The playground of possibilities: A guide to generative AI for public media

By |
More

Chad Davis, using DALL-E 3

Remember when you tried the internet for the first time? That first touch of a sprawling universe of information and connection? Generative Artificial Intelligence (gAI) is a paradigm shift. gAI refers to a category of artificial intelligence systems designed to generate new content, often in the form of images, text or other forms of media (including audio), that may be (and in most instances are intended to be) indistinguishable from content created by humans.

While this technology has captured the public’s imagination, it has also raised many questions for public media. It’s not just about fancy chatbots or photorealistic landscapes — it’s about pushing the boundaries of creativity and exploration, reshaping how we interact with technology and the world around us.

Any shift like that comes with a real risk of being on the wrong side of the AI divide, and the first step to landing on the right side of that divide is becoming AI-literate.

gAI has a place in both our personal and professional lives. You’ve probably heard of most of the major tools by now, such as ChatGPT, Google’s Bard, DALL-E, Midjourney and Llama. Some of these tools can help us with day-to-day work, and some can help with the process of creation. Public media can use these tools to improve: automating tasks, personalizing customer experience, and streamlining decision-making capabilities. All of that helps focus our time on mission efforts. But it is important to stress that they are just tools, and they need humans at the helm. 

We’ve written this Guide as and for practitioners of public media, and it will be a living document with future articles here in Current highlighting updates to the Guide. To kick things off, we’ll dig into chatbots like Bard, ChatGPT and Claude, that can hold conversations, write different kinds of creative content and even translate languages. Then we’ll dive into the world of generative art tools like Midjourney, DALL-E 3, and Imagen 2, where you’ll conjure landscapes from mere text. And going forward, this guide will evolve with the gAI landscape, adding new tools, exploring new categories of gAI and showcasing the possibilities that lie ahead.

Many public media organizations have already or are now determining their content and operational policies related to gAI. These have been presented to address ethical considerations internally and with some understanding that much output needs review. This is particularly notable for bilingual content in cases of organizations seeking to engage Spanish-speaking audiences. Needless to say, successful exploration of gAI is accomplished through cross-functional, deliberate conversations about the goals, outcomes and guardrails related to AI. We will continue to educate about tools and techniques but intend to foster strategic discussion as well.

While the lines between gAI categories are already blurring, for now these tools can be divided into a few categories, including Text, Data and Images. In all these categories, we’ve been selective in what to mention here. There are countless gAI tools on the market now and more coming online each week. If you hear about a tool, thoroughly investigate before subscribing or consider a tryout discount from third-party markets like StackSocial or Mashable. Also, if you’ve heard of a tool that isn’t mentioned, it means it’s not one we can necessarily recommend at this time. That may be due to not having tested it or because it’s simply not up to public media’s standards.  If you have a gAI tool you think deserves to be in this Guide, tell us in the comments below.

But a word or two of caution: Remember to be careful with uploading proprietary data. Be mindful of the considerations around accuracy, fairness and inclusiveness. Also, read the terms on the free tools — for example, Midjourney and OpenAI put liability on the user. If you are using gAI for anything more than experimentation, where you might take the output and publish it, be sure to follow the editorial guidelines set by your organization. And finally, try to ensure that the requests you make of these tools (aka “prompts”) are constructed in a way that doesn’t produce an obvious copyright infringement. If you are an organization with access to in-house legal counsel, you may want to ask them for risk evaluation.

So, whether you’re a seasoned gAI user, gAI-curious or simply someone who wants to peek into tools that may influence tomorrow’s public media, this guide is your permission to embrace the future. Let’s experiment, let’s learn, let’s push the boundaries together. It’s time to play. 

A word on terms

This guide uses several terms related to AI. We present them here to help you better understand systems, their strengths and areas for improvement.

  • Generative AI: AI systems that create content — like images, audio, video and text — that is unique. 
  • Datasets, also called Training Data: The data that AI models are trained on, determining their behavior. Models can inherit biases from problematic training data.  
  • Neural Networks: The computing systems modeled after the human brain that underlie most major AI breakthroughs recently. 
  • Large Language Models: Massive neural networks trained on huge text datasets, allowing natural language generation, comprehension and translation.
  • Parameters: The adjustable settings within a neural network tuned during training to optimize the model. More parameters mean higher capability.
  • Diffusion Model: A type of generative model that creates data by iteratively refining output according to patterns in the training data.
  • Prompt Engineering: Strategically designing text prompts provided to generative AI to steer outputs in a targeted direction.
  • Tokens: The unit of measure for interactions with chatbots. As of this writing, 100 tokens is about 75 words.

Text

Although generative art tools hit the scene about five months before chatbots, models that address text-based prompts captured the public’s imagination. Here are the ones you need to know now:  

  • ChatGPT Plus (from OpenAI):
    • TLDR: If you’re going to try only one, this should be it.
    • Description: The premium version of the original gAI chatbot offers a feature-rich portfolio of tools for gAI enthusiasts and professionals who are willing to pay for the best of the best. It has the most up-to-date data and offers a suite of tools that can be implemented as add-ons.
    • Availability: Desktop and mobile browser, and mobile app
    • Cost: $20/month for individuals; $30/month/user (or $300/year/user) for ChatGPT Team
    • Languages: Claims support for 50 languages, including Spanish, though output used for content and public-facing activities may miss cultural nuance
    • Pros:
      • Access to a more advanced model (GPT-4) under the hood than the free version (GPT-3.5)
      • Prioritized server access over the free version, for when the whole world seems to be using the bot
      • Access to generative art tool DALL-E 3 right in the interface (see more about DALL-E below)
      • Allows the use of images uploaded in prompts
      • Can accept input of up to 128,000 tokens (96,000 words)
      • Access to the recently launched nascent but growing marketplace of “GPTs”
        • GPTs are preset roles (aka “agents”) that serve as a shortcut to certain types of expertise or information. Think of them as the apps of the gAI world. There are already thousands, and they have the potential to make the tool ever more user-friendly. 
      • Available for free without multimodal components 
      • ChatGPT Team subscribers will also not have their data used in training (meaning corporate conversations with ChatGPT stay private)
    • Cons:
      • The $20 price tag may be steep for the gAI-curious
      • The $30/month price tag for Team users may be steep for CFOs
      • Image recognition isn’t as good as you’d hope for the price
      • Like other chatbots, its predictive text generation can result in citation of sources that don’t exist, or it may present false information as fact
      • Currently the defendant in multiple intellectual property lawsuits, including a major one from the New York Times
  • ChatGPT (free version from OpenAI):
    • TLDR: If you’re only going to try one, this should be it, but we recommend springing for the Plus version.
    • Description: ChatGPT comes in a free version and a paid “Plus” version ($20/month). 
    • Availability: Desktop and mobile browser, and mobile app
    • Cost: Free
    • Languages: Same number of languages as Plus
    • Pros:
      • Arguably, the best *free* chatbot on the market
      • Available both via the ChatGPT website/mobile app as well as via Microsoft’s Bing search engine
    • Cons:
      • As of this writing, ChatGPT offers responses with data that was last updated in Sept. 2021, so some material could be dated
      • Like other chatbots, its predictive text generation can result in citation of sources that don’t exist, or it may present false information as fact
      • Underlying model, GPT-3.5, lacks the training of Plus’ GPT-4
      • Can be slow or have access limited if OpenAI server usage is high 
      • Can only accept 16,000 tokens — 3,000 words — of input, meaning  your interactions with the chatbot will be limited 
  • Bard (from Google):
    • TLDR: Bard was an also-ran in the world of chatbots until late last year when Google launched its Gemini Pro multimodal model behind the scenes. Now it’s worth a try, especially if you are a Google power user.
    • Description: Bard can handle a wide range of tasks, from writing different kinds of creative content (poems, code, scripts, musical pieces, email, letters, etc.) to translating languages and answering your questions in an informative way.
    • Availability: Desktop and mobile browser, and mobile app
    • Cost: Free
    • Languages: Claims support for 40 languages, including Spanish, and adherence to AI principles
    • Pros:
      • No cost to experiment
      • You may have already interacted with Bard, or at least seen its summaries in your Google search results
      • Powered by the Gemini Pro model, which is highly competitive with OpenAI’s GPT-4
      • Allows the use of images in prompts and is better at identifying objects than ChatGPT
      • Arguably as powerful as ChatGPT 3.5
    • Cons:
      • Like other chatbots, its predictive text generation can result in citation of sources that don’t exist, or or it may present false information as fact
      • Because it’s Google, and it’s free, your data is the real product being monetized
  • Claude 2 (free version from Anthropic):
    • TLDR: Your friendly neighborhood chatbot. As a “constitutional AI” — or a model that its creators have aligned with ethical guidelines intended to make it more transparent — this tool is probably closest to public media in terms of underlying values. However, its feature sets may lag behind the competition for some users.
    • Description: Developed by Anthropic, a company started by ex-OpenAI researchers who felt their company was taking the wrong approach, Claude 2 comes in a free and enterprise version. We recommend trying the free version at this stage.
    • Availability: Desktop and mobile browser, and mobile app
    • Cost: Free for 50 messages/day (1500/month); $20/month for 5x more usage
    • Languages: Claims support for over a dozen languages, including Spanish
    • Pros:
      • Has a constitution under the hood that stands as a governing filter between what it “thinks” and what it “says,” plus Anthropic has been transparent with how their constitution was created
      • Users report it feels like a more human interaction; a kinder, gentler chatbot
      • As of November 2023, the team announced its latest version of Claude can handle up to 150,000-word inputs
    • Cons:
      • Doesn’t have the same multimodal integration (yet) that ChatGPT Plus and Bard do. So, you can’t upload images as part of your prompts.
      • Claude 2’s focus on safety comes at the expense of versatility. It may not be as adept at creative writing, humor or generating unconventional content compared to other gAI tools.
  • Microsoft Copilot:
    • TLDR: WIth OpenAI’s models under the hood, Copilot is just a different flavor of ChatGPT. If you are an Office365 power user (and especially tied to Teams) this will give you more bang for your buck than ChatGPT. 
    • Description: Copilot uses the same OpenAI training models under the hood. However, ChatGPT Plus is getting the latest product updates first. If that isn’t important to you, then you might be able to experiment with what is essentially ChatGPT through Microsoft’s Edge browser, the Bing search engine or, for a fee, the Office365 suite of tools. 
    • Availability: As an add-on feature to Windows 11 and 10, in the Office365 suite of tools, and as a feature in Bing search.
    • Cost: $20/month for individuals with Word, Excel, PowerPoint and OneNote access (no Teams); $30/month for full Office365 integration (with Teams); some features free through Bing and Windows OS
    • Languages: Claims support for about 24 languages through “bots,” with varied training levels
    • Pros:
      • Integration in the operating system and business tools used by much of the professional world
      • Potentially fewer privacy issues than using a tool like Bard
    • Cons:
      • Cost
      • Text training for local Copilot comes from your documents and may provide limited results

Data/Code

  • Github Copilot (from Microsoft):
    • TLDR: GitHub Copilot was the first mainstream AI coding assistant trained on billions of lines of open-source code from the largest repositories on GitHub. Just describe what you want or provide your code, and Copilot will analyze and offer suggestions.
    • Description: ChatGPT-style chatbot that you can use to ask questions within your integrated development environment. It is optimized to offer code suggestions for Python, JavaScript, TypeScript, Ruby, Go, C# and C++. But it can generate code for any language with public GitHub repositories. Security and privacy: Uses Transport Layer Security for secure data transmission, encrypts data in transit and at rest. Individual user data can be retained, but you can opt out or request deletion. Business user data is not retained.
    • Availability: Integrates with Visual Studio Code, Visual Studio, JetBrains IDEs, Vim, Neovim and Azure Data Studio.
    • Cost: Individual plan at $10/month or $100/year, Business plan at $19/user/month.
      • GitHub Copilot Individual Features: Telemetry, block any suggestions that match public code, integrate with your IDE or code editor, and multiline function suggestions.
      • GitHub Copilot Business Features: All features of the individual account, plus organization-wide policy management, audit logs and HTTP proxy support with custom certificates. 
    • Pros:
      • By far the most accurate and least annoying. GitHub Copilot is the most used among developers, so data is sourced from a larger pool of practitioners.
      • Better for niche languages that are in the GitHub repository ecosystem
    • Cons:
      • If you do not opt out, the code you provide could become the property of Microsoft
      • Copilot can suggest code that is unoptimized, inefficient or even insecure
  • Amazon CodeWhisperer (from AWS)
    • TLDR: CodeWhisperer is a free AI chatbot from Amazon that can accelerate software development by offering real-time, tailored code suggestions.
    • Description: By analyzing natural language code comments or partially completed code blocks, CodeWhisperer can intelligently infer and suggest code snippets to help you complete your task, including entire functions or logical blocks that match your coding style. Much like Copilot, Amazon CodeWhisperer was trained on billions of lines of code, but with additional code from Amazon. Supports Python, Java, JavaScript, TypeScript, C#, Rust, Go, Ruby, Scala, Kotlin, PHP, C, C++, Shell Scripting and SQL. Security and privacy: Uses TLS for secure data transmission, encrypts data in transit and at rest. Individual user data can be retained, but you can opt out. Professional user data is not retained.
    • Availability: Compatible with JetBrains IDEs, Visual Studio Code, AWS Cloud9, AWS Lambda console, JupyterLab and Amazon SageMaker Studio.
    • Cost: Individual plan is free, Professional plan at $19/user/month.
    • Pros:
      • Totally free!
      • If you work in the AWS cloud, it streamlines suggestions for APIs when building AWS applications, offers AWS code suggestions, performs security scans and has Amazon Q integration
    • Cons: 
      • Need to use short comments that map to small and discrete tasks 
      • Need to use intuitive naming conventions for variables and function names
  • Bard (with Gemini Pro)
    • TLDR: Google’s Bard is an AI that cannot only generate code, but help with debugging and code explanation.
    • Description: Google Vertex AI has a variety of generative AI foundation models that are accessible, including Bard and Gemini API. They provide advanced reasoning, multiturn chat, code generation and multimodal prompts.
    • Availability: Bard can handle more than 20 programming languages, including C++, Go, Java, Javascript, Python and Typescript.
    • Cost: Vertex AI is billed according to the compute resources and services used. Monthly cost could be around $1,000. 
    • Pros:
      • Bard has a nice UI and quick code generation, and it provides decent details about the code generated if prompted for them 
      • Bard is also proficient in changing a chunk of code into another code language
    • Cons:
      • Expensive
      • The details for the code aren’t quite as detailed as ChatGPT Plus
  • ChatGPT (free version from OpenAI)
    • TLDR: ChatGPT can generate code, but it also explains what the code is doing, which can be especially helpful if you’re new to the language.
    • Description: It is trained on Stackoverflow. Best to be used for a beginner. If you are just learning to code, this can be a teacher. It’s useful for getting an idea of implementation or an outline.
    • Availability: It can write code snippets in popular programming languages such as JavaScript, Python, C#, PHP and Java.
    • Cost: Free
    • Pros:
      • Fine for many general-purpose use cases 
      • It’s quick and easy to generate code using ChatGPT 
      • It can get you familiar with codes, modules and resources better than most readmes
      • Can be a good code-debugging companion. It also provides detailed explanations of what it did to fix the issue, making it a fantastic educational tool for programmers.
    • Cons:
      • Code isn’t always accurate, usually doesn’t work, about 90% correct
      • Can’t handle a complicated multistep problem, tends to struggle with answer accuracy and starts to lose context after multiple messages
      • If you copy and paste code from Chat GPT, you can be exposing your employer to copyright lawsuits
      • Don’t use it for cryptography or concurrency
  • ChatGPT Plus (from OpenAI)
    • TLDR: All the same features as the free version but includes access to the GPT Store, which makes it possible to share GPTs publicly and launch Assistants. 
    • Description: The new Assistants API provides new capabilities such as Code Interpreter and Retrieval as well as function-calling to handle a lot of the heavy lifting that you previously had to do yourself, enabling you to build high-quality AI apps. An assistant is a purpose-built AI that has specific instructions, leverages extra knowledge and can call models and tools to perform tasks. The Assistants Code Interpreter can write and run Python code in a sandboxed execution environment, generate graphs and charts, and process files with diverse data and formatting. It allows your assistants to run code iteratively to solve challenging code and math problems, and more.
    • Availability: It has the ability to write code snippets in popular programming languages such as JavaScript, Python, C#, PHP and Java.
    • Cost:  $20/month for individuals; $30/month/user (or $300/year/user) for ChatGPT Team
    • Pros:
      • ChatGPT Plus maintains great answer accuracy through lengthy conversations 
      • ChatGPT Plus also offers the added benefit of file upload and web-browsing capabilities 
      • ChatGPT Plus now offers a wide variety of user- and OpenAI-created custom GPTs for many different use cases
      • OpenAI is also introducing Copyright Shield, a new feature that will defend customers and pay the costs incurred if they face legal claims around copyright infringement. This applies to generally available features of ChatGPT Enterprise and the developer platform.
    • Cons:
      • Cost

Ernesto Aguilar is Executive Director of Radio Programming and Content Diversity, Equity and Inclusion (DEI) Initiatives at KQED in San Francisco. At KQED, he oversees radio broadcast content and DEI initiatives in the organization’s Content division. In his spare time, he writes OIGO, a newsletter on public media and Latino audiences.

Mikey Centrella leads the PBS Innovation Team as a Director of Product focused on rapidly building and testing ideas with emerging technology, helping PBS adapt to new services and skills for approaching AI, machine learning, VR, AR, voice and interactive video. 

Chad Davis is Chief Innovation Officer at Nebraska Public Media, where he leads Nebraska Public Media Labs, an R&D division focused on emerging media. Chad is also the inaugural Chair of the Public Media Innovators peer learning community at NETA and writes a weekly newsletter about the intersection of public and emerging media. To join the PLC, subscribe to the newsletter or talk to Chad about gAI and emerging media, you can email him at [email protected].

Leave a Reply

Your email address will not be published. Required fields are marked *