Procedure and House Affairs Committee on Oct. 20th, 2026

[Christopher Henry]

Those are very good questions.

The basic answer is that we don't have all of the capabilities and sophistication to AI-detect all content being submitted to the House. Therefore, the same aspect that we would impose on stakeholders checking a box, we also on our side as we're receiving briefs and information that come from the public.... While we could equip our people with tools to look at and, perhaps, assist with detection, ultimately, our human judgment is applied to make decisions based on what is coming in. Then we apply countermeasures when necessary, such as contacting the individuals or the person in their basement and stating that we may not accept those materials as such.

As to going so far as to track their location and others, we'll stay away from that one. However, we're certainly doing our utmost on the “human side” protection by our administrative staff, to make sure that we do detection there.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

I'd like to thank the witnesses, and I would like to thank the army of potential witnesses behind them.

We will suspend for a few minutes for the next panel.

[The Chair, Liberal, Chris Bittle]

Welcome back, everyone.

I would like to welcome our second panel of witnesses. From the International Association of Conference Interpreters, Canada region, we have Alionka Skup, president, and Marina Mistral, conference interpreter and member of the AIIC science hub. From the Westminster Foundation for Democracy, we have Franklin De Vrieze, head of practice and accountability.

We'll go first to our witnesses who are present here. Then we'll go to our witness online.

You have five minutes, please.

[Alionka Skup, President, International Association of Conference Interpreters - Canada Region]

Good afternoon.

On behalf of the International Association of Conference Interpreters - Canada Region, or AIIC-Canada, I would like to thank you for this invitation. I would also like to acknowledge that we are meeting today on the sovereign and unceded territories of the Algonquin Anishinabe people.

AIIC-Canada covers the Canadian region of AIIC and represents 3,000 members across the country. Many interpreters work for the highest authorities in the world.

Today, we're here to talk to you about artificial intelligence and interpretation. Marina Mistral is our expert here to tell you about this topic.

[Marina Mistral, Conference Interpreter and Member of the AIIC Science Hub, International Association of Conference Interpreters - Canada Region]

Good afternoon, Mr. Chair and committee members. Thank you for the invitation to participate in this study.

My name is Marina Mistral, and I'm an accredited parliamentary interpreter. I would like to say hello to my excellent colleagues in the booth, with whom I have shared the booth in the past. My background as an engineer and my interest in technology have led me to sit on the new scientific committee of the International Association of Conference Interpreters, which is looking at the issue of artificial intelligence in interpretation, among other things.

Artificial intelligence is undoubtedly a very powerful tool for interpreters, but we believe that entrusting it with the mike carries risks of misinformation, the erosion of bilingualism in Canada, the breakdown of democratic discourse and the increased vulnerability of our institutions.

First, the use of machine interpretation carries technical risks that other experts have already spoken to you about in this study. Outages, or even system shutdowns for maintenance, could jeopardize the continuity of Canadians' language rights. A breach could make everything that happens here behind closed doors public. Administrative friction could prevent terminology updates from being frequent enough to keep up with current events.

The greatest risk with AI interpretation lies in the fact that artificial intelligence is not, in fact, intelligent. It is a statistical pattern-matching black box that may sound human but doesn't actually think. Therefore, it cannot be trusted to make the kinds of decisions that interpreters make every day to communicate effectively. I'll give you some examples.

From a sense-making standpoint, we can distinguish information from noise and unimportant words. We can make sense of unfinished sentences and subtext. When we hear a homophone, we know which meaning is intended. We can follow someone's reasoning and use intonation to facilitate comprehension. We can understand and convey tone, irony, humour and cultural references.

On the practical side, we can adjust our pace or reformulate when interpreting from English to French, which requires more words to say the same thing. We understand regional accents, dialects and proper nouns. Unlike AI, we can notice when we make mistakes, and we correct ourselves. We will flag technical problems when they arise and let you know when your microphone is off. All these choices make communication reliable and effective and require real reasoning, which AI is incapable of by nature and unlikely to be able to do in the foreseeable future.

We believe AI should never be given the job of interpreter, because nobody deserves AI-grade interpreting. Interpretation that cannot be trusted comes with serious repercussions and reputational risks. With the risk of public-facing information being distorted in unpredictable ways, the dialogue that supports our democracy could be weakened.

When MPs cannot trust what they hear, they cannot presume to be heard. How can they, then, represent their constituents? If francophone constituents, as the linguistic minority, are forced to choose between leaving the conversation or having it in English, it means they cannot participate meaningfully, and certainly not equally. It's a violation of their rights.

To build Canadians' trust in the use of artificial intelligence and set an example for the rest of the world, Canada must promote the responsible use of these tools, meaning only in contexts where their output is verifiable. With that in mind, if we compare translation and interpretation, it's entirely possible to revise a machine translation to make it a high-quality final document. That is a possible use.

With interpretation, it would be physically impossible to listen to two audio feeds at the same time and make real-time corrections, in addition to all the human resources required.

Even though artificial intelligence is not suited to the dynamic process of interpretation, it can serve as a valuable ally to interpreters. At the preparation stage, these tools now enable us to translate and summarize texts, quickly consult a variety of sources and generate glossaries on subjects that we're sometimes called upon to master in under an hour. Artificial intelligence tools could also help quickly create reliable schedules and provide each committee with the best-suited team, while respecting the many constraints that are currently causing the coordination team a lot of headaches.

We believe that Canada should adopt these best practices for using artificial intelligence and provide employees and suppliers with training and support regarding the responsible use of these kinds of tools, while continuing to listen to them as the technology evolves.

We would now be delighted to answer any questions you may have.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

Before we get to questions, we'll turn to Mr. De Vrieze for five minutes, please.

[Franklin De Vrieze, Head of Practice Accountability, Westminster Foundation for Democracy]

Good afternoon. I'm speaking from London.

I'm Franklin De Vrieze.

Thank you very much for the opportunity to address you today. It's a true pleasure joining you.

I speak on behalf of the Westminster Foundation for Democracy. In terms of background, the WFD is the U.K.'s public body for international democracy support. We are linked to the FCDO, the Foreign, Commonwealth and Development Office, as an arm's-length body with the mandate to provide international democracy support. We have around 30 country offices around the globe, where we work with national parliaments and subnational parliaments in those countries on a wide variety of themes and topics relevant to democratization.

With regard to this topic we are discussing today, there are three themes that we are focusing on. These are the digitalization of parliaments, the introduction of AI in parliaments and the question of cybersecurity. The three themes are, of course, linked to each other, but they are still very distinct.

I have just a couple of words at the start of my contribution. From our observation, working with so many parliaments, what do we see regarding why parliaments consider using AI-enabled tools? It has to do, of course, with the expectation of enhanced efficiency, increased transparency, better public engagement, accessibility, etc., but there are also expectations that go much further. For instance, there's predictive analysis or expectations around enhanced legislative drafting. Of course, these are the issues that are much more sensitive.

Another thing we have observed is what we call the pacing problem, which is the growing gap between the speed of technological change and the speed of change at political and decision-making levels. We see this pacing problem increasing, and it's harder and harder for political institutions to keep up with the speed of change at the technological level.

Specifically for parliaments, we see this pacing problem at three levels. We see an external pacing problem between parliaments and industry and society, but often—and I think this is important here—there's a pacing problem between the different branches of government. In quite a number of countries, we see parliaments sometimes struggling to keep up with how governments or executive agencies are introducing AI in their functioning. The effect of that is executive agencies, public services and government departments in many countries sitting on a mountain of data, which, of course, parliaments do not have at their disposal, although parliaments are supposed to exercise oversight and some control functions.

I think the issue of executive dominance is an increasing problem in many countries around the world with regard to or on the occasion of the introduction of AI.

We see AI, as you've just discussed, introduced in various ways in parliaments. I think the most obvious one is at the administrative level in terms of transcription, translation, document classification and summaries. The more advanced and the more delicate one, as I mentioned, is the introduction of AI tools with regard to the analysis of amendments, checks of legal consistency and information retrieval.

There are a couple of interesting examples of this. For instance, the Italian Parliament has piloted AI-supported compliance checking of amendments. The French National Assembly, for instance, has created tools to model fiscal and social impacts of legislation and also to compare draft bills with existing legislation.

There are plenty of opportunities with regard to the introduction of AI in parliaments, but there are also a number of risks, and they have already been discussed in your previous conversations. As a short introduction, this is what I wanted to share with you today.

Thank you once again. I'm looking forward to the conversation today.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

We'll now turn to questions. We'll start with Mr. Calkins for six minutes, please.

[Blaine Calkins, Conservative, Ponoka—Didsbury, AB]

Thank you, Chair.

I'm going to start my questions with you, Mr. De Vrieze. Thank you for joining us. I don't know what hour it is where you are, but I'm sure you'd rather be somewhere else. Thank you for being with us here at the committee today.

I have a question for you because this is something that's on my mind a lot when it comes to the use of artificial intelligence, and you brought it up. I don't think we've had a good conversation about this yet.

Here in Canada, it takes anywhere from three to nine months for a regular piece of legislation to go through its process. This is from the time it's introduced at first reading. It then goes for healthy, robust debate at second reading, and if it passes, it goes to committee. The committee will then sit and take anywhere up to two to three months to go through deliberations on the bill, make its amendments, send the bill back to the House of Commons—in the Westminster system, of course—and have more debate at third reading just in time to send it all to the Senate, where it goes through the exact same process again. If there are no changes being made, it gets royal assent from the Governor General on behalf of His Majesty the King.

That all takes months. How fast does AI evolve in three to nine months? That's how long it takes us to pass one bill.

[Franklin De Vrieze]

That is a very accurate question, but it's also a question to which there are very few easy answers to give, because the question doesn't apply only to the functioning of Parliament but to the functioning of public services in general.

You described a process of six to nine months. AI can function in the legislative process best when it's very targeted. That means when it's looking at, for instance, text consistency, cross-referencing and checking legislation or amendments with existing legislation. I think it can also contribute by providing a basis on the ex ante impact assessment of legislation based on the available sources.

Specifically on the issue of impact assessment, the experience of the French is that human verification and cross-verification are very important. That's why I say AI doesn't provide an impact analysis. It provides a draft or a basis to conduct an impact assessment.

The easier elements to use AI in the legislative process are around terminology alignment, maybe translation support and also, I would say, identifying normative gaps. Of course, no normative gaps are intended, so we're looking for unintended normative gaps and sometimes unintended consequences or, specifically, consequences for specific groups in a society that might be overlooked or might not be accurately assessed. These are issues around predictive analysis where AI tools are being piloted and these things are explored.

I don't think we are yet at the stage where we have a reliable system and a reliable set of guidelines with regard to this. I think we are still very much at the level of piloting. Piloting is good and experimentation is good, but experimentation and piloting are not the same as transformation. We are not yet at the stage whereby the legislative process as such will be transformed. We are still looking at ways it can be improved with the help and assistance of these various AI tools.

I'm sure you have heard this comparison. I'll conclude my answer on this question with this comparison. AI tools are a bit like an intern. We all love interns. Interns do very good work if they are well trained, if they have had the time to familiarize themselves with various issues and if there is good oversight and supervision of the role of interns. It's the same with regard to the role of AI tools. In that sense, I would exercise caution with regard to the use of AI in the legislative process, but on various other issues, with regard to communication and administrative issues and citizen outreach, I think it can provide useful contributions.

[Blaine Calkins, Conservative, Ponoka—Didsbury, AB]

I would agree with that. I do believe, though, notwithstanding that every parliament, every country, tries to pass technology-neutral laws when it comes to governance, that the rapid growth and change in artificial intelligence is something that our Parliament, and many parliaments, will have to wrestle with. As far as being nimble enough, our parliaments and our democracies have to be nimble enough to respond to sudden and unexpected changes in terms of where artificial intelligence might be taking our societies.

Notwithstanding that, I am also concerned about information sharing with the public. We heard from witnesses at a previous meeting of this committee that in a recent sub-sovereign provincial election in Canada, I think about 48% of the information that was gathered and researched for that particular election was attributed to a website that was generated by artificial intelligence from a political hobbyist.

What should our Parliament be doing to make sure that the information that's verified, factual and true is actually the information that artificial intelligence is using when it comes to communicating what's actually happening with our democracy at the federal level in Canada?

[The Chair, Liberal, Chris Bittle]

That was very long, but we will need a very brief answer. I'm sorry.

[Franklin De Vrieze]

Thank you very much. It's an important question.

I would say that the role of Parliament as an institution—not individual MPs, but the role of Parliament as an institution—with regard to outreach to citizens can be helped with some specific AI tools, such as an AI-powered chatbot at the Parliament website. The key issue, of course, is that the information included in this chatbot to respond to citizens' inquiries needs to be based on Parliament information only, not on something that the chatbot will search on the internet, because then there is a risk for hallucinations. Chatbots in reaching out to citizens can be useful if the framework for the source of information is clearly defined.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

We'll turn to Madame Brière for six minutes, please.

[Élisabeth Brière, Liberal, Sherbrooke, QC]

Thank you, Mr. Chair.

Thank you to all our witnesses.

Thank you very much for your remarks.

Ms. Mistral, you mentioned the fact that artificial intelligence is a very powerful tool for interpreters. I imagine you use it for certain tasks. Can you give us some concrete examples of uses that don't compromise the quality of your service? For example, I'm thinking about looking up words in a specific sector or taking notes.

[Marina Mistral]

Here in Parliament, we use artificial intelligence a lot, because we're in a context where we often have to prepare very quickly for committees and meetings. Sometimes we only have a few minutes.

It's now possible to use confidential—hence, paid—versions of artificial intelligence. Those enable us to ask terminology questions. For example, I was working on a committee—I believe it was the Standing Committee on Fisheries and Oceans—for which I had to very quickly put together a termbase on a type of fish and on crab fishery. That means it was very specific to a given situation.

It's possible to ask questions and put together glossaries, but it's very important to then verify those glossaries using official Government of Canada resources and to consult studies previously done by the government to ensure that we're actually using Canadian terminology. That's really important. Acronyms can often be a problem, so we need to verify them. We need to verify regional ways of saying things, because artificial intelligence still tends to base its French on European French content. That's one example, then.

It's also possible to use transcription in some cases, but, again, it's important to really verify what we're reading.

My colleague can add examples of uses, but I would ultimately say that we can use artificial intelligence for terminology extraction. This can be very useful if we have two texts that have already been translated by the translation services here. It's possible to extract very frequent terms or technical terms.

That's how we normally use AI. We also sometimes use it to summarize a topic. If we provide material or documents, in particular, we can ask for a summary that, again, we can verify with our human intelligence.

[Élisabeth Brière, Liberal, Sherbrooke, QC]

I'll lighten the mood a bit. In my previous job, I once received a translation generated by artificial intelligence. The translation of “legal seal” came out as “phoque légal”.

I'm struck by another point. You spoke about the decline of bilingualism. I would like you to elaborate a bit on this issue. Do you think that the quality of interpretation might vary between French and English? We also know that indigenous languages are spoken in the forums. I would like to hear your comments on these two aspects.

[Marina Mistral]

Good question. I would like to provide an example, because I didn't have time to do so in my opening remarks. I can send you the sources.

The World Health Organization, a United Nations agency, has already looked into this issue. It assessed the quality of artificial intelligence in interpretation with the help of a team of professors from Geneva. Only 1 out of 90 speeches received a passing grade. All the speeches posed some kind of reputational risk, because the artificial intelligence often failed to identify proper nouns, country names or people's names. It also made gender‑related errors and did some truly objectionable things.

French, as the source language, received the lowest score at 36% . This is noteworthy because the interpretations into English carried the highest number of reputational risks. This was also compounded by a delay of over 30 seconds in the interpretation. The figures and technical information posed real problems.

I think that, at this point, the quality of this technology is nowhere near sufficient to support Canadian democracy, especially given that francophones are a minority here. The same goes for indigenous people, of course. We also know that Canada has a thriving francophonie that stretches from the Atlantic to Manitoba, including Quebec, of course. Each community has its own regional accents and terminology.

You can already see this in Moncton. You may have heard that the city's conference interpreter has retired. The city has had to turn to the largest provider of artificial intelligence‑based interpretation services. You can listen to the interpretation yourself. It's a total disaster. The city will hire an interpreter as soon as possible. In addition to a delay of up to a minute, the machine is basically unable to understand anything spoken in French or English. It often makes hilarious mistakes, which certainly pose a reputational risk.

[The Chair, Liberal, Chris Bittle]

Thank you so much.

Ms. Normandin, the floor is yours for six minutes.

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

Thank you.

Ms. Skup and Ms. Mistral, I would like to continue discussing two of the points that just came up.

First, as you said, interpreters are in short supply, especially in Moncton. Would it send the wrong message if parliaments started using artificial intelligence to recruit and train interpreters, who remain crucial? Are wrong messages already being sent? In your opinion, would this send another wrong message that could give us many headaches in the future?

[Alionka Skup]

I think that young people who have an interest in languages are already concerned. They don't want to pursue a career in this field given all the talk about artificial intelligence replacing interpreters and translators. This risk is already a reality.

We're already seeing some institutions—even international institutions such as the Organization of American States—conducting trials with artificial intelligence. The results have proven disastrous, particularly for French, once again. It seems that artificial intelligence is being trained less frequently with French.

This certainly sends a frightening message to people who are interested in getting into the field. Rather than being a solution—as we're being led to believe—it further exacerbates the current shortage.

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

Let's talk about delays in circumstances where, like here, speaking time is crucial. Sometimes we have only two and a half minutes to ask a question. If the interpretation takes an extra minute, we might miss the beginning of a sentence. When the conversation switches to another language, we might miss the rest of it. We could end up missing half the remarks as a result of poor interpretation timing.

I just want to hear you say it again. What's the added value of an interpreter who can adjust the length of sentences given that the timing must be right in order for us to catch everything?

[Marina Mistral]

The delay stems from the nature of machine interpretation. For now, and likely in the short to medium term, the process begins with a transcription. This step alone is prone to errors, especially for languages other than English. The machine then performs the translation and reads the new text. This whole process causes a delay, depends on the technology and remains vulnerable to attacks.

Obviously, the training program for interpreters here in Parliament includes the most demanding exam in the country to ensure that we can keep up with the conversation. Our delay is no more than five seconds, and that's really only in the worst‑case scenario. We're often able to summarize the remarks and rephrase them. We've received training. Our experience helps us convey the meaning of what people are saying, rather than simply translating words. This is really important. You're thinking individuals who convey ideas. We're also thinking individuals who convey your ideas, because we understand you.

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

You brought up various aspects. Interpreters can pick up on expressions, intonation and cultural differences. I would like to hear your thoughts on this. Is the non‑verbal aspect being addressed? When interpreters look at us, can they also pick up on some of the information through non‑verbal cues, which would be absolutely impossible for artificial intelligence to detect?

[Alionka Skup]

It's certainly one of the things that we use. Like everyone else, when we speak with someone, we respond to non‑verbal cues. We also understand certain aspects as interpreters and as human beings who can do this. Clearly, artificial intelligence can't access this information at all. It truly takes a human's understanding of other humans to pick up on non‑verbal cues. That's absolutely certain.