Easy on those ChatGPT prompts, buddy—marketers eager to incorporate AI into their business strategies need first to learn the key terms that articulate the space. It is a vocabulary that endlessly bends back on itself, requiring the knowledge of prior concept upon prior concept like a stack of nesting dolls. How else is one supposed to understand LLMs if they don’t already have neural networks down? Worse yet, how embarrassing it might be to discuss Perplexity when really you meant to talk about Anthropic …

For these reasons, Ad Age has compiled a glossary of all the words and concepts you need to know to get this space straight. And we’ll continue to update it as more terms become necessary to understand. 

AGI: Artificial general intelligence (AGI) is considered the holy grail of AI development. While precise definitions differ, it generally refers to AI that is as or more advanced than human cognition. AI companies such as OpenAI and Google are racing to build the first AGI systems because they could herald a new era of intelligent computing. But while advocates including OpenAI CEO Sam Altman claim AGI is just a matter of more computing power and is thus around the corner, others say AGI requires an entirely new paradigm of thinking and is therefore still a ways from being developed.

AI: Artificial intelligence, or AI, is the general set of technologies that use advanced computing to perform tasks. Neural networks, machine learning, large language models (LLMs)—these are all developments that belong under the AI umbrella. While generative AI is a subset of AI, when we say AI we don’t necessarily mean tools that generate content. The history of AI dates back to the 20th century, but advancements beginning in the 1960s led to the capabilities that are available today.

AI agent: A piece of software that uses AI to make rational decisions to complete specific tasks. Agents can interact with their environments and execute autonomously, meaning they don’t need continual prompting like a chatbot. These programs are expected to play a large role in automating a wide range of tasks for businesses.

AI washing: The act of falsely marketing or exaggerating one’s use of AI software, whether intentionally or unintentionally. In early 2024 the U.S. Securities and Exchange Commission began cracking down on firms for making such specious claims, a response to what many view as overhype within the AI space.  

Anthropic: The startup behind generative AI model Claude, and a primary competitor of OpenAI. Anthropic has attempted to separate itself by promoting its commitment to responsible AI development. The company has a partnership with WPP and is backed by Amazon and Google.

Related: A generative AI guide for brands

Chatbots: The programs running inside messaging apps and on websites that help consumers perform simple tasks. Chatbots were popularized by apps like Facebook Messenger, and they operate like apps within the messaging services. The recent boom in generative AI bots like ChatGPT is enabling these services to be far more advanced and human-sounding.

ChatGPT: A generative AI bot created by OpenAI, which uses GPT technology (see below) to produce an output based on an input. The platform is trained on vast quantities of text taken from websites, articles, books and other sources. It launched publicly at the end of 2022 and quickly exploded in popularity due to its vast array of capabilities, from screenwriting to joke-telling to disseminating knowledge on basically any topic under the sun. 

Related: ChatGPT writes Super Bowl ads

Context window: The upper limit of text that an LLM can process for any single input. LLMs differ in their context windows; Anthropic’s Claude 3 model, for example, can accept up to 200,000 tokens, or about 150,000 words. In comparison, OpenAI”s GPT-4’s context window is 128,000 tokens, or roughly 96,000 words.

DALL-E: Another generative AI bot created by OpenAI, though its function is to generate images. Like ChatGPT, a user queries the bot with an input (e.g. “Ronald McDonald swimming in a river of fire”) and out comes a corresponding image. The first version of DALL-E launched in January 2021; two more versions have since launched.

Deep Learning: A more advanced branch of machine learning, where a computer teaches itself with only minimal amounts of programming. With deep learning, marketers can make the most use of data and apply it to make predictions about consumer behavior.

Deepfake: Deceptive video and/or audio imitations of a person created without their knowledge or consent. The rise of AI has led to the increased proliferation and accuracy of this type of fraudulent media, which tends to impact celebrities and brands.

Dynamic Pricing: One of the common tasks that deep learning-based programs can perform is setting prices based on consumer data. Dynamic pricing means that each consumer is presented with a price based on their particular circumstances, depending on the time of day, their financial situation and other factors. Dynamic pricing can come into play in rates for plane tickets, for instance. AI can help craft personalized prices that are most likely to ensure a sale at the most efficient rate.

ElevenLabs: A text-to-speech AI platform, often used alongside other generative AI systems to incorporate voice. 

Generative AI: Generative artificial intelligence is a form of AI that, as its name suggests, generates content. Depending on the specific platform’s capabilities, this content can be images, videos, text, audio, as well as other mediums. The platforms that have recently become popular generate their content from a user’s input, such as “Produce an image of X” or “What is the answer to Y?”

Gemini: Google’s primary AI system, competing with OpenAI’s GPT. Gemini was formerly known as Bard but was rebranded in early 2024. The platform is used in most of Google’s consumer-facing products, powering its search overhaul dubbed Search Generative Experience. It’s also used in many enterprise products, such as ads manager Performance Max.

GPT: Generative Pre-trained Transformer, a generative AI model that is trained on reams of data to produce an output. The technology was created by OpenAI.

Hallucination: An instance of an AI system providing incorrect or biased information, such as when Google’s Gemini inaccurately depicted historic images. The term “hallucination” is used because the system presents the information as if it were correct, which makes spotting such an error even more difficult. Hallucinations arise from biases in the system’s training data or other malfunctions.

Image recognition: AI looks for patterns in images. Machines can analyze many more images than humans, and with machine learning, they can identify what’s in the images and reveal patterns that people would never detect. Brands can use image recognition technology, for example, to find every photo online in which their logos appear. That could help brands locate their most loyal customers and tease out other actionable marketing insights. “Computer vision” is a term associated with image recognition, and refers to computer programs that analyze and categorize digital images.

Related: How Snap, Google and Meta are embracing generative AI  

Large language models (LLMs): A class of AI technology whose primary function is to accomplish language-related tasks, such as text generation. The platforms are trained on reams of data and built on intricate neural networks that enable them to make predictions and generate corresponding media, such as text or images. OpenAI’s GPT model is an LLM.

LLaMA: Large Language Model Meta AI, a generative AI model developed by Meta. It is an LLM, like OpenAI’s GPT and powers much of Meta's AI applications.

LoRA: Low-rank adaptation (LoRA) is a way of fine-tuning an AI model for more custom media creation. Today’s most popular LLMs tend to be general-purpose, so a brand or AI company looking for more bespoke assets may want to train an open-source model using LoRA.

Machine learning: When a machine teaches itself with minimal programming needed. Machine learning can be helpful in direct marketing and email marketing, in particular, by ingesting vast sets of consumer data and using it to determine things such as the best times to send emails. Proponents say it can also identify the clients or customers that would be most receptive to given messages. And machine learning is used in ad targeting to help deliver messages to those audiences.

Neural Networks: Artificial intelligence programs modeled after the human brain. They incorporate deep learning and natural language processing to perform functions like recognizing handwriting and faces in photos.

OpenAI: An artificial intelligence company that has created some of the most popular generative AI tools, including ChatGPT and DALL-E. OpenAI started as a non-profit organization in 2015, founded by a team including Elon Musk, but expanded to include a for-profit arm in 2019. While technically independent, OpenAI is heavily backed by Microsoft; the tech giant invested another $10 billion in OpenAI at the start of 2023.

Open-source AI: AI software in which the underlying source code is widely accessible. Several AI providers have made their models open-source, including Meta and Google. However, the degree of this transparency is case-dependent; for example, Meta’s open-source AI does not reveal the code or data used to train the model.

Perplexity: An AI-powered search engine built to provide in-depth and conversational responses to queries. Perplexity is backed by notable investors including Nvidia and Jeff Bezos, but has faced controversy for allegedly plagiarizing material from news publications including Forbes, Wired and The New York Times. 

RAG: Retrieval-augmented generation (RAG) is a process by which LLMs can outsource for information that is beyond its training data. For example, an LLM can use RAG to surface real-time information from a news site or social media platform, or reference an organization's internal knowledge base, all without needing any extra training.

Runway: A text-to-video AI startup known for its short but realistic productions. Lionsgate has partnered with Runway to create a video model based entirely on its IP catalog. Runway’s 3rd and most powerful model, Gen-3, is in testing as of October 2024.

Small language models (SLMs): SLMs are compact AI models built for specific language tasks. Compared to LLMs, which have a much wider breadth of functionality, SLMs require less computational power but can perform more effectively for purposes that may be too nuanced for LLMs.

Synthetic data: Data created by AI that is used to emulate real data. Its purpose is to help companies conduct research and test products by skirting barriers they would experience if using real information, such as privacy restrictions and limited ability to generate new data. Synthetic data could help brands better understand their customers and how to engage with them.