AI Tools Glossary
35 terms. Plain English definitions for the AI and machine learning concepts that matter most to content creators, marketers, and writers.
Core AI
Large Language Model (LLM)
An AI trained on massive text datasets to generate and understand language.
A Large Language Model is an AI system trained on billions of text examples — books, websites, code, research papers — to learn patterns of language. It can generate coherent text, answer questions, summarize documents, and write code. GPT-4, Claude, and Gemini are all LLMs. The 'large' refers to the number of parameters, which can range from millions to hundreds of billions.
Example
When you type a prompt into ChatGPT and it writes an essay, it's an LLM generating text by predicting the most likely next word, over and over.
Tokens
The chunks of text an LLM processes. Roughly 1 token ≈ ¾ of a word.
AI models don't process text character by character or word by word — they process 'tokens,' which are chunks of text. One token is roughly 3-4 characters or ¾ of a word in English. 'Unbelievable' might be 3 tokens; 'cat' is 1. Token counts matter because AI APIs charge per token and have context window limits. Most AI writing tools display a token counter or impose word limits that map to token limits.
Example
OpenAI charges per 1,000 tokens. A 1,000-word article is roughly 1,333 tokens. GPT-4's context window is 128,000 tokens — enough for a full novel.
Hallucination
When an AI confidently states something false or made-up as if it were fact.
AI hallucination happens when an LLM generates plausible-sounding but factually incorrect information — fake citations, wrong statistics, names of people who don't exist, events that didn't happen. It's not intentional deception; it's a fundamental limitation of how LLMs work. They generate probable sequences of tokens, not verified facts. Hallucination rates vary significantly by model and task type. Research tasks and factual Q&A are highest risk.
Example
Asking an AI to summarize a scientific paper and receiving a list of citations where half the paper titles, authors, or DOIs are fabricated — even though each one sounds credible.
Context Window
The maximum amount of text an AI model can process in a single interaction.
The context window is the limit on how much text — including your prompt, conversation history, and documents you've uploaded — an AI model can 'see' at once. Once the limit is exceeded, the model starts forgetting earlier parts of the conversation. Measured in tokens. A 128K token context window is roughly 100,000 words. Larger context windows are better for analyzing long documents or maintaining coherence across long projects.
Example
If you paste a 200-page PDF into an AI with a 32K context window, it will only be able to read about 25,000 words of it — the rest is invisible to it.
Prompt
The text instruction you give an AI to tell it what to do.
A prompt is the input you give an AI model — a question, instruction, or example that tells it what you want. Prompt quality directly affects output quality. A vague prompt gets vague output; a specific prompt with context, format instructions, and examples gets much better results. This is why 'prompt engineering' has become a recognized skill.
Example
'Write a blog post about marketing' is a weak prompt. 'Write a 1,500-word SEO blog post targeting the keyword "B2B email marketing best practices" for a SaaS audience. Use H2 headers, a practical tone, and include 3 real-world examples.' is a strong prompt.
Prompt Engineering
The skill of crafting inputs to AI systems to get better, more consistent outputs.
Prompt engineering is the practice of designing, testing, and refining prompts to maximize the quality and consistency of AI outputs. It includes techniques like role prompting (telling the AI to act as an expert), few-shot prompting (providing examples), chain-of-thought prompting (asking the AI to reason step by step), and negative prompting (telling the AI what NOT to do). As AI tools proliferate, prompt engineering has become a core content production skill.
System Prompt
Hidden instructions that set the AI's behavior before a user sends their first message.
A system prompt is a pre-configured instruction that the AI tool's creators write to shape the model's behavior. Users don't usually see it. When you use Jasper as a 'copywriter' or Notion AI for 'meeting notes,' those modes are powered by system prompts that tell the underlying model how to behave. Most AI writing tools are primarily differentiated by their system prompts, not by which LLM they use underneath.
AI Writing
AI Writing Tool
Software using LLMs to help create, edit, or optimize written content.
AI writing tools are applications built on top of LLMs like GPT-4 or Claude that add a user interface, templates, and workflow features for content creation. They range from simple text expanders to full editorial platforms. The underlying model often matters less than the template library, workflow features, and output quality for specific content types. WriteTested has reviewed 148 of them.
Brand Voice
A set of guidelines an AI tool uses to match a company's writing style.
Brand voice in AI writing tools is a feature that lets you define your company's tone, style, vocabulary preferences, and writing rules — then apply them consistently across AI-generated content. Better implementations analyze your existing content to extract voice patterns automatically. Brand voice features significantly reduce editing time and are one of the key differentiators between AI writing tools.
Long-form Content
Articles, reports, and documents typically 1,500+ words where depth is expected.
Long-form content refers to articles, blog posts, reports, whitepapers, and similar documents that typically run 1,500 to 5,000+ words and cover a topic in depth. Generating coherent long-form content is harder for AI than short-form — it requires maintaining consistent argument structure, avoiding repetition, and keeping a logical flow across multiple sections. Not all AI writing tools handle long-form equally well.
Technical
RAG (Retrieval-Augmented Generation)
A technique that gives AI access to external documents to reduce hallucination.
Retrieval-Augmented Generation (RAG) is a technique where an AI model is connected to an external database or document set. When you ask a question, the system first retrieves relevant passages from those documents, then feeds them to the model as context. This significantly reduces hallucination for domain-specific knowledge because the model is generating from real sources, not from training data memory. Many AI research tools and customer support chatbots use RAG.
Example
Instead of asking GPT-4 to summarize your company's product docs (where it might hallucinate details), a RAG system retrieves the actual relevant sections from your docs and passes them to the model before generating the summary.
Fine-tuning
Training an existing AI model on new data to specialize its behavior.
Fine-tuning is the process of taking a pre-trained model and continuing its training on a smaller, specialized dataset to improve performance on a specific task or to match a particular style. A legal firm might fine-tune an LLM on thousands of legal documents to produce better legal writing. It's more effective than prompting for consistent style changes, but requires significant technical resources and data. Most AI writing tools for non-technical users don't offer true fine-tuning.
Parameters
The numerical weights inside an AI model that store its learned knowledge.
Parameters are the numerical values inside a neural network that are adjusted during training to capture patterns in data. When you hear about a '70 billion parameter model,' it means the model has 70 billion adjustable numbers that encode its understanding of language. More parameters generally mean more capability and better quality output, but also more computational cost. Parameters are what's actually downloaded when you run an AI model locally.
API (Application Programming Interface)
A way for software applications to communicate with each other programmatically.
An API is a defined interface that allows one software system to send requests and receive responses from another. AI APIs (like OpenAI's or Anthropic's) let developers build applications on top of AI models without running the model themselves. Most AI writing tools are built on AI APIs. When a tool offers its own API, it lets developers integrate that tool's features into their own workflows or products.
Grounding
Connecting AI output to verifiable, external sources of truth.
Grounding refers to techniques that anchor AI-generated content to real, verifiable sources — reducing hallucination by constraining what the model can say. Ground truth sources include web search results, uploaded documents, databases, or knowledge graphs. A grounded response includes citations; an ungrounded response is generated purely from the model's training data, which may be outdated or incomplete.
Inference
The process of running a trained AI model to generate a response.
Inference is what happens when you submit a prompt and the model generates output. Training is done once (or periodically); inference happens every single time a user asks the model a question. The speed and cost of inference is what determines how fast an AI tool responds and how much it costs to run. Inference speed is one of the dimensions we test in every WriteTested review.
Embeddings
Numerical representations of text that capture semantic meaning, enabling similarity search.
Embeddings are high-dimensional numerical vectors that represent text in a way that captures semantic meaning. Similar sentences produce similar embeddings. They're what powers semantic search (finding documents based on meaning, not keywords), RAG systems, and recommendation engines. When an AI tool can find 'related articles' or 'similar products' based on meaning rather than exact keyword match, embeddings are the underlying technology.
Semantic Search
Search that understands meaning rather than matching exact keywords.
Semantic search finds relevant results based on the meaning and intent behind a query, not just keyword overlap. If you search for 'how to reduce customer churn,' a semantic search returns articles about 'improving retention' and 'reducing cancellations' even if they don't use the word 'churn.' Powered by embeddings. AI SEO tools are increasingly optimizing for semantic search rather than just keyword density.
AI Image
Diffusion Model
The type of AI architecture behind most modern image generators (Midjourney, DALL-E, Stable Diffusion).
Diffusion models are the AI architecture behind most modern image generators. They work by learning to gradually remove noise from random pixel patterns until an image matching the text prompt emerges. Trained on massive image-text datasets, they can generate photorealistic images, artistic illustrations, or anything in between. Stable Diffusion (open source), DALL-E (OpenAI), and Midjourney all use diffusion-based approaches.
Negative Prompting
Telling an AI image generator what to exclude from the generated image.
Negative prompting is a technique used in image generation where you specify elements you DON'T want in the output. For example, adding 'blurry, watermark, extra fingers, low quality' to your negative prompt tells the model to avoid those common failure modes. It's particularly important in Stable Diffusion and similar tools where the model often defaults to low-quality patterns without explicit guidance away from them.
Example
Prompt: 'Portrait of a woman, natural light, professional photo' / Negative prompt: 'cartoonish, blurry, distorted, extra limbs, watermark'
Upscaling
Using AI to increase the resolution of an image while adding detail.
AI upscaling uses neural networks to increase an image's resolution — making it larger without the blur or pixelation of traditional scaling. Unlike simple interpolation, AI upscalers can hallucinate realistic detail at higher resolutions. Tools like Topaz AI, Magnific, and built-in upscalers in Midjourney can take a 512×512 image and produce a usable 4K version. Essential for AI-generated images used in print or large-format display.
Text-to-Image
Generating images from written descriptions.
Text-to-image is the core function of AI image generators: you write a description (prompt) and the model produces an image matching that description. The quality, style consistency, and prompt adherence vary significantly by model and tool. Commercial tools like Midjourney, Adobe Firefly, and DALL-E 3 have optimized for quality and safety filtering. Open-source models like Stable Diffusion offer more control but require more technical skill.
AI Video
Text-to-Video
Generating video clips from written descriptions or scripts.
Text-to-video AI generates short video clips from text prompts. The quality has improved dramatically — tools like Sora, Runway Gen-3, and Luma AI can produce realistic 5-30 second clips with coherent motion. Limitations include maximum clip duration (usually under 60 seconds), inconsistent character identity across clips, and physics artifacts. Most AI video tools used in production combine text-to-video with script-to-voiceover features.
AI Audio
AI Voiceover / Text-to-Speech
Converting written text into synthetic spoken audio.
Text-to-speech (TTS) converts written text into spoken audio using AI-generated voices. Modern TTS (ElevenLabs, Murf, Play.ht) produces human-like voices indistinguishable from recordings in most contexts — with natural pacing, emotion, and pronunciation. Quality is measured by naturalness, voice variety, SSML support (for controlling pacing and emphasis), and latency. Essential for video production, podcasts, and accessibility.
Voice Cloning
Creating an AI version of a specific person's voice from audio samples.
Voice cloning uses short audio recordings (typically 30 seconds to 3 minutes) of a real person's voice to create an AI model that can generate new speech in that person's voice. Used legitimately for content creators to produce voiceovers at scale, or for accessibility. Requires consent of the person whose voice is cloned — unauthorized voice cloning is illegal in many jurisdictions and violates most platform ToS.
SEO & Content
AI SEO Tool
Software using AI to research, optimize, or generate SEO-focused content.
AI SEO tools combine traditional keyword research and rank tracking with AI to automate or enhance content creation for search. Functions include: keyword gap analysis, content briefs, on-page optimization suggestions, AI content generation with SEO structure baked in, and competitive content analysis. The category includes tools like Surfer, Frase, Clearscope, and MarketMuse — each with different philosophies on how much to automate vs. assist.
SERP
Search Engine Results Page — the page Google shows after a query.
SERP stands for Search Engine Results Page. When you search for something on Google, the page of results is the SERP. SEO tools analyze SERPs to understand what kind of content ranks for specific queries: How long is it? What headings does it use? Is there a featured snippet? What questions do users also ask? Most AI SEO tools include SERP analysis to inform content briefs.
Keyword Research
Finding the words and phrases your target audience searches for.
Keyword research is the process of identifying the specific search queries your target audience uses, along with their search volume (how many people search per month), difficulty (how competitive they are to rank for), and intent (are they looking to buy, learn, or compare?). AI SEO tools have automated much of this research, clustering related keywords and identifying gaps in existing content.
AI Writing Quality
AI Detection
Tools that attempt to identify whether text was written by AI.
AI detection tools attempt to identify AI-generated text by looking for statistical patterns — unusually uniform sentence length, low 'perplexity' (how surprising each word choice is), and lack of the irregularities that characterize human writing. Detection tools are notoriously inaccurate, with high false-positive rates (flagging human writing as AI) and low true-positive rates on paraphrased AI content. Most academic integrity experts treat AI detection results as unreliable.
Perplexity
A measure of how unpredictable or surprising text is to an AI model.
Perplexity is a statistical measure of how surprising a sequence of text is to a language model. High perplexity means the model found the text unpredictable — suggesting it may have been written by a human who makes unusual word choices. Low perplexity suggests the text follows typical patterns — which AI text tends to do. AI detection tools use perplexity as one signal. AI humanization tools work by rewriting to increase perplexity.
AI Humanization
Rewriting AI-generated text to reduce AI detection signals.
AI humanization tools take AI-generated text and rewrite it to add variation, inject natural-seeming imperfections, and increase perplexity to reduce the likelihood of AI detection. Quality varies enormously — some tools just add typos or restructure sentences slightly, while others do deeper semantic rewrites. Use cases include students, content marketers, and anyone who needs AI-assisted content to pass plagiarism checkers.
Platform Concepts
Multimodal AI
AI that can process and generate multiple types of data: text, image, audio, video.
Multimodal AI models can process and generate more than one type of data. GPT-4V can analyze images and answer questions about them. Claude can read PDFs. Gemini Ultra can understand video. Multimodality is becoming the standard for frontier models. For content creators, this means AI tools can take a screenshot and rewrite the copy in it, analyze a chart and write commentary, or generate an image from a design brief.
Model
The underlying AI system a tool is built on (e.g., GPT-4, Claude 3, Gemini).
In the context of AI tools, 'model' refers to the underlying AI system powering the tool. The same model can power many different products — Jasper, Copy.ai, and dozens of other tools all used GPT-4 at various points. Models differ in output quality, context window size, speed, cost, and specializations. Some tools use multiple models or let users choose. Most AI writing tools don't disclose which model they use, but it's discoverable.
Temperature
A setting that controls how creative/random vs. predictable/focused an AI's outputs are.
Temperature is a parameter that controls the randomness of AI output. At temperature 0, the model always picks the most likely next token — output is deterministic and repetitive. At temperature 1+, the model samples from a probability distribution — output is more varied and creative but potentially less coherent. Most consumer AI tools hide temperature as an internal setting, but developer-facing tools expose it. For creative writing, higher temperature; for factual responses, lower.
Few-shot Prompting
Showing an AI 2-5 examples of what you want before asking it to do the task.
Few-shot prompting means including 2-5 examples of the desired input-output format in your prompt before asking the model to complete the task. It's one of the most reliable ways to improve output quality without fine-tuning. Showing the AI 3 examples of the email format you want is more effective than describing the format in words. 'Zero-shot' means no examples; 'one-shot' means one example.
Want the tools, not just the terms?
We've reviewed >148 AI tools so you don't have to.