What BC told Ottawa about AI transparency

Executive summary

Two things we want Ottawa to hear.

In July 2026, Innovation, Science and Economic Development Canada asked Canadians where AI transparency matters: knowing when content was made by AI, knowing when you are talking to a system rather than a person, getting clear information about what a system can and cannot do, tracking serious incidents, and tracking what agents do on someone's behalf. The government says plainly that transparency will not fix every problem. Two things came out of our room that matter to this consultation specifically.

01

What was said lines up with what the government already knows.

The government's own discussion paper includes a case study about an AI shopping agent that overspent a person's money, with no clear answer as to who was responsible: the person, the business, or the developer. Independently, a participant in our session described a closely similar situation from personal experience: an unauthorized, system-driven transfer between bank accounts, with no consent obtained, no clear record of what happened, and no straightforward way to get it corrected.

The paper also describes a car dealership's chatbot making an offer the dealership later said it was not authorized to make. Participants raised the same pattern on their own: when an AI system causes a problem, it is often unclear who is answerable, and there is often no record detailed enough to establish what happened.

02

The bigger concern isn't one of the five questions asked.

The most consistent thing said in the room was not about any specific transparency mechanism. It was doubt about whether this process changes anything. Versions of the same statement came up repeatedly: that this looks like an exercise to make people feel heard without changing what happens next.

Whether or not that is an accurate read of the government's intent, it is what a number of participants believe, and it will affect whether any transparency measure that comes out of this process is trusted enough to be used.

The five federal questions

Five questions, one room.

ISED's survey runs to 24 questions across five areas. The facilitator read the government's framing for each, then opened the floor. Pick an area to see what Ottawa asked, what the room answered, and what already exists that the room pointed to.

Cover of the Government of Canada discussion paper Enhancing trust in artificial intelligence through increased transparency, July 2026.

The document Ottawa put outEnhancing trust in artificial intelligence through increased transparencyISED discussion paper, July 2026. Read it on canada.ca

01 / What Ottawa asked

Would it help you trust what you see online if you could tell whether something was created by AI or by a person? When and for what kinds of content is it most important to know if something was created or modified by AI?

What the room said

Yes, but the useful question is how AI was used, not whether it was used. Hybrid media is already normal.
A label is a cigarette warning. Prevention and public-health-style education have to travel with it.
If most content ends up AI-touched, it may be more useful to mark what is human-made.
Real-time consent is the missing piece in every provenance standard. Quebec’s Law 25 was named as the model.

What already works

C2PA, an open technical standard for labelling AI-generated or AI-edited content, backed by Adobe, Google, Microsoft, and others.

01 / AI-generated content
Would it help you trust what you see online if you could tell whether something was created by AI or by a person? When and for what kinds of content is it most important to know if something was created or modified by AI?
- Yes, but the useful question is how AI was used, not whether it was used. Hybrid media is already normal.
- A label is a cigarette warning. Prevention and public-health-style education have to travel with it.
- If most content ends up AI-touched, it may be more useful to mark what is human-made.
- Real-time consent is the missing piece in every provenance standard. Quebec’s Law 25 was named as the model.
Already working: C2PA, an open technical standard for labelling AI-generated or AI-edited content, backed by Adobe, Google, Microsoft, and others.
02 / Knowing when you are interacting with AI
When and why is it most important to know that you are interacting with an AI system? Are there contexts where you do not need to know that you are interacting with AI?
- Every time. Until trust exists, disclosure is needed one hundred percent of the time.
- People want to know when their bank is using it.
- China’s rule that an AI cannot play the role of a human, such as a doctor, was noted as one existing hard line.
- Disclosure needs a working human override behind it, not a rubber stamp.
Already working: Quebec’s Law 25 real-time consent requirements, and the professional-liability model in medicine, where a practitioner is the filter on what an AI system suggests.
03 / Information about AI systems
What kinds of information about AI would you like to know in order to help you make informed choices about when and how to use AI? When and for which types of AI products is this information most important to you?
- Which organization the system comes from, and where its data came from.
- What the algorithm is actually optimized for.
- Whether a bureaucratic or financial decision about you was made with an AI system in the loop.
- Documentation a non-specialist can actually read.
Already working: Plain-language documentation requirements, and a broader legal definition of harm that includes harm to groups, not only individuals.
04 / AI incidents
How should a serious AI incident be defined, and what information would be most useful to better understand the causes and consequences, as well as the mitigation, of these incidents?
- A serious incident is one that harms a person, and the threshold for harm differs across cultures.
- Or simpler: it is serious if it breaks a law.
- Deepfakes of a real person should be a line not crossed. Each person should own their face and voice.
- Individual-focused law struggles to measure harm to a group. Credit scoring was the example.
Already working: Human review and override for consequential automated decisions, audit trails, and incident reporting modelled on existing safety practice in other sectors.
05 / AI agents that act on your behalf
Are you concerned about the use of AI agents? In what contexts is it most important to know when an AI agent is in use? Why?
- Is there a human who can override a decision an agent made?
- One participant’s family had money moved between accounts by an automated system with no consent, no audit trail, and no path to fix it.
- People want a built-in way to make things right, not a lawsuit.
- Federal money for transparency and sovereignty has to reach small organizations, artists, and community groups, not only the platforms.
Already working: Human override for consequential automated actions, audit trails that record who did what and when, and a path to reversal.

Try it

A label is not accountability. Build the receipt.

Ottawa is asking which pieces of transparency to require. The room kept answering with a different question: what can a person do afterwards? Pick a situation like the ones the room raised, switch on the pieces one at a time, and watch what you could do change. The situations are illustrative. The pattern is not.

Step 1 / Pick a situation

Step 2 / Switch on transparency

Transparency receipt01 / 05

A video of you, saying something you never said

It is in a group chat before you have seen it. It looks like you. It sounds like you.

□: A label that says AI was involved: Not disclosed.
□: What it is optimized for, and who built and deployed it: Not disclosed.
□: What data it used, and whether you consented: Not disclosed.
□: A person you can reach who can override it: Not disclosed.
□: A record of what it did, when, and on whose instruction: Not disclosed.
□: A way to reverse it, or be made whole when you cannot: Not disclosed.

What you can actually do

Not yet: 1. Understand. You know a system was involved and what it was doing.
Not yet: 2. Decide. You have enough to walk away, or not.
Not yet: 3. Contest. There is a person and a record to push back against.
Not yet: 4. Correct. What it did can be changed or undone.
Not yet: 5. Be made whole. Someone is answerable when it cannot be undone.

Nothing disclosed. You do not even know a system was involved. This is where several of the room’s stories started.

What the room asked for

Each person should own their face and voice. A deepfake of a real person is a line not crossed.

What was said

Fear and hope, often from the same person.

The group sorted its own responses into two piles: those rooted in fear or risk, and those rooted in hope, benefit, or attachment to what the technology could make possible. Both were present all evening, sometimes a few minutes apart in one voice.

Risk, harm, and what has already gone wrong.

No one consented to it. No one contacted either party. But the actions still happened. They had no traceability, no auditability, and couldn’t explain what happened, who did it, or when.

Fear / On an unauthorized AI-driven bank transfer out of a family member’s account

Am I suing the government? Am I suing the Anthropic? I don’t want to have to sue anybody. What system is going to be in place that has a built-in way to make this right?

Fear / On the absence of a clear path to accountability

This is just a red herring, a fake, pretentious way to appease us. How are they actually going to verify if something was made with AI or not?

Fear / On skepticism toward the consultation’s premise

Someone can just clone a whole me and put it on video. In the future we’re going to need some kind of ownership of our identity. It cannot be duplicated by someone else.

Fear / On deepfakes and identity

If the training data doesn’t include humans with a particular characteristic, a minority, a woman, a child, the system may simply fail to recognize them as human at all.

Fear / On bias embedded in training data

There’s no trust. That’s the whole problem. How do we repair the trust?

Fear / Said in direct response to another participant’s story

Quotes are a selected, high-impact sample from the recording, not a full transcript. Where a speaker could not be reliably identified, and that was most of the time, nothing is attributed. That is by design.

Fear-based responses

No one consented to it. No one contacted either party. But the actions still happened. They had no traceability, no auditability, and couldn’t explain what happened, who did it, or when.

On an unauthorized AI-driven bank transfer out of a family member’s account

Am I suing the government? Am I suing the Anthropic? I don’t want to have to sue anybody. What system is going to be in place that has a built-in way to make this right?

On the absence of a clear path to accountability

This is just a red herring, a fake, pretentious way to appease us. How are they actually going to verify if something was made with AI or not?

On skepticism toward the consultation’s premise

Someone can just clone a whole me and put it on video. In the future we’re going to need some kind of ownership of our identity. It cannot be duplicated by someone else.

On deepfakes and identity

If the training data doesn’t include humans with a particular characteristic, a minority, a woman, a child, the system may simply fail to recognize them as human at all.

On bias embedded in training data

There’s no trust. That’s the whole problem. How do we repair the trust?

Said in direct response to another participant’s story

Hope-based responses

There is a belief that each of us has a musical expression inside us, and AI has reduced the barrier to bringing it out. Now anyone can make that expression and the whole world can hear it.

On AI’s effect on access to creative expression

The idea of a personal data store: you bring your data to a service, and you decide. You can grant access, or cut it off, and keep control. You can’t be cut off from your own data.

On decentralized, self-managed data infrastructure

It doesn’t require 100% of society. If even a small group of people push back, that’s enough to make a difference.

On the scale needed for change

Empowering the human, rather than relying on the government, should be the guiding principle.

On where protection is expected to come from

I’m a technologist. I’m an optimist. I get excited about AI’s capabilities. I also have real fears, like everyone in this room. That combination is what makes me feel responsible to do something about it.

Closing reflection from one participant

Almost every month, we have new people show up. If we can give people something to connect to beyond just a chat thread, we can keep growing this into a real community.

On the value of the recurring discussion itself

Themes

Six threads, and what each one connects to.

For each recurring theme: the risk that was raised, what participants pointed to as a possible opening, and what already exists, whether or not they named it directly.

Identity and authenticity
The risk raised
Deepfakes and cloned voices and faces. One participant described the risk in personal terms: someone could put a false version of them on video, including as evidence in a legal proceeding.
The opening
Open, non-proprietary content-labelling standards exist and are already used by major platforms.
Already working somewhere
C2PA, an open technical standard for labelling AI-generated or AI-edited content, backed by Adobe, Google, Microsoft, and others.
Accountability and redress
The risk raised
No traceable record when an AI system causes financial harm. No agreed answer to who is responsible or how to get it fixed.
The opening
A model where every party in the chain, developer, deployer, institution, carries some responsibility, similar to models already used in other regulated industries.
Already working somewhere
Human review and override for consequential automated decisions, audit trails, and incident reporting modelled on existing safety practice in other sectors.
Consent and data use
The risk raised
Health, financial, and behavioural data used or shared without clear consent. Terms-of-service agreements treated as though they were binding law.
The opening
Interest in personal data stores that let a person or community grant or withdraw access to their own data.
Already working somewhere
Minimum consent standards set by government that override one-sided terms of service, similar to BC’s Residential Tenancy Branch standard lease terms.
Training data and group-level harm
The risk raised
A system trained on incomplete data may fail to recognize people outside that data. Harm to a group or a shared cultural resource, without one identifiable victim, does not fit current legal definitions of harm.
The opening
Interest in documentation of what is in training data, written so a non-specialist could actually read it.
Already working somewhere
Plain-language documentation requirements, and a broader legal definition of harm that includes harm to groups, not only individuals.
Who the benefits reach
The risk raised
Concern that funding and infrastructure investment concentrate among large platforms, while small organizations and local groups are left out, a pattern participants compared to what happened with local news and social media.
The opening
Local and community-run AI infrastructure, using smaller open-source models that run without sending data to outside servers.
Already working somewhere
Directing a share of federal AI investment specifically to community and grassroots organizations, not only large vendors.
Trust in the process
The risk raised
Direct statements that public consultations, including this one, may be symbolic rather than a real input into decisions.
The opening
This session itself, open floor, real time, no single voice dominating, produced substantive, specific input.
Already working somewhere
The group treated this as the central issue, not a secondary one. See the trust finding below.

Theme	The risk raised	The opening	Already working somewhere
Identity and authenticity	Deepfakes and cloned voices and faces. One participant described the risk in personal terms: someone could put a false version of them on video, including as evidence in a legal proceeding.	Open, non-proprietary content-labelling standards exist and are already used by major platforms.	C2PA, an open technical standard for labelling AI-generated or AI-edited content, backed by Adobe, Google, Microsoft, and others.
Accountability and redress	No traceable record when an AI system causes financial harm. No agreed answer to who is responsible or how to get it fixed.	A model where every party in the chain, developer, deployer, institution, carries some responsibility, similar to models already used in other regulated industries.	Human review and override for consequential automated decisions, audit trails, and incident reporting modelled on existing safety practice in other sectors.
Consent and data use	Health, financial, and behavioural data used or shared without clear consent. Terms-of-service agreements treated as though they were binding law.	Interest in personal data stores that let a person or community grant or withdraw access to their own data.	Minimum consent standards set by government that override one-sided terms of service, similar to BC’s Residential Tenancy Branch standard lease terms.
Training data and group-level harm	A system trained on incomplete data may fail to recognize people outside that data. Harm to a group or a shared cultural resource, without one identifiable victim, does not fit current legal definitions of harm.	Interest in documentation of what is in training data, written so a non-specialist could actually read it.	Plain-language documentation requirements, and a broader legal definition of harm that includes harm to groups, not only individuals.
Who the benefits reach	Concern that funding and infrastructure investment concentrate among large platforms, while small organizations and local groups are left out, a pattern participants compared to what happened with local news and social media.	Local and community-run AI infrastructure, using smaller open-source models that run without sending data to outside servers.	Directing a share of federal AI investment specifically to community and grassroots organizations, not only large vendors.
Trust in the process	Direct statements that public consultations, including this one, may be symbolic rather than a real input into decisions.	This session itself, open floor, real time, no single voice dominating, produced substantive, specific input.	The group treated this as the central issue, not a secondary one. See the trust finding below.

The finding

The room doubted the process. That doubt decides whether anything else works.

Distrust of the process is not a separate issue from the transparency questions being asked. It is the condition that determines whether any answer to those questions will be believed. A well-designed transparency measure introduced into a process people don't trust will still not be trusted.

One further observation is worth naming plainly. The same room that expressed distrust of government process also, without being asked to, ran a process with the features described as missing: open floor, no single voice dominating, and enough time for people to actually think together before answering. Whether or not that was intentional, it is a description of what a different kind of engagement could look like.

If you want to be cynical, you could say this is just an exercise so they can pat us on the head and say, we took your feedback, and then give the money to someone else anyway.

On the consultation’s perceived function

I don’t think they’re even going to look at these. There’s too much information. They’ll just feed it into an AI and ask, what’s the general consensus, and put that into a protocol.

On whether individual submissions are actually read

What repair would require

Three things came up repeatedly, from different people, describing the same gap.

01
Plain language
Several participants pointed out that the consultation’s own questions are written in language that assumes prior knowledge most people don’t have. Language that assumes expertise excludes the people it is asking to participate.
02
Time to think, not just time to answer
The group needed a full evening, open discussion, and room to disagree before members could say clearly what they meant. A single survey sitting does not allow for that. Skipping this step produces answers, but not necessarily the answers people actually hold.
03
A visible link between input and outcome
Trust is not repaired by being told feedback was considered. Participants who had filed complaints through existing channels, with banks, privacy regulators, financial regulators, described being met with process, not resolution. The same pattern, repeated at the level of a national consultation, produces the belief that consultation is symbolic.

Score a consultation

Three questions the room would ask of any public process.

Apply them to the ISED survey, a city hall open house, or the next consultation that lands in your inbox. Tick what is true.

Plain languageCould someone without prior expertise answer the questions as written?
Time to think, not just time to answerDid people get room to disagree and work it out before answering?
A visible link between input and outcomeCan participants see what changed because of what they said?

0/3

Consultation theatre. People will answer, and few will believe it mattered.

Methodology

How the room worked.

One recorded, facilitated session, the sixth in an ongoing series, held during the ISED public comment window. Rather than dividing into smaller groups by topic or expertise, the group chose to work through the government's questions together, so people with different backgrounds could respond to and build on what others said in real time.

The facilitator read the government's framing for each topic, then opened the floor. Everyone present was invited to speak, and no one voice took up more time than others. The first questions turned out to be shallow, the conversation kept drifting to deeper issues, and the facilitator steered it back to the list a few times. The group did not complete all 24 questions in the time available, and said so directly rather than rushing to finish.

When: Tuesday, August 4, 2026
Where: Parker Street Studios, Vancouver
Who: About twenty people
How long: One hour and seventeen minutes on the recording
Convenor: Kris Krüg, BC + AI
Facilitator: Peter Van Garderen
Host: Tanya Slingsby, Parker Street Studios
Report: Sara Taherzadeh
Format: Full-group discussion, recorded, with live notes

About twenty people seated in a circle of couches and chairs at Parker Street Studios for AI Ethical Futures Lab number 6, with a laptop open in the foreground and paintings on the gallery wall. — One room, one conversation. No breakouts.

Participants in conversation around a glass table at AI Ethical Futures Lab number 6, one person gesturing while others listen. — The facilitator read the government’s framing, then opened the floor.

Photos: AI Ethical Futures Lab, Tuesday, August 4, 2026

From the room to this page

The recording, the facilitator's live notes, and a transcript cleaned up with an AI tool were published the next day as a public repository. An AI-generated topic summary and a set of draft survey answers were added as inputs, clearly labelled as machine output. The written submission was then composed by a person from all of it. Quotes were chosen because they represent a view that was echoed, built on, or directly responded to by others in the room.

BC + AI files its written response by email to ISED's consultation inbox. Submissions are public documents and may be posted by ISED. Nothing here names a participant without consent.

Timeline

March 4
AEFL #1 runs the People’s AI Consultation
The first room. Jesi Carson designed the process and turned the worksheets into a nine-page public report.
July
BC + AI answers the AI for All strategy
Adoption Is Not the Goal: trust is infrastructure built through accountability, not a literacy deficit to be corrected.
July 23
ISED opens the AI transparency consultation
Five areas, a discussion paper, a survey, and a written route. Open until September 23.
August 4
AEFL #6 works the questions as one room
About twenty people at Parker Street Studios. Facilitated by Peter Van Garderen. Recorded, with notes.
August 5
The public record goes up on GitHub
Recording, transcript, facilitator notes, an AI-generated summary, and the five pre-read briefs.
August 24
The collective report
Sara Taherzadeh turns the room into the written submission this page carries.
September 23
The consultation closes
ISED then reviews submissions and publishes a What We Heard report.

Who we are

The AI Ethical Futures Lab, and why it exists.

The Lab is a recurring, open discussion group within the BC + AI Ecosystem, a network of people across British Columbia working on AI from clinical, technical, creative, legal, and community vantage points. It meets monthly in Vancouver to work through the practical and ethical questions AI raises as they come up, not only after policy or legislation has already been set.

Attendance is open and voluntary. There is no formal membership structure. People come because the questions are relevant to their work or their lives, not because they represent an organization. August was the group's sixth meeting.

It exists because consultations keep arriving with short windows and expert language, and the people most affected need a room to think together before they answer. The Lab carries public-safe notes into BC + AI's responses. Taking part does not mean endorsing a submission, and nothing is attributed without consent.

We have done this before

Come to the next room

The Lab meets monthly in Vancouver. Sessions are small, laptops-open, and built for disagreement that goes somewhere.

See upcoming Lab sessions Open the reading list

About BC + AI

BC + AI is the member-supported nonprofit association growing out of Vancouver AI. Membership keeps rooms like this one open and turns what happens in them into public work.

Join BC + AI

Sources

All sources were accessed directly. Links current as of August 2026.

Credits

Convened by: Kris Krüg, Executive Director, BC + AI Ecosystem Association
Facilitation: Peter Van Garderen, BC + AI member
Host: Tanya Slingsby, Parker Street Studios
Written submission: Sara Taherzadeh
Published by: BC + AI Ecosystem Association

Kris brought the consultation to the Lab, opened the public record with Peter, and carries the final submission to Ottawa. Peter ran the room and published everything the next day. Sara turned the night into the written response. Tanya opened the studio.

By all of us showing up tonight and having this conversation about AI transparency, I think everybody in this room is already doing something about it. That’s worth marking.

Closing statement from the session