Transcript
Transcript: Tech Demo by Formic AI
[00:00:08 Text appears on screen: Tech Demo by Formic AI.]
[00:00:12 The screen fades to John Medcof.]
John Medcof (Lead Faculty, Canada School of Public Service): Hello and welcome to the Canada School of Public Service. My name is John Medcof. I'm the lead faculty here at the School and I'm happy to be your host for today's session. We are here for the latest installment of the CSPS Tech Demo series. If this is your first time watching, this video series showcases practical demonstrations, innovative technologies developed by Canadian companies, highlighting how emerging tools like data and artificial intelligence can support new approaches to public service work and to service delivery, and I am delighted to be joined for today's session by Daniel Escott, who is the Chief Executive Officer at Formic AI.
[00:00:55 Daniel Escott appears on screen in a separate chat window.]
John Medcof: Daniel is going to give us a demonstration today of an AI platform that is designed and built by his company and specifically for the realities of regulated professions. So, Daniel, welcome to the Tech Demo series and maybe I'll invite you to start by introducing yourself and perhaps telling us a little bit about Formic AI and the product that you're going to demonstrate today.
Daniel Escott (Chief Executive Officer at Formic AI): John, thank you for having me. It's good to be here. My name is Daniel. I am the CEO of Formic AI. I am a lawyer based in Ontario. I live in Toronto, originally from Newfoundland, so you might hear a bit of my accent come out, but I'll try and keep it intelligible as much as I possibly can. Formic started about four and a half years ago on a mission to make AI trustworthy, which sounds ambitious and it is. The product of that research over the last four and a half years has been a new architecture for artificial intelligence that effectively removes information retrieval from generation, and what that really means is we're using large language models for what they're good at, semantic understanding, grammar, syntax, but not information retrieval. It was never really built for that in the first place. So, instead of trying to shoehorn it into doing everything, we figured, why don't we just build a system that does that better and plug the LOM in to what it does best?
And the result of that was an ability to completely eliminate hallucinations. We can provide citations to source material across any size of a database, the citations cannot be hallucinated, and we can guarantee that in our responses, it will never be able to invent information and it will not just give you what it believes to be the most probable answer, and the easiest way that I can describe that is the ability to prove a negative. AI, generally, if it cannot find what it believes to be the most accurate answer, it will just make one up because it's working on probability at the end of the day. Our system works on absolute certainty. So, if it cannot find what it definitively believes to be the correct answer to your question based on the information it has available to it in your database, then it will just say, I can't find that for you, I'm sorry, and that one change completely alters the nature of artificial intelligence as we know it, and that's what I'm excited to show you today.
John Medcof: Excellent. I really see a lot of applicability to that in our public service context. So, why don't you call up your screen and maybe walk us through the demo so we can get a sense of how the platform works?
Daniel Escott: Absolutely. So, I'll share my screen now. Let me know when it pops up for you. There we go.
[00:03:57 A webpage appears on screen that Daniel Escott scrolls through, displaying a list of documents that each have a series of highlighted category tabs below them.]
John Medcof: Yeah, I can see it, and maybe start by walking me through what it is that I'm looking at here.
Daniel Escott: So, what we're looking at right now is what we call the Formic Engine. This is, at least for now, our homepage, and we have it loaded right now with a whole bunch of test documents. This is our internal server right now. I have documents from federal court decisions, I have patent applications, ISO standards, Labour Board documentation, tax returns, a little bit of a hodgepodge of everything, and this was just to get an idea of what a representative sample of a diverse data set looks like. It's not just legal stuff, it's not just tax stuff, it's not just policy. It's a little bit of everything because in the nature of the public service and my own experience with it, that's what we work with on a daily basis.
John Medcof: Yeah.
Daniel Escott: So, what I wanted to show first and foremost is just how we get everything in here. When we upload documents, the first thing that we do is we scrape everything. We do not import any metadata. We don't try and figure out what we need to reconcile between the information on the file and the information in the database. We populate all of our own metadata based on the system that we developed, and we also summarize the contents of each file when we ingest it, and that's really to avoid my own pet peeve with LOMs which is if you give the same document to ChatGPT that I give to Microsoft Copilot and we ask it the exact same thing, it'll give us two completely different answers. This at least guarantees that everyone on this system sees the exact same information, whether you're looking at it today or in 20 years.
[00:05:51 Daniel Escott clicks on an arrow beside the first document titled "2025fc1144.pdf" whose highlighted tabs read: "Case", "en", "Administrative Remedies", "Appeal", "Evidence", "Judicial Review", "Practice and Procedure", "Public Administration", "Taxation", "Show More". A summary loads below it which reads: "Barbara T. Mendiola received CERB for eight periods between March 15, 2020, and September 26, 2020. The CRA determined she was ineligible because she earned over $1,000 during the applicable periods. Mendiola requested a review and, upon settlement of the initial judicial review, a second review was conducted. The second review confirmed her ineligibility. Mendiola sought a second judicial review, arguing the decision was unreasonable and that the CRA breached procedural fairness. The court dismissed the application, finding the CRA's decision reasonable and procedurally fair, as Mendiola earned over the $1,000 threshold despite her job loss and had been given opportunities to provide evidence of her eligibility. The court acknowledged Mendiola's financial hardship but emphasized the CRA's adherence to the CERB Act's eligibility criteria."]
Daniel Escott: When we have everything into the database, you can see we have the individual files here, the summary that I just mentioned pops up right here, and you see these highlighted tabs. We call them tags, but this is the metadata that we populate. We have a standard metadata scheme that we developed that can be modified for any kind of domain or organization or use case.
[00:06:11 Daniel Escott scrolls up to the top of the page to show filters:
"Language (EN, FR, ZH, HI, ES, AR, PT, RU, IT, DE)"
"Document type (Case, Legislation, Factum, Statement of Claim, Affidavit, Evidence, Patent, Trademark, Contract, Memo, Other, Intellectual Property, Notice of Application)"
"Area of Law (Access to Information and Privacy, Administrative Remedies, Appeal, Arbitration, Bankruptcy and Insolvency, Business, Child Custody and Access, Child Protection, Citizenship and Immigration, Commerce and Industry, Constitution, Contracts, Creditors and Debtors, Criminal or Statutory Infractions, Damages, Defences, Environment, Evidence, Family, Guardianship, Health and Safety, Indigenous Peoples, Insurance, Intellectual property, International, Interpretation, Judicial Review, Labour and Employment, Motor Vehicles, Municipalities, Negligence, Practice and Procedure, Professions and Occupations, Property and Trusts, Public Administration, Residential Tenancies, Rights and Freedoms, Search and Seizure, Sentencing, Support and Maintenance, Taxation, Torts, Wills and Estates, Young Offenders".]
Daniel Escott: But this is a metadata scheme that we had developed for a client of ours that's a law firm in Ontario. So, we display the language of the document with the type of document based on parameters that we set, what areas of law it might be relevant to, but this could be anything. It could be engineering specs, it could be legislation, it could be policies, it could be internal documents. All of this is done dynamically.
[00:06:34 Daniel Escott clicks on the "Show More" tab under "2025fc1144.pdf" and additional tabs are shown, "review", "applicant", "judicial review", "cra", "court", "cerb", "court found", "cra officer review", "cerb act", "Mendiola", "review officer", "job loss", "loss", "judicial", "eligibility", "cra officer", "procedural fairness", "job", "cerb eligibility".]
Daniel Escott: And we also populate, on an individual file level, the key terms, phrases, and concepts that appear on a regular basis in that document, both so we can find it but also so we can find other similar documents that discuss the same thing. Now, everything is in the database, there's only really two ways to interact with this data. The first is to query against everything. So, everyone in your organization would have access to a shared database, whether that's something like GCdocs or it could be as specific as just a folder of documents that your specific team is working with.
[00:07:17 Daniel Escott types "public policy on artificial intelligence" into a search bar at the top of the page and clicks "Search".]
Daniel Escott: But if I were to run a query here like public policy on artificial intelligence, topical.
[00:07:28 A new list of documents appears.]
Daniel Escott: The first thing that's going to come up very shortly is our non-generative search results.
[00:07:30 Daniel Escott clicks on an arrow beside the first document titled "AI_RMF_Playbook.pdf" and the first of its 54 sources loads below it. He clicks through the first nine sources.]
Daniel Escott: These are ranked search results, good old-fashioned, old-school Google style where we identify all of the relevant sources in every document in the database. It's roughly 80% accurate. It's not these are the most likely to be relevant, it's this is everything in the database and we show you exactly what in each of those documents is relevant.
[00:07:57 Daniel Escott clicks on an arrow beside the second document titled "1761729327847.pdf" and the first of its 32 sources loads below it. He clicks through the first three sources.]
John Medcof: Each of these lines I'm seeing now then is an item from the database that I can explore further. Is that the idea?
Daniel Escott: That's exactly it.
John Medcof: Okay.
Daniel Escott: Because the idea is we're not trying to let the AI decide what's relevant for you. If you're working with this kind of information in this kind of setting, you're probably trustworthy enough to figure out for yourself what's most relevant to what you're looking for. Our job is just to find it for you and present it in a way that cuts out all of the time scrolling through hundred-page PDFs looking for two different keywords. The next thing that comes up in the search is our briefs, and these are AI-generated.
[00:08:38 Daniel Escott clicks on an arrow beside the heading "Brief" at the top of the page and a short summary loads below it which reads:
"The provided documents discuss various aspects of public policy related to artificial intelligence (AI). The OECD, Boston Consulting Group, and INSEAD Business School have collaborated on research exploring the adoption of AI in firms, with a focus on policy implications, barriers to adoption, and potential solutions. Their findings, based on a 2022-23 survey of AI-adopting enterprises, indicate that AI adoption is relatively low and concentrated in larger firms and specific sectors. Obstacles include a lack of digital readiness, high implementation costs, and a scarcity of skilled workers. The study highlights the importance of public sector support, such as information services and training initiatives, in facilitating AI adoption, and emphasizes the need for clearer regulatory frameworks and better policy evaluation File 1761729327847.pdf. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14].
The National Institute of Standards and Technology (NIST) has developed the Artificial Intelligence Risk Management Framework (AI RMF 1.0) to guide organizations in managing AI risks. This framework is voluntary, rights-preserving, and non-sector-specific, offering flexibility for various organizations. It outlines characteristics of trustworthy AI and provides functions to address AI system risks, aiming to foster responsible development and use of AI systems File NIST.AI.100-1.pdf. [15] [16] [17]
In the context of legal systems, Alan Diner's report examines the digital strategy of the Federal Court of Australia and its lessons for Canadian courts. It discusses the current state of technology in Canadian courts, including AI, and provides recommendations such as developing AI guidelines and technology education for judges and staff File alan-diner dine…t public-v3.pdf. Additionally, a report on Generative AI (GenAI) suggests that court systems can enhance efficiency and accuracy by integrating GenAI solutions, emphasizing the need for tools designed for legal professionals, integration with existing technologies, and adequate staff training File how-genai-impro…s-tr4897289.pdf. [18] [19]
The adoption of Generative AI in enterprises is also a subject of study, with a report indicating that GenAI usage has become mainstream. This report highlights that accountability is a key consideration, with firms measuring ROI and shifting budgets to performance-proven programs. However, it also points out a shortfall in capability building, with challenges in training and hiring advanced talent File 2025-Wharton-GB…Full-Report.pdf. [20]."]
Daniel Escott: We have our short brief, which is more of a… I call it a quick and dirty, give me a file and two or three sentences to explain why it's relevant, file name, two or three-reasons why it's relevant.
[00:08:54 Daniel Escott hovers the cursor over the second footnote and additional text is shown which reads: "1761729327847.pdf (Page 133) THE ADOPTION OF ARTIFICAL INTELLIGENCE IN FIRMS ©… y of current and other support mechanisms."]
Daniel Escott: But you'll see here, we have a lot of highlighted footnotes and this is something that was proprietary to our system that we spent a solid three years developing. These hyperlink citations link to the document, the page number of the information that's being cited, and the specific text that's being cited, and these citations are not AI-generated because we were able to completely remove information retrieval from generation, which means completely separate from the hallucination proof of the response, these citations are the original source material for each part of the response.
John Medcof: That's really interesting. Can I maybe just pause there?
Daniel Escott: Absolutely.
John Medcof: So, this brief, the summary itself is generated by AI but those links to the original source document will take me to the document itself and not an AI generation of it.
Daniel Escott: That's exactly right.
John Medcof: Okay.
Daniel Escott: And we wanted to make sure that we're not just saying, here's all of the places that we found this information, like what ChatGPT or Copilot does, have fun looking for the actual content. So, we built it in a way where it's on a paragraph level. We can get more granular, but we found that a paragraph makes it a lot more palatable for users. So, these 14 citations were used to build this paragraph, these three citations were used to build this paragraph but not the first paragraph, and so on and so forth. So, we can very deterministically link, this part of the response came from these 14 sources and only these 14 sources.
[00:10:34 Daniel Escott clicks on the "LONG" tab next to the "Brief" heading and an alternate summary loads below it which reads:
"Consolidated Search Report: Public Policy on Artificial Intelligence
Public policy on artificial intelligence (AI) is a rapidly evolving and critical area of focus across various international and national bodies, legal frameworks, and industry sectors. The overarching themes emerging from the provided reports include the need for clear regulatory frameworks, robust risk management, public sector support for AI adoption, ethical considerations, and the application of AI in specific domains like the legal and criminal justice systems. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14]
1. Regulatory Frameworks and Policy Development:
Need for Clearer Regulations: Multiple documents (File 1761729327847.pdf, File The Sedona Cana… June 2025 2.pdf, File alan-diner dine..t public-v3.pdf) consistently emphasize the need for clearer regulatory frameworks and better policy evaluation to guide AI development and deployment. This includes addressing the evolving legal landscape surrounding AI (File The Sedona Cana… June 2025 2.pdf). [15] [16] [17] [18] [19]
International and Jurisdictional Approaches:
The EU's GDPR and AI Act are frequently cited as examples of comprehensive regulatory responses (File The Sedona Cana… June 2025 2.pdf, File AI RMF Playbook.pdf, File ISO 31000 2018.pdf). [19] [20] [21] [22] [23] [24] [25]
Illinois' BIPA is also mentioned as a relevant regulatory framework (File The Sedona Cana… June 2025 2.pdf). [24 [25] [26] [27] [28]."]
John Medcof: Why is that important? Why is that important from a client perspective?
Daniel Escott: If for nothing else, it's important for auditability and explainability. One of the biggest issues that I've seen, not just in law but in every sector that we've deployed this technology in so far, the biggest piece of feedback outside of hallucinations is the ability to understand why you're getting the answer you did. What did the AI consider when it's giving you a 37-page report? Where did it come from? Are these actual sources or did it just make any of this up? This was our way of safeguarding against it. So, the ability to not just present the information and the citations but to be able to cross-reference that with the source material in real time adds a layer of trust that AI in its current form is just incapable of.
John Medcof: That's great. Thank you.
Daniel Escott: And just to close the loop on our briefs, we also offer a longer version that's more of a thematic engagement with the contents of these files but this is more of an exhaustive explanation of everything in the search results and why it's relevant.
[00:11:48 Daniel Escott scrolls through the long brief.]
Daniel Escott: So, where the short brief is just a here's-what-you're-looking-for-and-here's-where-to-find-it, the long brief is a thorough and exhaustive answer to your question, whatever it is.
John Medcof: And I just go back and forth between those two just by pressing the short or long tabs I see in the centre of the screen there.
[00:12:07 Daniel Escott clicks the "SHORT" tab and then the "LONG" tab.]
Daniel Escott: That's exactly right.
John Medcof: Okay.
Daniel Escott: And the only other way to interact with this data is to put information into a project.
[00:12:15 Daniel Escott clicks the "ADD REFERENCED FILES TO A NEW PROJECT" tab at the top of the page and a new page loads, showing a selected document titled "Matcher and Complier Patent Application.pdf" on the left-hand side and the long brief on the right-hand side.]
Daniel Escott: And we can get there by loading a search result in there or we can just open a brand new one and start loading files into it, but I've taken that long brief and opened up a project with it and we see the query and the search result are the first things to come in, but all of the documents that have been cited with 158 citations in this result are now a part of this project folder.
[00:12:38 Daniel Escott clicks the "Documents" tab at the top of the right-hand side and a list of documents appears on that side.]
John Medcof: Okay.
[00:12:44 Daniel Escott clicks the "Reports" tab at the top of the right-hand side and a list of document summaries appears on that side.]
Daniel Escott: When we run our report, what we're going to see at the very top of this is a high-level summary of the entirety of the contents of this folder. So, we have one paragraph that succinctly explains the connected themes and tissues of all of these documents that are loaded in, and each of these files has a context-aware summary that we created for this document when we first ingested it. So, for something like an academic document like the Sedona Canada Primer, we have more of an academic abstract. But if we were to move over to something like a federal court decision, we have a whole legal research memo. And once again, if we move over to something like a patent application, we break out all the key information on the patent, and all of that is context awareness from our system. When we identify what type of document it is, that dictates what the most important information is going to be in that document and we pull all of that out when we first receive it.
John Medcof: On the right-hand side of my screen here, I'm getting a summary of every one of those documents, but then I can go in and look at each one individually. Is that it?
Daniel Escott: That's exactly right.
John Medcof: Okay.
Daniel Escott: And all of these summaries, as well as the conversation that we'll jump to in a second, are all backed by these same citations.
[00:14:10 Daniel Escott clicks on one of the citations from one of the summaries and the document shown on the left-hand side scrolls down to a highlighted passage. He clicks on other citations and the document scrolls to different highlighted passages.]
Daniel Escott: So, if I just want to cross-reference just the summary for this patent application, I can start jumping back and forth, and you'll see on the left-hand side of the screen, we're in the document right now. And as I navigate through these citations, we're jumping not just to the page that's being referenced but it's actually highlighting the specific text that the AI is referencing here.
John Medcof: Interesting. Wow!
[00:14:34 Daniel Escott clicks the "Conversation" tab which generates the long brief once again. He clicks on one of the citations and a new selected document titled "1761729327847.pdf" appears on the left-hand side, with one passage highlighted. He clicks on other citations and new selected documents appear, showing different highlighted passages.]
Daniel Escott: And if we were to jump over to the conversation and start doing the exact same thing with our long brief response, if I were to start cross-referencing this first point with need for clearer regulations, it's going to jump to each document that's being cited, it'll show me what exactly is being cited for each paragraph, and all of this is being done dynamically in real time. So, all of these citations are unique to this one response. And when I ask a second question in a moment, we'll get another response with completely different citations.
[00:15:15 Daniel Escott types "What does the Federal Court say about AI?" into a query bar at the bottom of the right-hand side and clicks "Send".]
Daniel Escott: So, if I wanted to zero in on, what does the federal court say about AI? Because we do have, I believe, two federal court decisions and a report from a federal court judge in here, this is going to exhaustively comb through the contents of this entire folder. It'll come back with a couple of paragraphs explaining what is in this folder that's relevant, why is it relevant? And with those citations, it'll show me exactly where I can find that, and the last thing that we can close with is once we get this response, I'm curious to try and prove the negative as I mentioned earlier. So, I'd like to run one last query just to see what that would look like in this context and show everyone that we are in fact able to prove that something doesn't exist.
John Medcof: Yeah, I'm really curious to see that part of the activity or that part of the demonstration.
[00:16:14 An extensive answer to the query is generated on the right-hand side.]
John Medcof: So, on the right-hand side there now, I see the search has finished. So, it has identified the sources to your more specific query?
[00:16:25 Daniel Escott clicks on various citations and new selected documents appear, showing different highlighted passages.]
Daniel Escott: That's exactly right, and you'll see we've started at number one again because all of the citations are unique to each response, but it's showing me in real time, here's what I found, here's why it's relevant to what you're looking for, but most importantly, where is it in your source documents? And that I found to be most important in contexts that are relevant to the public service. Because at the end of the day, it's not sufficient in most use cases to just find what you believe to be the right answer. You have to be able to show where in the policy document or where in the legislation or where in your internal documents you're finding that information, and why is it relevant? And a lot of that right now is just done either through institutional knowledge and people who've been around the block a lot or you're doing some sort of fancy keyword search or in some cases a Boolean search and you have to go through and audit all of these results yourself.
John Medcof: I can see where if I'm trying to determine from where I, for instance, have obtained a specific authority, having a summary of the authority is one thing but being able to go in and actually see the specific paragraph without having to do the combing through of a large piece of legislation myself, I would see as being highly applicable.
Daniel Escott: Exactly, because, I mean, you see here the study leave report that we're seeing on the left-hand side of the screen is 149 pages. For me to try and find all of the reference material in this myself, that's going to be a solid hour of reading, and I'm a fast reader. But here, not only do I have access to all of that information but I can interact with it in real time and dynamically.
[00:18:12 Daniel Escott types "What does the United States Executive Branch think about AI?" into the query bar and clicks "Send".]
Daniel Escott: Now, to close by return, if I were to ask, what does the United States Executive Branch think about AI?
John Medcof: And you're guessing that there will not be a response to this particular question in the data set. Is that it?
Daniel Escott: I would be very, very surprised. I will admit I had a meeting with the team at the Canada Revenue Agency two weeks ago, and the system actually proved me wrong (laughs).
John Medcof: Okay (laughs).
Daniel Escott: Which is quite fun because that's the nature of this. It's not trying to do work for you. It's trying to make doing your work faster and easier, because all of this time spent just looking for information is not time adding value or performing actual tasks. It's the legwork that has to be done.
[00:19:08 An extensive answer to the query is generated on the right-hand side.]
My God, I was wrong indeed. I'll admit I was wrong but just to close my thought, the whole idea is that all of this time not working, just finding the Lego blocks of what I need to be able to actually do my job, we don't need to waste that time anymore. It's work that no one likes doing, it's usually very difficult to do well, and this allows us to do that work in a way that AI is being used ethically and responsibly. It's secure. It can be deployed on cloud or on premise. But at the end of the day, it's not trying to replace anyone. It's trying to make everyone's lives easier.
John Medcof: So, it's an enhancement of a human capacity as opposed to a full replacement, although I could see it saving us a lot of time. Now, in this example you've provided, it turns out there was.
Daniel Escott: It turns out there was. I was wrong.
John Medcof: That responded to your query. But if I understood your point from your introduction clearly, if you were to ask it something to which it did not have an answer, rather than trying to invent one, it would just give you the transparent response that nothing has been found. Is that correct?
[00:20:34 Daniel Escott types "Can artificial intelligence replace Canadian workers?" into the query bar and clicks "Send".]
Daniel Escott: That's exactly right. I want to try maybe one quick thing asking, can artificial intelligence replace Canadian workers? I'm curious to see what the response on this is going to be because at the very least, this requires a level of analysis that I don't think would be in these documents. But again, I stand to be proven wrong, but all of this is to say we are trying to provide access to this kind of technology in responsible forms in ways that actually add value without creating more risk. I don't think we need to sell what the value of AI is at scale. We need to sell ways to access that value without jeopardizing people's rights, people's privacy, people's security.
[00:21:40 An response to the query is generated on the right-hand side which reads: "The provided documents do not directly state whether artificial intelligence can replace Canadian workers. However, File 1761729327847.pdf discusses studies on the effects of AI on labour demand, noting that some studies suggest significant disruption as cognitive tasks are substituted by AI. One early example estimated that 47% of total US employment could be automatable over one or two decades. Other studies, accounting for the distribution of automatable tasks, concluded that around 9% of jobs around 21 OECD countries are automatable. More recent research suggests that almost 40% of global employment is exposed to AI, rising to around 60% in advanced economies due to the prevalence of cognitive-task-oriented jobs. The document emphasizes that these results depend on assumptions about the rate of AI adoption. [1] [2] [3] [4] [5] [6]"]
Daniel Escott: See, there we go. The provided documents do not directly state whether artificial intelligence can replace Canadian workers. However, it's provided a source here that discusses a much more abstract level and it explains what it says, why it's relevant, and once again, where I can find it. We've proven the negative, and I suppose what I didn't anticipate is it actually showed me where I could find related material anyway.
John Medcof: Okay, that is very interesting. Now, you have mentioned that you're a lawyer by background and that you've done work with different kinds of organizations. Are there specific sectors that you're sort of envisioning as the ones where this kind of tool would be particularly relevant or helpful and why?
Daniel Escott: Well, John, it's a great question. The first place that I wanted to start was obviously law, because hallucinations, explainability, security are all highly prized in the legal sector, but what we found is that areas like accounting, banking, finance, policy, national security, even heritage to a certain extent all have an element of, for lack of better phrasing, if the AI system that we are using makes something up or we can't find where it came up with this information, bad things happen. It's not a scenario where if ChatGPT writes your newsletter and it types someone's name wrong, you fix the newsletter, you move on with your life, but these are areas where decisions are made every day, funds are being deployed, and they have consequences in the real world that affect thousands or millions of people. So, we can't afford the kind of risk that traditional AI, which sounds weird to say only five years out, but these are the areas where bad things do in fact happen when AI inevitably messes up. So, we wanted to have a specialist alternative where the parameters that affect these people can be directly addressed and mitigated. It's not an afterthought. It's not an add-on. It is at the very core of the architecture that this technology is built on, and that's what I think our public service needs.
John Medcof: Okay, so, that sort of trust and confidence in the validity of the information and being able to go directly to the source, as you said at the beginning, leveraging the best of what AI can do but not asking it to do the things where other formats or other technologies might be better suited to the role.
Daniel Escott: That's exactly it.
John Medcof: And maybe one last question for you, Daniel. With the organizations you're working with now, is this something that that platform or that interface would appear on the desktop of every employee in the company? Is this something that is designed sort of for everybody to be accessing the data that is relevant to them? Is it really sort of people in specific roles in the organization? Is it sort of general use or specified use within a company or an organization?
Daniel Escott: We found that the value of this technology scales exponentially the more data you have and the more people that need to access it. So, our best practices that we would typically deploy, if we're deploying in cloud, it would be accessible through a website. If it's deployed on premise or on a local network environment, it would be accessible either through the network environment or the desktop itself. But in either scenario, we try and go as broad as possible and we try and include as much data as possible because the system is not trying to cater to a specific audience or bias toward certain use cases. I call it a digital librarian. It's, at the simplest, most abstract level, an interface for you to come to and say, I am looking for this information, is this information here, and if so, where can I find it? And it'll go back in the stacks, it'll comb through everything, and it might come back and say, you know what? There's nothing here about that and I can't provide any of that information for you, but it might also come back with a stack of books and say, here's exactly what you're looking for, here's the pages that you'll find all of this in each document, and if you have any other questions about it, let me know. It's not doing your work for you, it's not filing your taxes, it's not providing legal advice, but it'll certainly take all of that legwork going through the digital stack, so to speak, and put all of that in one digestible format for you.
John Medcof: That is a great summary. Thank you very much. And maybe before I close, any final messages you'd like to share sort of about the platform, its usability, its application? Any sort of final words from you on what we should be thinking about here?
Daniel Escott: Certainly. Well, I have one note on Formic in particular and one perhaps on AI more broadly. On Formic, we are part of the Innovative Solutions testing stream. So, any teams that would be interested in trying out this technology, we can go through that testing stream and get people's hands on this to start playing within different use cases, whether that's on the engine or through an API, because all of that functionality is modular, we can carve that out one way or another. On AI more broadly, I think we need to be aware of the shifting circumstances about the AI sector in particular. We now have instances where even Canadian companies that are developing AI, they're training their models on data that they've stolen from American copyright holders and there's litigation on that and there's concerns around privacy and ethics and security and ultimately who owns not just the data that's being trained on but the data that's being provided to all of these different models, and that's not just a concern for sovereignty. I think there's a concern about just the viability of these kinds of tools for systems that interact with the Canadian public.
When we know that something like ChatGPT or Claude, when they're trained on mostly American data from the internet, it is providing a 97% default bias towards U.S. information and U.S. institutions instead of Canadian sources and Canadian material. There's only so much that prompt engineering or guardrails can do without raising those kinds of concerns. And when we're talking about not just A.I. systems that are being integrated in government environments but that do interact with regular everyday people and the services that they're trying to access or the rights that they're trying to protect, I think we need to put that on a pedestal, because the reality is, whether we're supporting Canadian innovation or just innovation in Canada, what this technology can do is exceptionally transformative if it's done right, if it's done responsibly, if it's done ethically, and if it's done securely, but all of that starts with trust. If we can't trust it and the public can't trust it, you can have the best guardrails in the world but it's not going to get you over the line.
John Medcof: That's maybe a great reflection to close on. Daniel, thank you very much for joining us today and for coming and demonstrating the platform and sort of walking us through how it works and explaining the value of understanding the sources of our information and the tools that we can use to access it. I really appreciate you taking the time. Thank you so much.
Daniel Escott: Thank you very much, John. Happy to do it.
[00:29:49 The CSPS logo appears on screen.]
[00:29:55 The Government of Canada logo appears on screen.]
