**TLDR:** AI tools like ChatGPT are powerful assistants for PR professionals that can automate routine tasks and enhance productivity, but they require proper prompting skills and human oversight, especially for sensitive communications. # Introduction to AI for PR & Comms 📣 As AI progresses, it becomes more useful. It becomes more relevant. Not just for PR, but for for every industry. The way we work is changing. The key to maximizing AIs potential lies in understanding how it works, applying it judiciously, and maintaining a balance between automation and human touch. So today, for the good of the team, I'm going to address a few common questions about AI: 1. What is AI? What does it do? We hear about it constantly, but it often seems like magic. I'm going to try to demystify that a little bit. 2. What is ChatGPT? Many of us use it every day, but do you know the difference between the available models? How should we choose which model to use? 3. How can PR professionals, especially writers involved in content production, use AI? When *should* we use it? When should we *not*? 4. How can we improve our experience with AI? 5. Lastly, we'll address some AI hype. ## Defining Terms First, let's define some terms. Does anyone know what... 1. **Artificial intelligence (AI)** is a field concerned with building systems that simulate intelligent behavior resembling human actions but typically limited to specific tasks. 2. **Machine Learning (ML)** is an area within AI that is concerned with teaching mathematical models how to do stuff by giving them a ton of examples. And I mean a ton. This process can take a lot of computational resources, which is why they you may have seen months ago OpenAI made a $10B deal with Microsoft to build a super computer. 3. **Deep Neural Networks (DNN)** are a type of ML model. These are the most powerful and practical ML models you'll encounter regularly. They learn to create new data that are statistically indistinguishable from the training data 4. **Generative AI (Gen AI)** refers to a class of DNNs that are designed to create new, plausible instances of data such as text, images, or audio based on learned patterns, much like humans. For example, they can write essays, solve math problems, and even win art competitions. It's kind of a hype word. I'm mention it here, though, because you'll likely encounter it in the wild. ## Demystifying Types of AI DNNs encompass a variety of architectures for different tasks. DNNs that understand and generate text are called **large language models (LLMs)**. Ex. BERT, OpenAI's GPT-4T (Generative Pre-trained Transformer), and Meta's LLaMA 2 Most LLMs are built with the **Transformer Architecture**. It's like reading a book with friends, each highlighting key points, combining everyone's highlights, and then using this combined understanding to summarize, translate, or explain the book: 1. **Convert words into numbers**: Imagine you are turning a book into a list of codes, where each code represents a different word (embedding the words). 2. **Remember the order of the words**: As you read, you also note the position of each word in the book because their order matters (positional encoding). 3. **Pay attention to the important words**: While reading, you highlight key words in each sentence that seem most relevant to understanding the context (self-attention). 4. **Use multiple perspectives to decide importance**: You then ask several friends to read the same sentences and highlight important words from their unique perspectives (multi-head attention). 5. **Mix all this information to understand the context**: You combine everyone's highlights to get a comprehensive understanding of which words are crucial in each sentence (feed-forward networks). 6. **Repeat the process several times for better understanding**: You read through the book multiple times, each time refining your understanding based on previous readings (transformer layers, including residual connections and layer normalization). 7. **Use this understanding to perform various tasks**: Finally, you use your thorough comprehension of the book to summarize it, translate it into another language, or answer questions about its content (encoder and decoder functions). DNNs that understand and generate audio are called **ASR (Automatic Speech Recognition) models**. Ex. OpenAI's whisper can understand audio and convert it to text. DNNs that understand and generate images are called **CNNs (Convolutional Neural Networks)**. Ex. OpenAI's DALL-E 2 (or Midjourney's Model Version 5) can generate images from text. DNNs that can take all these inputs (text, images, and audio), understand them, and generate various outputs (not just the source type), are called **multimodal models**. OpenAI kind of made up this word. Ex: OpenAI's SOTA (state of the art) GPT-4o ("o" for "omni"). OpenAI is just one provider of DNNs. There are other closed source competitors as well as many open-source alternatives. OpenAI was first to the consumer market and has been mostly dominating since they released ChatGPT to the public. ## Demystifying ChatGPT and Available Models OpenAI's ChatGPT provides a user interface (UI) for their closed-source LLMs and, most recently, it's multimodal GPT-4o. Here are the latest models compared: | **Feature/Model** | **GPT-4** | **GPT-4T (Turbo)** | **GPT-4o (omni)** | | --------------------------- | ---------------------------- | ------------------------------------------------------ | -------------------------------------------------- | | **Release Date** | March 14, 2023 | November 2023 | May 13, 2024 | | **Context Window** | 8,192 tokens | 128,000 tokens | 128,000 tokens | | **Knowledge Cutoff** | September 2021 | April 2023 | October 2023 | | **Input Modalities** | Text, limited image handling | Text, images (enhanced) | Text, images, audio (full multimodal capabilities) | | **Vision Capabilities** | Basic | Enhanced, includes image generation via DALL-E 3 | Advanced vision and audio capabilities | | **Multimodal Capabilities** | Limited | Enhanced image and text processing | Full integration of text, image and audio | | **Cost** | Standard | Three times cheaper for input tokens compared to GPT-4 | 50% cheaper than GPT-4 Turbo | ### Comparing GPT-4/T to GPT-4o - GPT-4o is a new model trained end-to-end across text, vision, and audio, meaning the same neural network processes all inputs and outputs. The training set for GPT-4 was likely used for GPT-4o's text portion, along with new vision and audio datasets, making it a multimodal model. While this is impressive, it's not particularly relevant for our use case as writers. - GPT-4o was trained with more data, which might enhance the model's reasoning capabilities. OpenAI's self reported benchmarks support this, but practical performance can vary. For now, I'll consider this a win (ref. https://arxiv.org/abs/2302.14045), though I personally prefer GPT-4T's answers over GPT-4o's. - GPT-4T has a cutoff date six months later than GPT-4o, which is a drawback for the newer model. GPT-4o is, however, one month ahead of the base GPT-4, which is not much of an advantage. - GPT-4T has a much larger token limit (context window) than the base GPT-4, and GPT-4o matches this. So, this aspect is a tie. - GPT-4o is 50% cheaper (API) than GPT-4T, which is a win. However, since we pay a subscription fee, this is less relevant. In summary, despite the impressive multimodal capabilities of GPT-4o, for text-based PR tasks, the benefits over GPT-4T are marginal. ## How PR Pros Can Use ChatGPT Using Gen AI is by no means required to write well. But it can save a lot of time writing boilerplate text, generating outlines, or writing drafts (where there are ample examples or well established processes). You can create press releases, outline articles, emails, or explain and simplify any text. It doesn't matter what kind of text you need to create or interpret; If you have your task outlined properly and include the right information, the LLM will perform. However, it's important to understand when to use AI and when you need human insight, especially in sensitive communications. For example, generating a response to a sensitive question from a journalist via email and sending it without much thought is not appropriate. ## How To Improve Your Experience with LLMs like ChatGPT ### #1 Improve How You Interact with the System The first way to improve your experience with LLMs like ChatGPT is to improve how you *interact* with the system. Although Gen AI is extremely useful, it requires humans to direct it on what to do. Often, Gen AIs are like new company interns. They are very capable, but they need clear instructions to do well. Being able to properly instruct Gen AIs is a very powerful skill. Telling Gen AI what to do is called "prompting." Unfortunately, there's a lot that goes into prompting, so it's something you'll have to look into on your own and learn with practice. I'll share a few resources to help you get started. ### #2 Improve the System The second way to improve your experience with LLMs like ChatGPT is to improve the system itself. There are a few ways to do that, but the simplest is using Custom GPTs, which OpenAI released on Nov 6, 2023. This feature allows anyone to create custom versions of ChatGPT that combine instructions, extra knowledge, and any combination of skills: 1. **Custom Instructions:** let you set some chat preferences and ChatGPT will keep them in mind for all future conversations. The model will consider the instructions every time it responds. This limits manually copying and pasting your prompts into ChatGPT. 2. **Extra Knowledge:** GPTs can now fetch up-to-date or relevant documents you upload. This new architecture helps solve three fundamental problems of LLMs: - **LLMs lack up-to-date knowledge**, so RAG provides recent information. Information may be current news, an industry best-practices manual that was just published, etc. - **LLMs lack out-of-domain knowledge;** this could be knowledge on a unique topic that LLMs do not understand or internal knowledge such as company documents. - **LLMs are prone to hallucination.** Researchers have shown RAG reduces the likelihood of hallucination even on data that the model was trained on. 3. **Developer Actions:** allow GPTs to integrate external data or interact with the real-world. **Note:** Default model for Custom GPTs is GPT-4, for now. #### Retrieval Augmented Generation (RAG) ![[Pasted image 20230926092023.png]] RAG is considered an "equalizer" of models; RAG with any model is likely to offer more significant advantages than the greatest model alone: > RAG with various models (including open source, smaller) outperform SOTA models like GPT-4 (Reference: Pinecone RAG Study). #### Custom GPT: Communications Pro Gregory made a Custom GPT for PR & Comms called "**Communications Pro**", which you can see under my account if you're logged in, or go to this link: https://chatgpt.com/g/g-xkLmXlaFL-communications-pro This GPT has access to a collection of books that cover strategies for marketing, public relations, and communication: 1. Influence by Robert B. Cialdini 2. Made to Stick by Chip Heath and Dan Heath 3. Strategic Planning for Public Relations by Ronald D. Smith 4. Public Relations Strategies and Tactics by Dennis L. Wilcox and Glen T. Cameron 5. This ls Marketing by Seth Godin 6. Positioning by Al Ries and Jack Trout 7. Marketing Management by Philip Kotler and Kevin Lane Keller 8. The New Rules of Marketing & PR by David Meerman Scott 9. Crossing the Chasm by Geoffrey A. Moore 10. Reputation Management by John Doorley and Helio Fred Garcia 11. The Practice of Public Relations by Fraser P. Seitel 12. Contagious: How to Build Word of Mouth in the Digital Age by Jonah Berger 13. That GPT would be good to interact with when you're interested in strategy. I've made some others with custom instructions which you can check out, too: - Prompt Engineer - Try asking "What are your best prompting tips?" 2. Outline Generator 3. Copywriter ## AI Hype OpenAI is very good at what they do, but they tend to be dramatic. Oftentimes Sam Altman is caught embellishing on X, and he's been accused of outright lying to stakeholders. Unfortunately, this kind of attitude spreads ends up permeating consumer's collective consciousness. But the reality often falls short of the hype, leading to unrealistic expectations and misunderstandings about AI's actual value. For example, the idea that AI will replace all junior marketing roles is an oversimplification. AI will not suddenly steal thousands of jobs. However, as AI systems automate repetitive tasks, junior employees will have more opportunities to focus on tasks that require human ingenuity: creativity, critical thinking, ethics, and communication skills, etc.—all things that machines can never replace. ​​Furthermore, eliminating juniors is self-defeating, because juniors are the future of the organization. Therefore, instead of replacing workers, AI will act as a catalyst for upskilling (learning new skills to improve performance in current job) and reskilling (learning new skills to do a new job or adapt to changes in current one), but these require a culture of continuous learning. Put simply, AI is raising the bar. Software has been automating repetitive tasks for a long time, so, in terms of qualifying a threat, this is nothing new; to a software engineer, AI is another addition to his toolbox. ## Recap For the good of the team, we: 1. We began by demystifying AI 2. We compared different versions of ChatGPT 3. We reviewed AI's applications in PR & Comms 4. We talked about how to improve your ChatGPT experience 5. And, finally, we addressed the AI Hype AI, specifically Generative AI like ChatGPT, is a powerful tool for PR professionals. But understanding *how* and *when* to use it is crucial. You should also understand it's limitations. As AI tools become more ubiquitous, PR professionals must continuously develop their skills to remain competitive. This shift underscores the broader trend of technological advancement driving the need for continuous learning and adaptability in all industries. ## Q&A