---
id: "workshop:generative-ai-for-public-sector-communicators-tools-ethics-and-best-practices"
type: "InnovateUS workshop"
title: "Generative AI for Public Sector Communicators: Tools, Ethics, and Best Practices"
url: "https://innovate-us.org/generative-ai-for-public-sector-communicators-tools-ethics-and-best-practices"
date: "2025-06-10"
duration: "60 minutes"
language: "English"
instructors: ["John Wihbey"]
recording: "https://www.youtube.com/watch?v=77KbhQFISzg"
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"
---

# Generative AI for Public Sector Communicators: Tools, Ethics, and Best Practices

InnovateUS workshop · 10 June 2025 · 60 minutes

This workshop introduces public sector communication professionals to a range of generative AI tools, with a focus on their capabilities, limitations, and real-world applications. Participants will explore how these tools can be thoughtfully integrated into public communication workflows while upholding standards of quality, transparency, and ethics. The session will offer practical guidance on choosing the right tools, developing safeguards, and adapting workflows to take advantage of emerging capabilities. Together, we’ll explore two guiding questions: Which AI tools and features are most useful in public communications work, and what skills do professionals need to use them ethically and effectively?

## Learning goals

- Understand the capabilities and limitations of leading large language models (LLMs) and generative AI tools in the context of public sector communications
- Compare commercial AI models and evaluate how they can support ethical, effective workflows and content strategies in public-facing work
- Apply generative AI tools to core challenges in public communications and assess methods to improve strategy, efficiency, and engagement

## Speakers

- **John Wihbey**, Director of the AI-Media Strategies Lab (AIMES Lab) & Associate Professor, Northeastern University

Recording: https://www.youtube.com/watch?v=77KbhQFISzg

## 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/generative-ai-for-public-sector-communicators-tools-ethics-and-best-practices.txt (plain text) and https://innovate-us.org/generative-ai-for-public-sector-communicators-tools-ethics-and-best-practices.json (every segment with its start and end second).

**Agueda Quiroga**

[0:00](https://youtu.be/77KbhQFISzg?t=0) So, David.

**John Wihbey**

[0:03](https://youtu.be/77KbhQFISzg?t=3) Thanks again, I really appreciate it. And thanks everyone for being here. I'm really thrilled. I think we had several hundred people signed up on the registration and it's really healthy attendance. So I'm just really honored that everyone's here to learn a little bit more about generative AI and some of the uses of it in communications practices, whether you're doing media and content, or whether you're sort of more upstream on in terms of doing research and other things that support communications work. I'm John Whibbey. I've been at Northeastern for about 10 years now. I teach a course called AI and Media Industries. I'll talk a little bit more about that as we sort of move along. I'll actually show some examples of the work my graduate students have been doing. I have a sort of very new lab, very fledgling lab called Ames Lab. Strategies Lab, please do reach out. I'm interested in connecting with more professionals out there.

[1:06](https://youtu.be/77KbhQFISzg?t=66) And if you're interested in a talk or some advice or you have a research project in mind, please do reach out, ameslab.org. Okay, so our agenda today, I structured this with a kind of caveat in mind, which is I don't know the levels of all the learners here. Some people may be quite sophisticated with Gen AI and may have workflows built up that are quite complex, while others may just be sort of new learners. So what I decided to do was to kind of front load some of the more, I think, sort of cutting edge, sort of leading trends stuff, and then we'll kind of back into some of the basics and the fundamentals, and then we'll kind of return to some of the more sophisticated stuff. So it's a bit of a sandwich here in terms of the level of the content and the level at which it's pitched. So I'm gonna start with a kind of point of view on the state of the art, and talk a bit about customization and integration.

[2:08](https://youtu.be/77KbhQFISzg?t=128) We'll then talk about some of the fundamental sort of skills and the landscape, some key terms and concepts. I'll look at some use cases in visual images and data visualization. I'll talk a little about this sort of next level of sort of simulations and audience persona, sort of synthetic audiences, and how that can be used in communications workflows. And then finally, I'll end with some shout outs to some tools that my network is sort of lifting up. So I have a lot of people who are in the sort of communications business, biotech, environmental ethics and climate change, as well as like all manner of different like spaces within the social sector. And so I'll shout out some tools, but I'll also shout out a couple of podcasts that I think would be really helpful in terms of keeping up with the latest. Okay, so our goals today, and I wanna sort of debunk some sort of issues around the use of generative AI in communications work.

[3:17](https://youtu.be/77KbhQFISzg?t=197) Our goal really is to make you, your team, faster, sharper, more informed by context research and data, and to leverage tools that can adhere to and boost your core values and your brand identity. This is mostly about early to middle stages of workflow. So workflow upstream, whether that's press releases, media advisories, newsletters, social media captions, data viz, meeting notes, research. So not necessarily the stuff that reaches the public, which is where I think the human has to be in the loop. And we'll talk a lot about sort of information ethics as we move through. And so we're not talking about automated content farms, we're talking about human directed workflows. Okay, so if we were coming at this from an academic point of view, these would be the big kind of ethics principles. I wanna foreground ethics right away, even as we get into kind of practice and the nitty gritty. And in bioethics, there's sort of four core pillars that are long established, beneficence, non-maleficence.

[4:23](https://youtu.be/77KbhQFISzg?t=263) You know, this is the kind of do no harm sort of concepts, autonomy, respecting individual autonomy, and achieving goals of justice. Luciano Floridi and Josh Cowles philosophers have added this explainability or explicability variable to this particular hierarchy or framework. These are pretty abstract and I'm not sure how useful they are for kind of line communicators, people in the public sector, you know, issuing press releases and posting on social media. So what I wanted to do is to kind of bring it down, right down to the ground about what I think maybe some applied AI ethics principles for communicators would be, what do we owe our stakeholders, whether it's, you know, public funders, you know, other kinds of constituents out there, expert audiences. First of all, I think, and I say these to my students as well, these are sort of evolved out of my own practice. First of all, someone needs to, on the team needs to own the information integrity of the particular outputs, just like in real life.

[5:32](https://youtu.be/77KbhQFISzg?t=332) Second, transparency and disclosure. And I think, you know, there are new norms around this kind of hybrid product workflow, you know, where AI has played a significant role, but is not the only input. Humans maybe have, you know, more than 50% input, but AI played a role. And I do think we have to get comfortable with disclosure and transparency. And then finally, establishing, I love this phrase, communities of practice within our own organizations, systems of fact-checking, spot checks, collaborative critiques, sanity-checking workflows. I think we need to do that throughout. So I think having someone own it, making sure to disclose it, and then having a team that can help you fact-check and do sanity-checking, I think is really what we owe our stakeholders. Not every output's gonna be perfect. And certainly we've all written press releases and, you know, created social media captions that are less than stellar. And the same is going to follow from workflows that have generative AI in them. But I do think these are some really core things that we owe them.

[6:33](https://youtu.be/77KbhQFISzg?t=393) Okay, let's get right to the AI itself. So it's a busy landscape, and obviously there's a lot of hype coming out of Silicon Valley, and people are trying to sell all kinds of things. The way it's sort of shaken out so far is that there's really sort of three big model players, maybe four, and they all sort of have their kind of frontier or cutting-edge model right now. This would be Gemini 2.5 Pro, the OpenAI, ChatGPT products of 03 and 04 mini, and then finally, Cloud4, which just was released. Perplexity is the other kind of player. There are certainly many others, you know, DeepSeek, Mistral, others from around the world that we can mention. But if you're trying to, you know, sort of jump in and get a sense of, you know, what are some of the main tools? These are three that you should be paying attention to. Another myth I wanna dispel right up front is that when we first started talking about artificial intelligence as a concept, and this comes out of the 50s and it has a long history, there was really this idea of a kind of single kind of broad, omniscient guru who would just kind of tell you everything, and that this sort of like chat interface was the kind of dominant paradigm coming out of the GPT-3 release in 2022.

[8:03](https://youtu.be/77KbhQFISzg?t=483) But what it's turned out, and this is especially true for communications workers, in our mind's eye, we have to sort of get rid of this idea that generative AI is going to be about this sort of omniscient single intelligence, and really thinking about it as a whole bunch of specialized worker bees who can help at various points in the workflow, various kinds of highly trained kind of intervention, various specialists, dynamic interventions within workflow, and I'll talk a little bit more about that. So I wanna sort of dispel that right away. The big trends that I see, and here's sort of my point of view in the comms and media fields right now, is we're sort of bringing AI back home. We're bringing it to the sort of the personalized, institution-specific workflow. Again, away from this idea of just sort of like just a single general intelligence and more a series of like worker bee small intelligences.

[9:03](https://youtu.be/77KbhQFISzg?t=543) So bots that perform certain micro tasks, has become like, I think, a real area of practice for a lot of the AI consultants out there, and I'm certainly happy to recommend names for people who are working in this area of kind of specialists and specialization and creating custom bots. Other things that are going on, AI is being embedded in so many of the applications that we use as communications professionals that would include like Sprout Social, I'm just shouting out a few here, Meltwater, Notion, Jasper, Adobe. AI, generative AI is being embedded in so many of the things that we are, that we were already using. The other thing, and I've alluded to this at the beginning, is that we're starting to see people testing content against simulated audiences, creating a series of personas with data. I mean, one of the great things about many of our audiences they've said a lot of things online, right?

[10:04](https://youtu.be/77KbhQFISzg?t=604) They've made a lot of comments, they've made a lot of, they've issued a lot of press releases. We can actually think about creating synthetic audiences and then testing messages against them. And people, I've talked with some of the leading folks on Madison Avenue who are doing this for big clients, and it's pretty incredible the detail at which they are creating these kinds of simulations in order to try to fine tune their messages, and to red team their messages as well. So we're moving kind of from this general purpose paradigm to this custom paradigm. Of course, what do we want to do as communicators? We want to preserve and establish consistency in tone, style, and voice. And information integrity and trustworthiness are absolutely paramount when we're dealing in the public sector. And that means, it's actually a beautiful thing that we're starting to become more about customization because customization provides and solves for this need for control, stability, and precision.

[11:13](https://youtu.be/77KbhQFISzg?t=673) And there are kind of three customization fundamentals, and I'll walk us through that and I'll show some examples. So we want to codify a knowledge base, like what do we know as a communications professional embedded in an institution? What are the things that we know? And that could be a whole body of knowledge, and we can help the LLMs, the models, sort of learn what we know. And then we can provide them with instructions and we can articulate style guides. So those are going to be three things we're going to look at. Just stepping way back on, sort of looking at the kind of learner journey here. So again, I don't presume to know where everybody is at in terms of on their learning journey, but I've kind of constructed two ways of thinking about it over here on the left-hand side We start with, as a learner, in a kind of descending order. needing to have some orienting knowledge, in order to choose wisely, whether it's the model or whether it's the input-output sequence, we have to have some idea of what's going on, right?

[12:17](https://youtu.be/77KbhQFISzg?t=737) How does the memory work? How is this different from search? What are these different terms like token and feature and function and all these different kind of terms, jargon that are out there. And once we get oriented, I feel like we can start getting serious about this evolving discipline called prompt engineering. And so there's sort of some basics, we'll walk through that. There's some optimization techniques in terms of prompt engineering, but I want you to think, for those of you who are beginners about taking seriously this new evolving discipline we're then starting, as you sort of work down of prompt engineering, this sort of learning ladder or work up, you can look over the right-hand side too, you start to think about how to integrate AI into workflows. Then you start to think about how can I customize GPTs? And GPT is a generic term, although it's often associated with open AIs products.

[13:18](https://youtu.be/77KbhQFISzg?t=798) And then there's just the integrating with tools which is starting to become a reality, and the very frontier, which is to create agents, which are sort of semi-autonomous bots that go out and do things and perform tasks. And then I think at the very highest level for organizations that are really doing heavy duty, uses of agents and integrations and whatnot, providing some kind of systematic oversight structure, which is super complicated, involves responsible AI frameworks. I think really that's kind of almost the highest level, but it's only really for a few types of organizations right now. All right, let's think about customization. Like I said, we're gonna get really, this is sort of complex stuff, and then we're gonna kind of work back to some more basics. But I wanted to, for those of you who are more in that middle tier of the learning ladder, I wanted to give you something right away to kind of chew on. But I think for beginners too, this will be interesting.

[14:20](https://youtu.be/77KbhQFISzg?t=860) So, all of those big three major models, whether it's Gemini and GEMS or GPTs and custom GPTs or cloud projects, they all have this capability of creating something that is custom. So think about putting all of your press releases and all of the news articles about your organization and all the research reports and a ton of the emails and the email and the meeting transcripts into a single engine for which that will become the knowledge base, right? You're sort of creating a brain. And the major models are all moving in this direction. I think the sort of best-in-class, sort of state-of-the-art with communication. Relations and public relations firms are all doing this. So let me just walk through a couple of examples. So if you were in this, if you were in the chat GPT interface, you would ultimate and you clicked create under GPT, you would see that there are a bunch of different fields, instructions, knowledge, and even web search that you can create.

[15:26](https://youtu.be/77KbhQFISzg?t=926) So think about creating this brain that has very tailored instructions, has a knowledge base, and then has the ability to actually go out to the web. And that could mean that you assign it to go out to your own website and read your updated website or your competitor websites or other agencies or institutions that are peers, right? So there's like lots of opportunities to make this a dynamic intelligence as you do that. Okay, so this would be the Google version of this, which is called the GEM. You can see the same kinds of fields. There's instructions, there's knowledge, but let me walk you through actually the Cloud version. So if you were over in the Cloud interface and I'm stealing a few slides from a talk I gave a little earlier for the AI in action series at Northeastern, you would see that there are a series of possible projects that you can create down here. Sometimes they're suggested, they're sort of like off the shelf projects, but you can also just kind of create your own, come over to here to Cloud projects, create one.

[16:31](https://youtu.be/77KbhQFISzg?t=991) What are you working on? So you name it, you give it some goals. You can be more or less specific on this, though I always advise being more specific. And then you might put in your project knowledge. So you can see down here, I kind of have some dummy documents, the Instagram prior posts, our mission statement, our recent news releases, relevant scientific reports. We can put that all in and you can see you have a lot of memory there. It's only 1% of capacity used for these five, four documents that I put in there. There's not that much text, but you get the point. And you can see you add all these reference materials that become your knowledge base. And then there are the project instructions. So this is going to be where you want to very carefully lay out a series, a sequence of things to do for that particular chatbot, that custom GBT, or in this case, we call it a cloud project. And so when you put together knowledge and instructions, you can basically create a repeatable workflow, right?

[17:38](https://youtu.be/77KbhQFISzg?t=1058) Let's generate some new social media captions. Let's generate new media advisories. Let's generate different versions of a email newsletter that we're going to create. So these custom chatbots can really be helpful for internal workflow. They can be fine-tuned in so many different ways, and we can talk about that, but I want to keep moving. But I just want to state, just one more time, that this kind of custom chatbot, custom GBT is a kind of generic term for it, is really becoming the state-of-the-art in communications work across the economy. And I think public sector has particular requirements around sort of data privacy, what sorts of things that can be shared and not shared. You know, worth noting that, you know, an awful lot of public sector use Microsoft and Google products. So, you know, getting enterprise licenses and having data privacy agreements is pretty standard with some of these different kinds of tech companies.

[18:39](https://youtu.be/77KbhQFISzg?t=1119) The generative AI companies and products, you know, are going to have their own kind of push and pull. I mean, at our own institution at Northeastern University, we just inked a deal with Anthropic, one of the first university partners to do so, and, you know, still working out lots of things. But, you know, I think you can get, you know, you can get really, I think, careful, thoughtful people who are thinking about, you know, data privacy involved, and, you know, the general counsel's office will figure that out. But I do think there's a lot of power in these custom GBTs. Okay, the other thing I will say before we sort of move on is just that generative AI is being integrated all over the place. So, you know, this is just Anthropic Cloud just offering to connect with your Google Drive, with your Gmail, with your calendar. So there's a lot of new functionality that's being rolled out, which can provide all kinds of different advantages.

[19:41](https://youtu.be/77KbhQFISzg?t=1181) And we're going to have to think about, you know, your particular use cases and your applications and whether you want to do that. But I think a lot of communicators out there are finding it really, really helpful. All right, let me pull back again to ethics, and then we're going to get really specific on like basic terms and fundamental things. Again, I started at like the more like, whoa, you know, like kind of cutting edge level, and now I'll kind of pull back a bit. So in higher education, we're having the same kinds of conversations that everybody across, you know, the social space, the public sector, the economy writ large are having. And this is like, how much are we just seeding our own judgment, intelligence, and decision-making to these automated tools? And how much, you know, how much are we sort of cheating? How much are we sort of giving away autonomy? And then how much are we actually just increasing the personalization capacity of our workflows, the ability to access large bodies of knowledge and data to help inform our decisions, and ultimately to increase our productivity?

[20:48](https://youtu.be/77KbhQFISzg?t=1248) So, you know, there is a push and pull about, you know, especially at that sort of grand level of narratives, is this going to replace, you know, are we going to be all replaced by AIs? Are we, is it going to be more of a worker-centric revolution where actually everybody's productivity rises? And I don't pretend to know the answer of how it's going to shake out, but it is a very active debate for sure. I mentioned that my graduate students and I were doing some things together and that I thought might be of interest and might give you kind of a flavor for the kinds of things that can be done in the kind of civic and social space, which I think also analogizes pretty well to the public sector. And so what we did in this past semester is we partnered with five local NGOs, one of which was the Boston Housing Authority, you know, in this case, more public sector, you know, straight up public sector and not necessarily NGO. I'll show you some NGOs as well. And so I'll give these slides to everybody so you can kind of look at, you can kind of look at what, you know, we actually did, but I'll just give a brief summary.

[21:58](https://youtu.be/77KbhQFISzg?t=1318) So we're the Boston Housing Authority. We use GPT and CLAWD. Basically Boston Housing Authority was rolling out digital literacy curricula for public housing residents, especially seniors. And so we help them build out different kinds of curricula, printable guides, tiered activities. We built a curriculum support website. We used AI tools to do many of these different things. And I think overall it was a pretty satisfying engagement. Here is a sort of youth focused NGO, Get Lit. And so they wanted, we wanted to help amplify the visibility of a grassroots nonprofit. They have a literacy focused mission. And so for them, we created a multi-platform digital campaign, you know, in some respects We created an AI generated outreach set of prompts using custom GPTs. to assist with personalized communication to influencers, sponsors, and donors.

[23:01](https://youtu.be/77KbhQFISzg?t=1381) And we designed social media strategies to help that organization. Let me give you just a few more examples. The Fenway Memory Project is dedicated to doing local history and sort of lifting up the really rich history of the Fenway area. We know it from Fenway Park and the Red Sox, but the neighborhood itself has an amazing history. And they have a lot of oral history. So a lot of interviews, some video, newspaper archives and whatnot, and needed help kind of organizing that material and producing installations. And we were able to use AI tools to help do that. And we also, you know, use AI enhanced content to design material from the archive that would engage different types of residents, visitors, and students. A couple more. In this case, this is like a tour company, but they're doing like public history. So, you know, particularly in the neighborhood of Roxbury, which has a very rich African-American history.

[24:04](https://youtu.be/77KbhQFISzg?t=1444) And so we use a bunch of different models there to sort of analyze different curricula that were being, that was sort of a new offering from this organization. We actually helped them kind of triangulate between some state standards, which are quite complex and subtle, and sort of what they needed to deliver in classrooms. We also provided a set of hands-on sort of AI prompts to help partner, explore, and apply their AI in their workflow. So we helped them with their workflow. It turns out that a lot of the kind of AI communications consulting that we did ultimately was very much about, was helping sort of upstream in that workflow, not necessarily like end products. And I had mentioned that. And this final organization was doing advocacy around dairy and alternatives to dairy. We helped them identify key legislative supporters, informed advocacy targeting, just a whole bunch of different approaches there.

[25:05](https://youtu.be/77KbhQFISzg?t=1505) And that was a pretty iterative project, but it worked out really well. I'm happy to answer questions about any of these. All right, so we're gonna step way back now to what I think are kind of the two major things that you need to do, especially if you're at the beginning stages of that learning ladder. And one thing that I think is very difficult for people who first engage with an LLM is they first wanna see it as Google search, right, as a search bar. Whereas actually it's something very, very different. And each of the models, each of the LLMs has a different kind of personality, a different way of behaving. We almost have to think of it like almost like therapy. So I think an absolute requirement for using these models effectively is just to spend a lot of time with each of them.

[26:06](https://youtu.be/77KbhQFISzg?t=1566) Maybe you just choose one as your sort of go-to, and then you extend over to a couple others. But what makes them tick is a great question. And you'll see that as you experiment, it's a real journey of discovery. It's a real, there's experimentation, iteration is really required. It is somewhere between a tool and an organism when we're communicating with machines like this. Like literally, it feels like an alien intelligence sometimes, sometimes it feels like a human intelligence sometimes. Sometimes it makes mistakes and hallucinates and does bad things, but often it's like a really wonderful, rich set of outputs that can challenge us as communicators. And so I would just encourage you to spend time. I don't think there's a set number of hours, but I would say, at least 10 or 15 are required to really kind of start to get it and to be able to think like the model. The other thing we need to do, and I had mentioned at the top is the importance of just understanding some of the jargon, right?

[27:12](https://youtu.be/77KbhQFISzg?t=1632) There's a whole new world. We're all kind of learning. There's this idea of prompting, which is really just putting inputs into the model. There are all these ideas around data and concepts around data, tokens, context windows, embeddings, memory, worth kind of unpacking what those are because as you understand each concept, as you understand each piece of jargon, it helps lock into place. Like it gives you a little bit more leverage in terms of using the tool. And then the models, there's lots of really technical, really math-y stuff and really computer science stuff, some of which we don't need to know. But if we wanna get really good as professionals, probably worth following some of the more technical stuff and paying attention to that. And yeah, happy to answer questions about that, but I won't spend too much time. Here is just some basics though. So the kind of the coin of the realm in this world are tokens, which is literally the smallest unit of text processed by the models.

[28:17](https://youtu.be/77KbhQFISzg?t=1697) So not every word is a token. You can see some examples here. The model literally thinks in tokens. It's a probabilistic sort of machine learning environment, which basically associates words with the next possible word. And so it's a probability machine in many ways. The newest models are getting quite complicated in the way that they do this, but just know that tokens are the sort of like basic the comic unit in this entire world. The context window, if you're doing some experimentation and you start just uploading a whole bunch of documents and you start putting in a ton of texts and you start having a long chat transcript with one of the models, you start to find sometimes that as you move along, it starts to get worse, right? It sort of deteriorates. This is a pretty common phenomenon. And often that means that what you've done is you've really stretched the context window, which is the maximum number of tokens a model can consider at once in a single session.

[29:19](https://youtu.be/77KbhQFISzg?t=1759) And the context windows are getting much bigger over time. It's becoming much less of an issue in some cases, but just know that like tokens are the atomic unit. Context window is like the bucket for how many tokens you can be having. You can be sort of, can be in play before things kind of start to go haywire. And just know that there are some limits here. These are vast data systems, but they're not infinite. The other thing is you may wonder, hey, Google personalizes my results. TikTok personalizes my feed. Can't the GPTs do the same thing? Can the LLMs do the same thing? And the answer is yes. They do have some longer-term memory, but for the most part they kind of forget what you said in your last session when you start a new session. There are some things that you can do in some of the settings to give it particular, to help it understand who you are and what you're doing. It will remember some stuff about you from past transcripts. You can of course delete transcripts and you can of course delete anything that it's remembering about you, but if you're seeking some kind of personalization, just know that it's sort of uneven and it does depend on the model. Okay, prompt engineering. I said this is like a new discipline. It's one of the things that my students and I work on over and over again. In fact, every class will have a sandboxing session where we just work on different kinds of prompts. I think you have to think of it as a skill, almost like an athletic skill, which is this ability to do prompt engineering really well.

[31:06](https://youtu.be/77KbhQFISzg?t=1866) I have some more slides in here about that, but I'll kind of rush through that piece so we can get some other stuff. In general, we want to be clear and concise. These elements cannot read our mind. We have to be really specific about what we're looking for, what kinds of formats do we want, do we want a list, do we want a table, do we want seven bullet points, etc. So be really specific. Also think about kind of deconstructing what you want to do. So don't maybe put all of your requests and detail in a single prompt. You can kind of move it through step by step, and I think increasingly that's a best practice. So OpenAI has said for a while, include as much detail as you can. Ask the model to adopt a persona. We'll get to that in a minute. Provide examples. We'll get to that in a minute. But the slides here contain a bunch of the sort of best practices as seen by, in this case, this is Claude.

[32:12](https://youtu.be/77KbhQFISzg?t=1932) So just to take, you know, just the top line from this little table here. Worse, how do I add numbers in Excel? Better. How do I add up a row of dollar amounts in Excel? I want to do this automatically for a whole sheet of rows with all the totals ending up on the right in a column called total. You can see how specific this command is, and that's much easier for the model. So this is really weird. This is just a fact about transformer models, which is a kind of the technical term for all of these new generative LLMs, which is that they have so much data in them. They have so many parameters. Think of it as a giant spreadsheet with like a trillion rows and a trillion columns that actually helping them narrow what they're doing by pointing their attention to something is really helpful. And one of the best ways to point them to a particular kind of piece of their overall mental universe is to ask them to play a role. So you are a science communicator. You are a legal expert. You are a marketing professional.

[33:14](https://youtu.be/77KbhQFISzg?t=1994) And then you can provide more detail. This actually helps the models pay attention to the particular kinds of things that you want to get them to help you with. It sounds a little odd, right? Because you may have heard like the big danger with LLMs is hallucination, right? They're going to make something up. So in this weird way, you're actually asking them to make up a kind of role or a persona in order to do something. It sounds like you're setting them up for failure and more hallucination. But in fact, what it does is it narrows their scope. It narrows their attention. It actually helps them process a little bit better. Some more detail here if you're looking to think about context windows, memory, token limits. I'll leave these in here for reading later. I did mention decomposition as a kind of or prompt chaining as some people are calling it as a technique. This is this idea that we really want to break down complexity and ask them. For example, we upload a spreadsheet.

[34:25](https://youtu.be/77KbhQFISzg?t=2065) We say, how many rows? How many columns? What's in it? I'm uploading a spreadsheet. I'm uploading a CSV file. I'm really starting to walk the model into what you're doing and then stating, I'm looking to do change over time for cities of 100,000 people or more since 1960. I'll get into data analytics and database, but the point is a broader point about decomposition, which is we don't want to put everything right away into the input window. We want to put in concise but fairly detailed inputs, let the model work, and then keep walking it through steps. We're doing really complicated stuff. There's these kind of jargony ideas about zero-shot prompting and few-shot prompting. Zero-shot prompting is basically like not giving it examples. Few-shot prompting means you give it some examples. There are various schools of thought about which to do and you're going to have to experiment depending on your use case and your application. Let me look at a couple use cases. Image generation is obviously quite controversial. I'm not going to spend a lot of time on it here. The degree to which this might replace photographers, deceive people through missing disinformation, these are all big live issues. But there are lots of ways in which AI-generated images can be super helpful to us. You can see throughout this presentation, there's a ton of clip art, essentially, that I generated through Dali or Midjourney or others. Let's just think about that for a second. One key thing to know is that the models are trained on a massive amount of image and text. There are biases and gaps in that training, what we call pre-training data.

[36:25](https://youtu.be/77KbhQFISzg?t=2185) The AI companies and their engineers and computer scientists try to train out some of the bias through reinforcement learning with human feedback, but it's not always perfect. There are lots of questions about the ethical sourcing of images and this pre-training, I'm not going to get into that. We want to be really specific sometimes about what kinds of images we want and what... Excuse me. I'm going to come right back here. I'm just going to stop sharing for a quick second. Let me just bring this back up. Okay, here we go. There are different kinds of... It's quite complicated what goes into this generative process. It's a little different than the text generation. We can get a little bit into diffusion models if you'd like. But one of the things, just basic tips for prompting if you're trying to create images, like be specific, provide a clear subject, minimalist, mention the color palette, use simpler language at times. You can prompt these models to do things and iterate, but you have to be really specific. You might indicate the camera angle, top down. You might indicate, if you're a little more technical, the aspect ratio, specify focal points, et cetera, et cetera. There's lots of things that you can do that are essentially photojournalistic terms and techniques that if you can communicate them directly to the model, they're going to be much better than if you're just like, draw me a picture of environmental pollution or something like that. There's a lot of style transfer going on, so you can use different types of aesthetic words to get the models to adopt certain kinds of styles. There are limitations with current

[38:30](https://youtu.be/77KbhQFISzg?t=2310) artists and photographers and other kinds of where there are a lot of IP concerns, the models are going to be more hesitant to do that. But with lots of classical aesthetic styles, you can have the model adopt those and everything from cartoonish to photorealistic. What are the pitfalls in an area like this? Well, we call it the uncanny valley, which is you end up with too many fingers or too many eyes or something weird. A lot of the images continue to have a lot of distortions, though you'll see some of the photo, some of the image generators now, including some of the video generation engines are getting really good at avoiding this uncanny valley. But there is a big potential for bias or offensive content, issues of race, gender, global culture. You'll see a lot of different perils as you ask for different kinds of, for example, job roles. What does a nurse look like? What does a librarian look like? What does a president look like? You'll see immediately all of the built-in assumptions that that pre-training data carries, and that's something we have to really keep in mind from an ethical standpoint. I mentioned data viz and data analytics, and I just want to say this is a growing relationship. None of the models are perfect yet.

[39:55](https://youtu.be/77KbhQFISzg?t=2395) If you load in a spreadsheet or a CSV file and you ask it to do something really sophisticated, correlate rows and columns and run regression, and you ask it to do something that's really complicated, you're likely to get garbage out. You have to really do it patiently through steps. So just beware. But that said, I actually think many of the models are getting more useful, better, more sophisticated. They're making fewer errors. If you can take that time, and here's a workflow, which is that you upload a CSV file, for example, to ChatGPT. You explain what you're interested in. I'll give you some prompts in the next slide. You ask it to pick a format. Maybe it's a stacked horizontal bar chart like this one, or you might ask it to give you a bunch of possibilities. The models are terrific at generating different possibilities. And then I still recommend, especially if you're trying to do higher-end, higher-polished data viz, using another app like a Data Wrapper or Flourish to finish it off, but letting the models help you explore the data. It's essentially like Excel without having to know formulas or any of the particular tools in their toolbar. It's really helpful, and you can just use plain language to do that. Then I think another fantastic use case is to finish the cycle of data analysis and data visualization by uploading the more polished product. Just upload the image out of Flourish or Data Wrapper or Excel or wherever you're producing the viz, and then ask GPT or Claude or Gemini to make suggestions for improvements and critiques.

[41:53](https://youtu.be/77KbhQFISzg?t=2513) So here would be some sample prompts. You upload a CSV file. What are two or three key insights on this data? Select the chart type. Anyway, I'll leave you with those. I know we're running a little short on time, so I'm just going to speed through some things. For those of you who are a little higher up that learning ladder and you're starting to think about custom chatbots beyond just creating these really specific internal projects, but you're actually starting to think about should we create something that can talk with customers or with the public, which is a much more typically more expensive but higher risk proposition, the custom chatbot workflow I kind of lay out here, and I sort of lay out what it means, the different kinds of models that might help you do it. I mean, you're going to need an engineer or some kind of AI consultant to come in to help you do it, but I sort of walk you through what that would look like if you're looking to create sort of external facing, non-internal custom chatbots. And so there are some slides in here with that. But in the name of time, I'm going to just finish up with some tool tips and then some podcasts that I might recommend. So overall, I would say you want to use ensembles of models. You want to be doing Gemini and Perplexity and using GPT for this. Once you get up that learning ladder a ways, like you should be using multiple models. They can cross reference each other. They can be used adversarially. You can do something in one and then ask the other if it thinks that the output was reasonable. They're very good, actually, and they have different strengths and weaknesses. And we'll talk about that a little bit more if you'd like.

[43:51](https://youtu.be/77KbhQFISzg?t=2631) Many apps and specialized sites are being created. I'll shout out a few in a second that can be helpful. And there's so much rapid change right now. That kind of best in class, best practices, all of that, it almost is changing week to week. And so you need to have some kind of news and information flow where you're keeping up as a professional. In my talks with my network, which is a lot of former students and a lot of people who are out there in the public sector, in the social sector, on Madison Avenue, wherever, a few apps came to the surface. And so I wanted to sort of highlight those. One, if you're looking to do really high end presentations, probably worth checking out this gamma.app, which is really cool. I was kind of experimenting with it. And it's a way of making really high end, pretty fantastic presentations. If you're looking for a solution to meetings and meeting notes, I know there's lots of different solutions right now out there. But a few folks I know were which is one of my favorites. And Asar produces a ton of content. There are really interesting guests all the time. He's keeping up with all the latest models. Every weekend, he just tells you sort of what's the latest in the space, but he also talks with really thoughtful people

[45:52](https://youtu.be/77KbhQFISzg?t=2752) who are doing communications and enterprise and business work of all different kinds. And then AI Explored, a podcast that I came upon fairly recently. And the guests here, if you go back even a couple months, just guest after guest, he's providing a ton of insight, solo entrepreneurs, people in big shops, all kinds of different people. And there are just so many tips for communications professionals. So with that, I'm going to stop and take questions. I hope this was of interest to you. There's something like 75 slides in here. So you know, I didn't get to everything, but I hope it was useful.

**Agueda Quiroga**

[46:31](https://youtu.be/77KbhQFISzg?t=2791) Thank you so much.

* * *

Original page: https://innovate-us.org/generative-ai-for-public-sector-communicators-tools-ethics-and-best-practices · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org
