time and money.

We start with Copilot, because that's where our information is, and we can't be entering that information anywhere else. That's the challenge. That's why we're with Copilot in that sense.

[Christine Normandin, Bloc, Saint-Jean, QC]

Thank you.

[The Chair, Liberal, Chris Bittle]

Mr. Moore, you have five minutes, please.

[Rob Moore, Conservative, Fundy Royal, NB]

Thank you, Mr. Chair.

Mr. Stoddart, we've seen examples in recent campaigns in which an outsized amount of information comes from very few people or even one individual. As members of Parliament, we sometimes weigh considerations in our riding by the interest that we're receiving and the correspondence we're receiving. Hearing from you on some of these AI-generated campaigns and then having AI-generated answers, we're led to situations in which we could have AI just talking back and forth with AI without the proper oversight.

I think this goes to something else you mentioned: the democratic integrity of what we do as parliamentarians.

Can you expand a bit on those types of scenarios? What are some of the safeguards you would say we should put in place to ensure that we're able to do our jobs to get to the root of the issues facing our constituents?

[Alistair Stoddart]

Thank you, Mr. Moore.

I thought about the AI loop as a thought experiment. You could have a consultation process in which AI does the full process all the way through, and no human is really interacting, and it is a risk. That's when human judgment and critical thinking come in.

In terms of guardrails, there are aspects to do with thinking about the.... One thing is that, if it's a classic postcard campaign, then really AI is just a supercharged postcard campaign. Then you maybe handle that in the same way if it's the same information packaged in that way.

As AI gets more sophisticated, they could try to hide that in some ways. There are things like looking at digital fingerprints, which doesn't take into account just IP addresses and whether it's coming from a particular server; it will look at all other aspects in what that is to identify whether that's a human. There are also things about how it's written. The way AI writes may be very different from how a human writes, so there are different things you can think about to make that judgment, as I said, about whether this is an authentic, genuinely held view or something that's, at the very least, magnified.