---
id: "workshop:how-to-write-a-generative-ai-policy-for-your-jurisdiction"
type: "InnovateUS workshop"
title: "How to Write a Generative AI Policy for Your Jurisdiction"
url: "https://innovate-us.org/how-to-write-a-generative-ai-policy-for-your-jurisdiction"
date: "2023-06-14"
duration: "90 minutes"
language: "English"
instructors: ["Santiago Garces"]
recording: "https://www.youtube.com/watch?v=Y4t-EznsPYQ"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# How to Write a Generative AI Policy for Your Jurisdiction

InnovateUS workshop · 14 June 2023 · 90 minutes

Join Boston's Chief Information Officer, [Santiago Garces ](https://www.boston.gov/departments/innovation-and-technology/santiago-santi-garces)for this 90-minute virtual training, where participants will learn about how he developed the Boston guidelines on using generative AI in public service. Participants will learn about several scenarios in which government workers might want to use AI to improve how they work, as well as the process, decision-making, and recommendations involved in creating a generative AI policy for government.

## Learning goals

-  Learn about several scenarios in which government workers might want to use AI to improve how they work
- Understand the process, decision-making, and recommendations involved in creating a generative AI policy for government

## Speakers

- **Santiago Garces**, Former Chief Information Officer, City of Boston

Recording: https://www.youtube.com/watch?v=Y4t-EznsPYQ

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/how-to-write-a-generative-ai-policy-for-your-jurisdiction.txt (plain text) and https://innovate-us.org/how-to-write-a-generative-ai-policy-for-your-jurisdiction.json (every segment with its start and end second).

**Santiago Garces**

[0:00](https://youtu.be/Y4t-EznsPYQ?t=0) Thank you for taking the time to talk a little bit. So what I thought for the conversation today was, I put together some slides mostly so that I wouldn't forget to make some points, but there are pieces of, hopefully this is a little bit of a conversation. I wanted to share with you what was going through my mind and through some of the conversations that we were having with the mayor and with the leadership team at the city around when we were putting together the guidelines. And there's a couple of points where hopefully it becomes kind of a shared exercise where I can hear from you as you start identifying what are the things that you would want to guard against and what are the things that you think that are opportunities as well. So there'll be some opportunities for participation and then towards the end, we will use some of the tools just to demonstrate again.

[1:03](https://youtu.be/Y4t-EznsPYQ?t=63) So two quick disclaimers, we're not endorsing any particular tool. There's a number of different tools in Boston. We have been, and you'll see it in the guidelines, that we are using BARD from Google, partially because we have an enterprise subscription from Google. So we have the ability of managing that at an enterprise level, but we haven't restricted either. People have been using ChatGPT and other stuff, but we make sure that people know that it is their responsibility to comply with public records requests and all sorts of other things. But let's get started. Share this. All right. Okay.

[2:04](https://youtu.be/Y4t-EznsPYQ?t=124) Everybody can see the screen? Okay. Yes. Super. Yeah. All right. So generative AI is a set of technologies that are relatively new and that leverage really large amounts of data to create models for different things, for vision or for language or whatnot, to try to predict what the right, to make a prediction based on all of the data that it has because of the underlying technology that's being used. Because of the underlying technology, it's proven to be very powerful when it comes to creating or analyzing language and images. And it is a different iteration built on the same type of technology that has been helping us with voice-to-text and Siri and Alexa and all these tools.

[3:04](https://youtu.be/Y4t-EznsPYQ?t=184) But now it gives us, my Alexa's now complaining in the background, but it's a different iteration of tools that we've been using for a while, but that allows us to create something new and novel that didn't exist before. For the most part, we interact with this technology using PrEPs. So we've seen a number of, a couple of months ago when we first started having conversations about generative AI, and there's rich discussion, I think, across the board, but also within City Hall, where like, why should we do guidelines? Why should we think about, why is this important? And there's a number of people that I know that their initial reaction was more of aversion in the sense there's some concerns about whether it was actually a useful tool.

[4:08](https://youtu.be/Y4t-EznsPYQ?t=248) And the number of governments were starting to come up with bans or blocking the access to the tool. But we started to see first, there's a large number of people that reported that it was useful. And I remember my initial reaction was to be skeptical, and it wasn't until I follow some professors that I admire on LinkedIn and other social media, and I started to see some of the posts that they were sharing about how they were using the tool and how useful they were finding the tool, and they're people that I respect. So I figured, you know, I'm gonna set aside my skepticism and use it. So I found it useful. So first, they're useful tools. They can do things that we wouldn't be able to do without those tools easily. Secondly, we were thinking, this is something that is very widely available. So I was looking at some statistics. It took, I think it took Netflix a long amount of time.

[5:11](https://youtu.be/Y4t-EznsPYQ?t=311) I think it was like almost five years to get to a million users. It took ChatGPT five days to get 1 million users. So the adoption and the exposure across the board to these models has been incredible. I get ChatGPT is not the only tool, but when we have technology that is so pervasive throughout society, it is kind of hard to ignore from government because it's out there. Just as a point of curiosity, I built this chart using ChatGPT. I asked it first to give me a chart of the number of users of ChatGPT, and it said that's proprietary information, so it wouldn't make the chart, but then I found some information elsewhere that the data around the usage, and I just copy and paste it without any formatting and said, I want you to help me build the chart, and it built this, which is pretty cool. But third, it is something that is risky.

[6:13](https://youtu.be/Y4t-EznsPYQ?t=373) We know that there are risks and there's potential, there's downsides to the technology. There's also downsides of the technology that will manifest regardless of whether we regulate it or not, or whether we acknowledge it or not, because other people are going to be using it. Again, anecdotally, and we know that there's an increase, for instance, in phishing over the past year, and the quality of phishing has gone up, and why maybe some of these people are using these tools to make. One of the things that we would tell people when we're doing cybersecurity awareness training is look at the quality of the writing as a proxy for whether a message is authentic and true, but if it's authoritative or not, or someone who's pretending to be that person. Now, with some of these tools, you have the ability of having really well-crafted information. You could even have images and seals, and you could have someone who looks like the mayor going in and telling you to give them the information of the city's bank account.

[7:22](https://youtu.be/Y4t-EznsPYQ?t=442) So regardless, the risks are out there, and there's nowhere where we can hide from some of these risks. So we thought the benefits are too valuable, and the risks are too pervasive to just ignore this. On the other side, I forgot to make a slide for this. We have an amazing mayor, Mayor Michelle Wu, who one of her three guidelines for how we should operate is to embrace the possibilities. So there's kind of this risk assessment and opportunity assessment, but fortunately, we have a mayor that was telling us, don't shy away from this stuff that is happening. Seek it out and embrace it and try to understand it. So broadly, where we've landed, and kind of the framing of the guidelines has been one of responsible experimentation. So because the technology is still evolving so quickly, and the ecosystem around the technology is evolving so quickly, we know that if we come up with policy and structure, if things are too structured, maybe the technology itself will change and we'll miss the mark.

[8:39](https://youtu.be/Y4t-EznsPYQ?t=519) So we said, let's create an environment where we can work as a society, as a community, both internally within the city, but also with our constituents, with people from academia, from industry, to create an environment of collective learning where we're controlling the risks, but we are actually fostering and creating an opportunity for people to, we won't be able to understand the risk unless we're experimenting, unless we're trying some of these things out. And we will also not reap the rewards unless we're experimenting and trying things out. Right. So as promised, here's a little bit of, and I think that some of these answers will vary depending on your context and where you are, what country you're in, what organization you work for. Even within the city, when we were putting together the guidelines, I felt that in public education, in education, there's some considerations that I was not going to do justice to, and I was thinking mostly about like, how does the tool fit for public employees rather than for students?

[9:47](https://youtu.be/Y4t-EznsPYQ?t=587) So I said, the guidelines cover only, it does not cover the Boston Public Schools, but let's open it up, maybe put in the chat, or if you wanna just speak out loud, what are some of the things that could go wrong? What are the things that you're trying to, it's almost like drawing a red line. You say, what are the things that you really don't want to mess up when we're experimenting and working with generative ed? So maybe take 30 seconds, put it in the chat. You can also feel free to just unmute yourself and let us know directly or use the chat

**Unknown speaker**

[10:32](https://youtu.be/Y4t-EznsPYQ?t=632) and I'll read it out to Santia as they come in. So, plagiarism, sharing personal sensitive information, racist and other discriminatory content, false information, getting duped, privacy, worry, shutting down experimentation, false empathy, incorrect responses. Santia, we have one participant who's saying they're not getting a clear voice. Is everyone having trouble hearing Santia? Accuracy, bias, and disclosing confidential information, becoming over-dependent on technology, destroys value for money in many professions, negative impacts on marginalized communities, privacy, AI hallucinations.

[11:44](https://youtu.be/Y4t-EznsPYQ?t=704) Okay, thanks for letting us know that the audio issues are not an issue for everyone. Santiago, you're good on the microphone. Widening the gap between school and real work worlds, proper disclosure of uses, perspectives of the human input, HITL, professions getting replaced by AI, efficiency gain going to the wrong pockets, capitalization of user data without permission. Thank you all for submitting these. Santiago, can you just make sure that you're speaking sort of head-on into the microphone? I think the volume is okay, and then some of the words might get a little choppy.

**Santiago Garces**

[12:42](https://youtu.be/Y4t-EznsPYQ?t=762) Yeah, I tried to switch, but somehow Zoom wasn't catching my headphones, so let me, hopefully that makes things more clearly. All right, hopefully this makes the audio sound a little bit better. Do we, so let's pause for a second, and this is great. And a lot of the things that you've talked about have come up with, were things that we discussed. Again, some of you mentioned the societal issues that might precipitate from the use of generative AI, and we talked about in Boston, and the mayor cares deeply about unions and about workers, so thinking about the erosion or the impact in the labor force is one of the things that came up, and many of the things that you've mentioned. But let's talk now, okay, what are things that you think that are opportunities?

[13:45](https://youtu.be/Y4t-EznsPYQ?t=825) Like what are things that we want to make sure that we're trying to achieve? So it's not only the negative, but what are the positive things that we would want to, and more than specific outcomes, what are fundamental principles or values that you think that we would want to end up with? Another, maybe 30 seconds of brainstorming. I'm seeing efficiency, automation, adding capacity to understaffed teams, reduced employee workload for menial tasks, saving time, reaching each learner, more accuracy, more efficiency, more capacity for the important work left. Access to information in unstructured data, high availability of services, file organizations, assisting with writing briefs.

[14:47](https://youtu.be/Y4t-EznsPYQ?t=887) Becoming a curator rather than a creator will allow us to do more creative work. Higher evaluation of human wisdom and wise AI, broader range of information sources all at once. These are great, thank you. Yeah, so okay, I'm going to show you the ones that we were thinking about. So we're thinking there's potentially unintended consequences, there's like we won't be able to think about all of the things that could go wrong, and we figured there's a number of things that could go wrong, and in government we manage public goods and services, and our employees have a fair amount of power over other people, like we're reminded of that, right? There's the potential of losing trust in people. I think that some of you mentioned false empathy, like what happens if the response that you're expecting from government and from an official is not the actual reaction of the official? We care and we value about like having a connection between the people and their public servants. Waste of time and money, so there's a lot of concerns from people, especially from some people saying, is this just overly hyped? People are maybe a little bit too excited about this, but we don't think that this will have a lot of value. Some people, I don't think that I've seen, you all mentioned it, but there's some people that brought up the environmental cost. It takes a fair amount of computing power to train these models, and whenever we use it, we're using some electricity and there's an environmental impact to the use of the models as well. So we started thinking as well, what are things that we care about? What are things that we want to preserve? So we said, look, ultimately this has to be about people. So it has to be about the people that we serve and the outcomes that we can reach, but also the employees. And in some sense, these are kind of meant to be a little bit tied between

[16:51](https://youtu.be/Y4t-EznsPYQ?t=1011) the left column and the right column. And we thought, we place a lot of trust and our public servants have a lot of power in there. Even the people that we don't think have a lot of power, have a lot of power. If you're doing maintenance in a building, if you leave the floor slippery and you don't put a notice, someone might hit their head. And in a lot of public buildings, we have vulnerable people that have mobility issues or whatnot. So again, we're constantly dealing with power from a public servant and the potential impact. So we said, people are both the end and also a means of controlling some of these unintended consequences. When it comes to loss of trust, let's be transparent. I think one thing that people often don't like is having a politician or someone else tell them that they have all the answers when they don't. So let's acknowledge what we know, acknowledge what we don't know, but do it transparently so that others trust us as we keep building this. And again, if there's a sense of risk around the potential for false empathy or for misinformation, let's be transparent about how is it that we use the tool.

[18:06](https://youtu.be/Y4t-EznsPYQ?t=1086) Innovation and risk management. Every time that we do work in government, we are trying to find the right balance of providing value and solving a problem and then containing the risk and the consequences of what we do. So in our case, we thought it starts to seem that there's high enough number of people that are saying that this is valuable, but likely there's a return on investment. A return on investment. There's a reason why doing the experimentation will pay off. And what we have to do is figure out how to contain and manage the risk. And finally, around environmental costs and around some of the other pieces, again, thinking. And I think about this in the context as well. There's a lot of conversations around the potential of AI for bringing an extinction event and some kind of big visionary questions around like, is AI going to replace human intelligence?

[19:10](https://youtu.be/Y4t-EznsPYQ?t=1150) And the thought is, at this point, we already use gasoline in trucks that haul it trash and in police vehicles, in fire vehicles, and those vehicles contribute to emissions. In some sense, again, we already exist in a world where we make trade-offs, where the things that we value in terms of immediate outcomes of public safety, of sanitation, offset environmental costs and offset other costs because we know that we have to operate in society. So again, I'm not claiming these are perfect, but these are some of the things that we're thinking. And I think that as you think about your guidelines or your version of policy and regulation, this was just a helpful exercise on what are the things that we want to steer away from and what are the things that we want to steer towards. So I think one of the pieces that was, I think, the most helpful was thinking this is a tool, like AI, generative AI is a tool, like a spreadsheet. We could do our payroll by hand, but it would take a lot more people and a lot more time. And maybe we wouldn't be able to do payroll as regularly as we do without this tool. So the thought is, the tool serves people. And if we make these, especially at this point, make sure that people are in control and they know that they are expected to be held accountable, that this revolves around them, that this will be, that this helps offset a lot of the concerns and a lot of the potential for negative risks. Giving people and trusting people to experiment, which is not something that we always do in governments. And we tend to be, especially in some parts of government, very hierarchical. But again, it's acknowledging the fact that there's a lot of wisdom in our

[21:13](https://youtu.be/Y4t-EznsPYQ?t=1273) public servants and it is when we help support them that we get the best outcomes. So that's also something that ended up coming up in the guidelines and feedback that I got from some of the collaborators and the people that helped with this, was make the guidelines and make the regulation as simple as possible. If you're someone working in government, and you all mentioned it, we tend to be understaffed with a lot of pressure. The last thing that you want is to create something that is so abstract and so difficult to access. We want simplicity because we want to activate and empower these people. And again, one of the final outcomes was to learn. We want people to have exposure to the tool so that they can see not only how they can use the tool, but how other people in society are going to be using the tool in a way that might impact them. So if you're in IT, people are going to use these tools to try to do phishing attacks and to have like social engineering attacks and to try to... And there's restrictions that some of the tools try to place in some of the usage, but that's a reality, and there's ways of circumventing and using models that are not regulated. Two, our community groups and others are going to use it to create opposition research and to try to think about other alternative points and whatnot, which is not bad, but we have to anticipate that there's other uses and that all of a sudden people have access to other resources to be able to engage with us. So again, these are the guidelines. There's only three guidelines that we have on the rest of the documents that I'm sure Hannah and others will share with you. The rest of it is around things that we recommend that people do, things

[23:14](https://youtu.be/Y4t-EznsPYQ?t=1394) that we recommend that people steer away from and some examples, but at the heart of it, there's these three guidelines. So the first one was fact check and review all the content generated by AI, especially if it was going to be used in public communication and decision-making. And there's another paragraph that we added at the beginning saying, this is a tool. You as a public servant are responsible for the good and the bad that comes out of the usage of the tool. If BARD or Chad DPT puts out something that you use, it's you using it. You are going to be held responsible. The same as if we had a spreadsheet and we had used the wrong formula and all of a sudden we added an extra zero to our budget, it's your fault. It's not the spreadsheet's fault, it's the accountant's fault. We have to verify and we have to be vigilant and we have to be trying to understand what looks good and what doesn't look true and leverage your expertise. Second, at this point we said it's still also new that it is ready to disclose when we use AI to build trust so that people know when it said that it's you and when it said that you've used it. And you can tell others that you've used it to edit, like a lot of times you might use it to help prepare a statement or prepare something. You could say this was edited with the help of Chad DPT or with BARD, but we want to build trust and showing the process is part of the building trust. And the third guideline is making sure that we do not share sensitive information or private information in the prompts. A lot of people are going to be accessing these tools through their free version and based on how we know that the tools work, that information that you put in the prompt is fair game for the providers to take and incorporate into their model

[25:15](https://youtu.be/Y4t-EznsPYQ?t=1515) to try to improve. So if you put your company's proprietary code or you put personally identifiable information or you put information that is regulated, you're actually disclosing information with this third party and you have no control over that. So the second reason why we put that guideline was thinking it also starts discouraging some of the riskier potential uses until we have better mechanisms of controlling for those. But like you think about it, potentially this being used to help support police officers in writing cases and investigation reports and other things. Maybe we're not there yet or if it's going to help with that putting together like this sensitive information about potential suspects and victims, we're not quite there yet and we would want to have better controls around like how is it that the data is being transferred to enable those use cases. So let's pause here and again we can have a little bit of a conversation again. I was going to jump and show maybe together some of the use cases and the things that we found it to be useful. And as a disclaimer again, we'll do it live but there's some sometimes when we're using the tools the output that they generate is not exactly what we expect but that's part of what's good about the tools. We have the opportunity of iterating and giving it feedback to improve. But what are some things that you think that we have missed or what are some things that you think I'd love to hear your reactions or what resonates with you that were that worked particularly well. Maybe like a couple of minutes of discussion and feel free

[27:16](https://youtu.be/Y4t-EznsPYQ?t=1636) to unmute yourself and or raise your hand and we can call on you. And I'll just pause the recording for that. All right so really fantastic conversation. So we didn't again partially because we acknowledge that there's so much that we didn't know that we didn't even assign like a risk a mechanism when we first launched the guidelines. But I just wanted to share with you again as we work through some of these there's applications or ideas that have come up that have different levels of risk. As I mentioned I think that for us like the ultimate control is making sure that there's a human being that there's a public servant that intermediates the output of the of a generative AI tool with the with its use. Whether it's code, an image, text, that there's a human being that is responsible as the person that is soliciting and requesting the tool to output. There's conversations about creating automated chatbots and and whatnot that we if you look at the guidelines we're kind of restricting people of restricting people from doing that because we're not someone would have to review the content before it goes out in under the guidelines that we have. So I figured that we would for especially for those of you that have been curious about the tools but haven't been thinking and some of these are tools are examples that I've learned from other people from Mitch Weiss and from employees of the city. I think I mentioned Mitch is a Harvard professor at the business school who used to be the chief of staff in the city under Mayor Menino. So he's been really thinking

[29:17](https://youtu.be/Y4t-EznsPYQ?t=1757) creatively about this and one of the things that I've learned from him is usually if we're more limited by our creativity rather than what the tool can do and in some cases it is partially there's a change in the paradigm of how we think about how we interact with technology. What is interesting about particularly ChatGPT and BARD is we're used to technology where you have to develop technical skills before you interact with it and that's what kind of sets how many functions or what programming paradigms you know are what limits you and what you can do whether you know data science and you know random forestry. How to train that convolutional neural network. That's what limits you. With these tools, partially what limits you is your imagination of what you can prompt, and some of the most effective people using the tool are actually the people that are using it in plain language. So these are some that we have used so far in the city. Some of them we were using before the guidelines, so when we used it, we actually hadn't disclosed it yet, but now that we have the guidelines, people are starting to disclose. So I'm going to change how I share my screen. I'm going to do some examples in BARD, which is the Google tool, and then some examples in ChatGPT. I've been using the paid version of ChatGPT, but some of these have worked well with the 3.5 version of ChatGPT, or you can get access to the latest version of ChatGPT if you access it through Bing. Let me change how I share my screen.

[31:08](https://youtu.be/Y4t-EznsPYQ?t=1868) With BARD, I'm just going to use this one. There we go. All right. So I was just preparing a little bit here, but I'm going to clear it up so that you just wanted to double check that it was going to do what we wanted to do. So one of the things that I found this to be really helpful for is when we're creating job descriptions. So again, our department has a lot of vacancies, and we are hiring, and in technology, we have the exciting reality that the job descriptions that we wrote 30 years ago are probably not relevant right now in terms of the skills and types of positions that we want to have. So I'll give you an example of one that we wanted to. So we wanted to hire a senior manager of software engineering. So I'm going to ask it, help me write a job description for a senior manager of software engineering. This is where the more specific that you are, the more helpful it can be. So I'm just going to do this prompt, and then you'll see sometimes it gives us some variables. It gives us some stuff that is not specific enough. I'll show you how we can iterate to try to get. So in a few seconds, it's created this job title, this template that looks pretty similar to what we would submit to human resources.

[33:05](https://youtu.be/Y4t-EznsPYQ?t=1985) So it has its title, who it reports to. It says that it's responsible for leadership and management of the software engineering team, responsibilities, its qualifications, and its benefits. So let's say that I want to make it a little bit more specific for my department. Please rewrite the job description. The senior manager will report to the chief engineering engineering officer. This position will be staffed in the department of innovation and technology of the city of Boston. Please add language that emphasizes experience in accessibility tools.

[34:25](https://youtu.be/Y4t-EznsPYQ?t=2065) I will leverage another great automated tool, autocorrect, and I'll see what it does. All right. So again, I don't really have to think too much about how I'm wording stuff. It goes on automatically, interpret, and figure out that the reporting field needed to change to the chief of engineering. It started mentioning in the summary that the software is intended to be accessible to all users. And again, it added. So this is where we say that it gets you 80% of the way there or 90% of the way there. There's still edits that you'd probably want to make specifically around the tools that you use. You can use the tools to keep iterating and making it better. The way that we've done this, the way that I used this in the past, and again, before the guidelines came out, I copied the job description, put it in Google Docs, and then shared it with my team and let them have a round of iteration. And it was amazing because it ended up converging to kind of the original job description, just with a little bit more specificity. And it saved so much time. We would have spent a few hours. What we would probably typically do is go research online about the position, find a couple of different template job descriptions. What the tool is doing is gathering information about all of the positions that it knows about, and then creating something that aggregates all of that. And sometimes when you're writing a job description, you kind of want that collective intelligence You see? It's pretty cool. So let's do another use case that has come up. So as we start building our team, we've been trying to figure out, okay, we want to have engineers that

[36:26](https://youtu.be/Y4t-EznsPYQ?t=2186) work at different levels. So we want to have junior engineers, mid-level engineers, and senior engineers. And both for HR and for a union, we want to make sure that there's very clear criteria around those different groups. So let's go and do, let's say, please make a table that contains some differences in skills, experiences, and compensation for a entry-level software engineer, a mid-level software engineer, and a senior software engineer. I'm going to anticipate, I'm going to say these positions would staff the Department of Innovation and Technology in the city of Buffalo.

[38:02](https://youtu.be/Y4t-EznsPYQ?t=2282) Okay, it's put together a table, which is great, and actually there's a export function that gives us out. It's created a matrix where it's giving us, per each of those verticals, different types of skills. It tells us about the types of programming languages, the software methodologies, software testing, software design. Let's say that I wanted to add, and it's giving us some compensation and experiences. So I'm going to ask it now to see if it can give me, yeah, I'm already gendering this, that it can give me a, I don't know why it seems it doesn't have, but let's assume that I want to get information about compensation specific to the public sector and that I want to get also supervisory. I think that it under other skills, but I want to call it out to give me supervisory duties. Please rewrite the table. I'm going to try to see if I can, I'll just please rewrite the table and add a row for compensation ranges based on for these decisions in the public sector and add another row that shares the differences in supervisory responsibilities.

[40:03](https://youtu.be/Y4t-EznsPYQ?t=2403) You can go make some tea while it's thinking of the, didn't give me even enough time to make some tea. So while you see here, it messed up instead of adding me another row, it created a, it gave me, it gave it to me as a column. So it did get a little bit hung up like last time that I had asked for it, it had done it correctly. But again, it starts giving us some of the information that we wanted. It gave us information about compensation in the public sector and it gave us information about the responsibilities in terms of supervision. The data got scrambled, but again, it got us 80% of the way there. But this is, for instance, one of those places where the fact checking comes in. I know that some of these salary ranges look a little bit suspicious, they don't match what my organization has. So I'm going to have to do a little bit more work. I'm going to have to go to the compensation scales that the city has and try to adjust.

[41:11](https://youtu.be/Y4t-EznsPYQ?t=2471) But this is something that will be really helpful for me as I start planning and designing my organizational structure and thinking about the job descriptions and the duties that they're going to have at the different levels. So I wanted to share that with you. Again, let's do one around support for translation and then maybe some co-generation. So I'm going to switch screens again. I'm going to go to chat GPT.

**Unknown speaker**

[41:46](https://youtu.be/Y4t-EznsPYQ?t=2506) And Santi, just a note that in about three minutes, we'll move to the Q&A.

**Santiago Garces**

[41:51](https://youtu.be/Y4t-EznsPYQ?t=2511) Perfect. I'll do one quick one. So let's do a requesting to generate some code. So let's say that maybe because we just did a technical example, I'm going to do really quickly one around community feedback. So the city of Boston is considering a policy to dispose, to sell a city-owned property to the developers that can't afford to build affordable housing. These pretend to be three different people that might have different perspectives on the policy.

[43:26](https://youtu.be/Y4t-EznsPYQ?t=2606) So again, there's ways in which you can actually create this to make it more interactive. But if we're going to a public presentation, a public meeting, and we want to just even start thinking about the types of perspectives that will be shared, these are tools that are helpful. Big disclaimer, this does not replace community engagement because this is just a computer simulation. And we know that part of the virtue of community meetings is the trust and the relationships that we're building with the real human beings. But I wanted to bring this example because this is the kinds of tools that we know that community members are also going to have. So in some sense, we want people to familiarize themselves with the tools as well, because they will be encountering... These are the kinds of tools that people are going to be using on the other side as well. So if you don't know how they work, it's going to be hard for you to respond. Really quickly, I just wanted to share some quick tips. I know that we're almost out of time. Things that were helpful.

[44:37](https://youtu.be/Y4t-EznsPYQ?t=2677) Getting feedback from other people when we're designing the guidelines and the regulations. So again, I put the Google Doc and I started sharing it with some other people. Initially, people that I thought that had perspectives that were well-informed around this issue, academics and other thought partners. Then we started expanding it to others to help shape. And we wanted it to be... So, first of all, we need to change the guidelines, partially because we know that it should evolve and we have this governance, and a way of making it better, not just like regulations, it's a little bit harder to change. Secondly, iterating incrementally, so we expanded the group of feedback and it helped inform particularly, as I mentioned, it was a little bit drier, a little bit more focused on resources, and people gave me the feedback that it should be more useful, a little bit more practical around like what the people are going to be encountering. Thirdly, creating mechanisms for participation, so we've also created some Google Forms and some tools that allow people to submit concerns that they might have, so we heard people wanted training, so we've started delivering training.

[45:45](https://youtu.be/Y4t-EznsPYQ?t=2745) We also created spaces where people can share about what's working and what's not working. We can tell people we don't necessarily have all of the answers but we have some resources on how to think about like the impact of these tools on graphic design and on the artist communities and others, so we did point people in directions but in some sense there's this sense that we're collectively learning. We also started training some hosting some training sessions so we had a first training session that had over 100 people. And it was really, really great and we hope to continue to be able to do that and also start thinking, how can we do training outside of City Hall and with community members. So, pieces like be collaborative. This is a learning opportunity and I've learned from the conversation that we've had today from your questions and from your perspective so that's it.

* * *

Original page: https://innovate-us.org/how-to-write-a-generative-ai-policy-for-your-jurisdiction · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org