What is AI? A plain-language guide to artificial intelligence

A plain-language definition of AI

Most definitions of AI either sound like science fiction or like a textbook. Neither is helpful when you just want to understand what is going on under the hood of the chatbot you started using last week. So here is the version we'd give a smart friend over coffee.

Artificial intelligence is software that performs tasks normally associated with human intelligence by learning from data instead of being explicitly programmed. That second half is what makes the modern wave different from older software. A traditional program is a set of instructions a person wrote: if X then Y. A modern AI is a program that started as a blank statistical structure, was shown billions of examples, and slowly tuned itself until its outputs matched what those examples suggested they should be.

That is why AI feels different from regular software. You're not running a recipe. You're querying a compressed summary of an enormous amount of human-produced content, and asking it to extend the pattern.

What AI isn't

It helps to clear away a few myths up front. They show up in headlines and product copy and they make the rest of the topic harder to understand than it needs to be.

• AI is not conscious. It has no awareness, no feelings, no inner life, and no agenda. The fact that an LLM can write a sentence like "I am worried about that" does not mean it is worried. It means humans wrote that sentence often enough in the training data that it is statistically appropriate to produce in context.
• AI does not "know" things the way you know things. When you ask an LLM a factual question, it is not looking up the answer in a database. It is producing the response that the patterns in its training data suggest is most likely. Most of the time those patterns reflect reality. Sometimes they don't, and you get a confident, fluent, completely wrong answer — that's a hallucination.
• AI is not magic and it isn't a single thing. "AI" today refers to many different model types — language models, image models, video models, speech models, recommendation models — each trained for a specific job. They share underlying math but they are not interchangeable.
• AI is not "the algorithm" in the conspiracy-theory sense. Real AI systems are not deciding what's good for you. They are optimizing a measurable objective set by their developers. Who chose that objective and how it was measured matters far more than any spooky agency on the model's part.
• AI is not new. The term was coined in 1956. The current boom is roughly the seventh wave of AI hype since then. What's new is scale: vastly more data, vastly more compute, and a particular neural-network architecture (the transformer) that scales unusually well.

The main types of AI

People split AI up in two useful ways. The first split is by capability:

• Narrow AI — does one specific task well: detecting tumors in X-rays, transcribing speech, ranking search results, generating images from prompts. Every AI system in real-world use today is narrow AI, including the very capable ones.
• Artificial General Intelligence (AGI) — a hypothetical AI that could perform any intellectual task a human can, across any domain, without retraining. AGI does not exist. Whether and when it might exist is one of the most contested questions in the field.
• Artificial Superintelligence (ASI) — an even more hypothetical AI that exceeds human ability across all domains. This is a thought experiment, not a product category.

The second split is by what the model is trying to do:

• Discriminative AI — classifies, predicts, or scores. "Is this email spam?" "What's the next word likely to be?" "Is there a stop sign in this image?" Discriminative models powered the AI products of 2010–2020 (search ranking, fraud detection, recommendation feeds).
• Generative AI — produces new content. Text, images, audio, video, code. Generative models are what made AI go mainstream in 2022–2024 and are probably why you're reading this page. Wurt.app's Picwurt and Vidwurt are generative AI; so are ChatGPT, Claude, Gemini, Midjourney, Sora, and Runway.

The main subfields of AI

The capability/function split above is one slice. The other slice — the one academic textbooks open with — is by which kind of input the system is built to handle. These subfields existed long before the current boom, each with its own conferences, jargon, and history. Modern deep learning has eaten most of them, but the divisions still tell you what a system is for.

Computer vision

Computer vision is AI that processes images and video. Detecting objects in a photo, segmenting a tumor in an MRI scan, reading a license plate, deciding whether a face matches an ID, generating a frame from a text prompt — all computer vision. The field used to depend on hand-designed feature extractors (edges, corners, gradients). After 2012, when AlexNet won ImageNet by a wide margin, the entire field migrated to deep convolutional networks, and more recently to transformer-based vision models. Picwurt, Vidwurt, every face-unlock on a phone, and most of what self-driving cars do is computer vision.

Natural language processing (NLP)

NLP is AI that processes human language — reading it, writing it, translating it, summarizing it, answering questions about it. The field has gone through three eras. First came hand-written grammars and rules. Then statistical NLP starting in the 1990s, where models counted word co-occurrences and learned shallow patterns. Then, after the 2017 transformer paper, large language models swallowed the entire field. Today "NLP" and "LLM" are nearly synonymous in industry, though academic NLP still covers parsing, tagging, and the unglamorous infrastructure that sits underneath.

Speech and audio

Speech recognition turns audio into text; speech synthesis turns text into audio; speaker identification figures out who is talking. Modern systems handle all three with deep neural networks trained on huge audio corpora. Whisper, Siri, Alexa, the captioning on a phone call, the voice clone in a deepfake — all the same family. Music generation and sound-effect generation are newer applications of the same toolkit.

Robotics

Robotics is the part of AI that has to deal with physical reality. A robot needs to perceive its environment (computer vision and other sensors), plan a sequence of actions, and execute those actions through motors and joints, all while the world keeps changing. This is genuinely harder than software-only AI, which is why robots that can reliably pick up arbitrary objects in a warehouse remain a much smaller industry than chatbots. Self-driving cars are the most ambitious deployed robotics project so far, and they remain a work in progress after twenty years of effort.

Expert systems and symbolic AI

For decades — roughly the 1970s through the early 2000s — most "AI" in industry meant expert systems: hand-coded rules and logic that captured the decision-making of a human specialist. Tax software, early medical-diagnosis tools, and a lot of business logic still works this way. Modern deep learning has displaced expert systems for any task with enough training data, but the symbolic tradition is not dead. It's where ideas like knowledge graphs and constraint solvers live, and there's active research on combining symbolic reasoning with neural networks.

Recommendation and ranking

The least glamorous subfield is also the one that touches the most users. Every social-media feed, video-streaming queue, e-commerce homepage, and search-results page is ordered by a recommendation or ranking model. These are usually some flavor of deep learning trained on billions of user interactions. They are AI; they are simply not the part that gets press releases.

How AI actually works (without the math)

Almost every modern AI system follows the same three-stage life cycle: data, training, inference. If you understand what each stage is doing, the rest of the field clicks into place.

1. Data

Modern AI is the data it was trained on. An LLM was shown a large fraction of the public internet plus books, code, and curated text. An image model was shown hundreds of millions or billions of image-and-caption pairs. A speech model was shown audio with transcripts. Whatever signal the data contains — including its biases, gaps, and stylistic quirks — ends up baked into the model.

This is why "the data" is genuinely the most important part of a model. A model trained on bad data is a bad model, no matter how clever the architecture.

2. Training

Training is where the model becomes a model. You take a neural network — billions of internal numbers ("weights"), all initialized at random — and you start showing it examples. For each example you check the output, compute how wrong it was, and nudge every weight a tiny amount in the direction that would have made the output less wrong. Repeat trillions of times.

That iterative nudging is called gradient descent. It is the single most important idea in modern AI. The whole field is, mathematically, applied gradient descent on huge neural networks running on enormous fleets of GPUs.

For chatbots, training has a second stage called RLHF (Reinforcement Learning from Human Feedback). After the base model has learned to predict text, humans rank pairs of responses ("which of these two is better?") and the model is fine-tuned to produce more of the kind people prefer. RLHF is what turned raw LLMs into helpful, polite, instruction-following assistants.

3. Inference

Inference is just running the trained model. You give it an input, the input flows through all those tuned weights, and an output comes out. When you type a prompt into ChatGPT or hit "generate" on Wurt.app, you are doing inference.

Inference is much cheaper than training, but it isn't free. Every prompt costs the operator some compute on a GPU somewhere. That's why most generative-AI products charge by usage, by tokens, or by credits.

The "model" — what is actually inside the box

When people say "the model," they usually mean a large neural network — a stack of layers, each layer a grid of those weights, with non-linear functions in between. The two architectures you'll see named most often are:

• Transformers — introduced in the 2017 paper "Attention Is All You Need". Transformers power virtually every notable LLM (GPT, Claude, Gemini, Llama) and an increasing share of image and video models. The key idea is the attention mechanism, which lets each part of the input look at every other part to figure out which parts matter for a given output.
• Diffusion models — the dominant architecture for image generation and a fast-growing one for video. A diffusion model learns to gradually turn random noise into a coherent image that matches a prompt, by reversing the process of corrupting clean images with noise during training. Stable Diffusion, Midjourney, DALL-E 3, and the image models inside Wurt.app's Picwurt are all diffusion models.

How models are trained: the four main paradigms

"Training" is one word for several different procedures, depending on what kind of signal the model gets to learn from. The four labels below come up constantly in AI writing, and most production systems combine more than one.

Supervised learning

The model is shown labeled examples — images tagged with what's in them, sentences tagged with their language, emails tagged spam-or-not — and learns to predict the label for new examples. This is the workhorse of practical machine learning. It is conceptually simple and very effective, but it depends on having someone label the data, which is often the most expensive part of the project.

Unsupervised learning

The model gets data with no labels and has to find structure on its own. Clustering customers by behavior, detecting anomalies in a stream of credit-card transactions, compressing a dataset into a smaller representation — all unsupervised. Useful when you have lots of raw data and no patience to label it.

Self-supervised learning

This is the trick that made modern LLMs and most modern image models possible. Instead of asking humans to label the data, you take the raw data and hide part of it from the model, then ask the model to predict the hidden part. For text: cover up the next word and ask the model what it should be. For images: cover a patch of the image and ask the model to fill it in. The "labels" are generated automatically from the data itself, which means you can train on essentially the entire internet without paying anyone to annotate it. Almost every notable model since 2018 was pre-trained this way.

Reinforcement learning

The model is treated as an agent that takes actions in some environment, and gets a numerical reward for actions that turn out well. Over many trials it learns a policy that maximizes long-run reward. Reinforcement learning is how AlphaGo learned to beat human Go champions and how robots learn motor skills in simulation. RLHF, the technique that turned raw LLMs into helpful chat assistants, is reinforcement learning where the reward signal comes from human preference rankings instead of a fixed game score.

Semi-supervised and other hybrids

Real systems mix paradigms. A typical modern LLM is pre-trained with self-supervision on web text, then fine-tuned with supervised learning on instruction-following data, then polished with RLHF. An image model might be pre-trained self-supervised on uncaptioned images, then fine-tuned supervised on caption pairs. In practice, "what kind of training did this model get" usually has three or four answers, not one.

A short, honest history of AI

You don't need this to use AI, but it helps you cut through the marketing.

• 1950 — Alan Turing publishes "Computing Machinery and Intelligence," which proposes what we now call the Turing Test and asks whether machines can think.
• 1956 — The Dartmouth Summer Research Project, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, gives the field its name: artificial intelligence.
• 1960s–1970s — Symbolic AI dominates: hand-coded rules and logic. Strong predictions of imminent human-level AI are followed by the first "AI winter" when those predictions don't pan out.
• 1980s — Expert systems briefly take off in industry. Another winter follows.
• 1986 — The backpropagation algorithm (the trick that makes training neural networks practical) is popularized by Rumelhart, Hinton, and Williams. Quietly, the foundation of the modern era is laid.
• 1997 — IBM's Deep Blue beats world chess champion Garry Kasparov. Narrow AI, but a milestone.
• 2012 — AlexNet, a deep neural network trained on GPUs, wins the ImageNet image-classification competition by a huge margin. This is the big bang of modern deep learning. Everything since flows from this.
• 2014–2016 — Rapid progress in image recognition, speech, machine translation. Generative adversarial networks (GANs) make AI-generated faces newsworthy.
• 2017 — Google researchers publish "Attention Is All You Need," introducing the transformer.
• 2018–2020 — OpenAI releases successively larger language models (GPT, GPT-2, GPT-3). Each scale-up surprises researchers with new emergent abilities.
• 2022 — Stable Diffusion is released as an open model in August. ChatGPT launches in November. The current AI boom begins.
• 2023–2026 — Capable LLMs and image models become commodities. Multi-modal models that handle text, images, audio, and video together arrive. Image-to-video models (Sora, Kling, Runway, Wurt.app's Vidwurt) cross from research demo into usable product.

Glossary of AI terms

The terminology is most of the difficulty. Here is the short list, in plain language. Skim it once and the rest of the AI internet becomes much more navigable.

Artificial Intelligence (AI)

Software that performs tasks normally associated with human intelligence by learning from data.

Machine Learning (ML)

The branch of AI where the system learns to perform a task from examples instead of being given explicit rules.

Deep Learning

A subset of machine learning that uses large multi-layer neural networks. Deep learning powers nearly all of today's notable AI.

Neural Network

A mathematical model loosely inspired by the brain. Inputs flow through layers of small processing units, each applying a learned weight, until the network produces an output.

Large Language Model (LLM)

A neural network trained on huge amounts of text that can read and write language. ChatGPT, Claude, and Gemini are all LLM-based.

Diffusion Model

A generative model that creates images (and increasingly video and audio) by starting from noise and gradually refining it into a coherent result.

Transformer

The neural-network architecture introduced in 2017 that underpins almost every modern LLM and most state-of-the-art image and video models.

GPU

Graphics Processing Unit. Hardware originally designed for video games that turned out to be ideal for the parallel math involved in training and running neural networks.

Training

The process of feeding data through a neural network and adjusting its weights so its outputs better match the expected results.

Inference

Running a trained model on new input to get an output. When you ask a chatbot a question or generate an image, you are doing inference.

Fine-tuning

Continuing the training of an already-trained model on a smaller, focused dataset to specialize it for a task or style.

RLHF

Reinforcement Learning from Human Feedback. Humans rank model outputs and the model is trained to prefer the higher-ranked ones.

Hallucination

When an AI model produces output that is fluent, confident, and false. A consequence of sampling from a probability distribution rather than consulting a database.

Prompt

The input you give to a generative model — the question, instruction, or description it should respond to.

Embedding

A list of numbers that represents a piece of content in a way that captures its meaning, so computers can compare meanings mathematically.

Token

The chunks an LLM actually reads and writes — usually a few characters or part of a word, not a whole word. LLM context limits and pricing are usually measured in tokens.

Context window

The maximum amount of text (in tokens) an LLM can consider at once when producing a response.

Generative AI

AI models whose job is to produce new content — text, images, audio, video, code — rather than just classify or predict.

Narrow AI

AI that performs one specific task very well but cannot generalize beyond it. All AI in production today is narrow AI.

AGI

Artificial General Intelligence. Hypothetical AI capable of performing any intellectual task a human can. Does not currently exist.

Computer Vision

The branch of AI that processes images and video — detecting objects, segmenting scenes, reading text, generating new pictures.

Natural Language Processing (NLP)

AI that reads, writes, translates, and analyzes human language. Today nearly all production NLP is done with large language models.

Supervised Learning

Training a model on labeled examples so it can predict the label for new ones. The workhorse of practical machine learning.

Self-supervised Learning

Training where the labels are generated from the data itself — for example, hiding the next word and asking the model to predict it. The technique behind almost every modern LLM and image model.

Reinforcement Learning

Training where a model takes actions in an environment and gets numerical rewards for good outcomes, learning a policy that maximizes long-run reward.

Parameters

The internal weights inside a neural network. "Model size" almost always means parameter count, which today ranges from hundreds of millions to hundreds of billions.

RAG

Retrieval-Augmented Generation. Combining an LLM with a document search step so the model answers from real source text instead of training-data memory.

AI Agent

An LLM given the ability to call external tools and pursue a multi-step goal autonomously, deciding what to do next at each step.

Multimodal

A model that handles more than one kind of input or output — typically text plus images, audio, or video.

Open-weight model

An AI model whose internal weights are publicly downloadable, so anyone with enough hardware can run, modify, or fine-tune it locally. Often loosely called open-source AI.

Claude

The large language model made by Anthropic. Known for very long context windows and a writing style focused on clarity and reluctance to invent facts.

Codex

OpenAI's brand for code-specialized models. The 2025–2026 Codex is a cloud-based software-engineering agent that edits codebases and opens pull requests.

One useful tip from working with this terminology every day: the AI field has a habit of using ordinary words for very specific technical things. "Attention," "memory," "understanding," "learning," "reasoning" — when you read them in an AI context, mentally translate to "the technical mechanism the field is calling that, which is not the same as the human version."

AI vs machine learning vs deep learning vs generative AI

This is one of the most-searched questions in the entire AI category, and the relationship is actually simple: each one is a subset of the previous one.

• AI is the umbrella — any software that performs tasks usually associated with human intelligence.
• Machine learning is the subset of AI where the system learns from data instead of being explicitly programmed.
• Deep learning is the subset of machine learning that uses large neural networks. Today, almost all impressive AI is deep learning.
• Generative AI is the subset of deep learning whose job is to produce new content rather than classify or predict.

So when somebody says "we use AI," in 2026 they almost certainly mean machine learning, almost certainly mean deep learning, and very likely mean generative AI. The terms are nested like Russian dolls.

The modern AI vocabulary you'll keep running into

A handful of concepts come up constantly in AI writing in 2026 that were either obscure or non-existent five years ago. Skim them once and most of the field's recent jargon stops being mysterious.

Parameters

A parameter is one of the internal weights inside a neural network — one of the billions of numbers that get nudged during training. Modern frontier LLMs have anywhere from tens of billions to hundreds of billions of parameters; smaller open-weight models run from a few hundred million to a few billion. More parameters generally means more capacity to memorize patterns, but it also means more compute to train and more memory to run. "How big is the model" almost always means "how many parameters does it have."

Fine-tuning

Fine-tuning takes a model that was already trained on general data and continues training it on a smaller, focused dataset. The base model brings broad capability; the fine-tuning step specializes it — for a particular tone of voice, a particular domain (legal, medical, code), or a particular task. Fine-tuning is much cheaper than training from scratch and is how most companies that "have their own AI" actually have it. Wurt.app's prompt enhancer, for example, does not require Wurt.app to train a model from zero; it builds on a fine-tuned LLM under the hood.

RAG (retrieval-augmented generation)

RAG is the standard pattern for giving an LLM access to information that wasn't in its training data. The setup is: you keep your documents in a database, you index them as embeddings, and at query time you retrieve the most relevant chunks and put them into the model's context window before it answers. RAG is how most "chat with your documents" products work, how customer-support bots ground answers in a knowledge base, and how a model can answer questions about events that happened after its training cut-off. It also reduces hallucination, because the model is reading actual source text instead of guessing.

AI agents

An agent is an LLM (usually) that has been given the ability to call tools — search the web, run code, query a database, send an email, edit a file — and instructed to pursue a goal across multiple steps without supervision at every step. The model decides what to do next, executes the tool call, looks at the result, and decides what to do next after that. Agents are the part of generative AI that is changing fastest in 2026. They are powerful when the loop works and frustrating when it doesn't.

Multimodal models

"Multimodal" means the model handles more than one kind of input or output. A text-and-image model can read images and write captions, or read a screenshot and answer questions about it, or do both at once. The latest frontier models are increasingly multimodal across text, images, audio, and video. This is what lets ChatGPT, Claude, and Gemini accept a photo as part of a question, and what lets video generators take an image and a prompt at the same time.

Open-source AI

Some powerful models are released with their weights freely downloadable — Meta's Llama family, Mistral, the Stable Diffusion family, Qwen, and many others. These are usually called "open-source" or, more precisely, "open-weight" models. Anyone with enough hardware can run them locally, modify them, fine-tune them on their own data, and ship products built on top. This is the counterweight to the small handful of closed labs that train the largest frontier models, and it's why a hobbyist with a single GPU can do real AI work in 2026.

The major AI assistants and tools you'll hear named

The same handful of products come up over and over once you start paying attention to AI. Here is the lay of the land in 2026, focused on what each one is actually for.

ChatGPT (OpenAI)

The product that made generative AI a household word. ChatGPT is OpenAI's chat interface on top of the GPT family of large language models. It does general-purpose conversation, drafting, summarizing, coding, image generation through DALL-E, and image input. It is the largest chatbot by usage and the default for most users who want one chat product.

Claude (Anthropic)

Claude is the LLM made by Anthropic, a lab founded in 2021 by former OpenAI researchers. The Claude family — Claude 3, Claude 3.5 Sonnet, Opus, Haiku, and the Claude 4 line — is the most-cited alternative to GPT and the model many engineers reach for when they want long-context reasoning or careful writing. Claude is known for a few specific traits in 2026: very large context windows (handling hundreds of thousands of tokens at once), a writing style that tends toward clarity and reluctance to make things up, and explicit emphasis on what Anthropic calls "constitutional AI" — a training process where the model is taught to follow a written list of principles instead of being purely tuned by human raters. Claude has a consumer chat product at claude.ai, an API that thousands of products are built on, and a coding-specific product called Claude Code that runs in a terminal and edits files directly. In the developer-tool ecosystem, Claude is the model behind a large fraction of agentic coding products and IDE assistants. If you are picking a second AI to keep open in another tab next to ChatGPT, Claude is the usual pick.

Codex (OpenAI)

"Codex" is OpenAI's brand for code-focused models and products. The original Codex was a 2021 fine-tune of GPT-3 specialized for programming, and it powered the first version of GitHub Copilot. The name was retired and then revived: in 2025 OpenAI launched a new Codex — a cloud-based software-engineering agent that takes a task description, runs in its own sandboxed environment, edits a codebase, runs tests, and opens a pull request when done. The 2026 Codex sits alongside ChatGPT for plain conversation and is aimed at the niche where the work to be done is "modify this repository" rather than "answer this question." The underlying model is a code-specialized variant of OpenAI's frontier line. Practical use ranges from generating routine boilerplate to running long autonomous coding sessions with light human review at the end. Codex is one of two products (Claude Code is the other) that defined what a "coding agent" looks like in practice in 2026.

Gemini and the Google AI umbrella

Google's AI work spans more brands than is convenient. Gemini is the current flagship LLM family — it powers the Gemini chat product, the AI features inside Google Search ("AI Overviews"), the AI inside Google Workspace, and the API that third parties build on. Before Gemini there was Bard, and before that LaMDA and PaLM; all retired. Separate from Gemini, Google's research arm DeepMind develops AlphaFold (protein folding), AlphaGo (board games), and a stream of academic models. There is also Imagen for images and Veo for video. The whole stack is "Google AI," but in 2026 most things a user actually touches are Gemini-branded.

The rest of the landscape

• Llama (Meta) — Meta's open-weight LLM family. Free to download, free to fine-tune, the foundation of a huge open-source ecosystem.
• Mistral — French lab known for compact, capable open-weight models.
• Grok (xAI) — Elon Musk's lab. Conversational chatbot with looser content policies than the major labs.
• DeepSeek and Qwen — Chinese labs shipping competitive open-weight models that anyone can download and run.
• Midjourney — image generation, originally Discord-only, now with a web app. Famous for a particular painterly aesthetic.
• Stable Diffusion (Stability AI) — the open-weight image-model family that started the open generative-AI ecosystem in 2022.
• Sora (OpenAI), Runway, Kling, Pika, Wurt.app's Vidwurt — video and image-to-video generators.
• GitHub Copilot and Cursor — code assistants embedded in editors. Copilot is the original; Cursor is the IDE that grew up around AI-first coding.

What AI is actually used for

Stripped of the marketing, here are the categories where AI is genuinely doing work in the real world today.

Information work

• Drafting and rewriting text — emails, reports, marketing copy, code.
• Summarizing long documents, meetings, and research.
• Translating between languages with quality close to professional human translators for major language pairs.
• Answering questions in natural language — replacing a lot of "search and skim" workflows.

Creative tools

• Generating images from text prompts (Picwurt, Midjourney, DALL-E, Stable Diffusion).
• Generating short videos from text prompts or starting still images (Vidwurt, Sora, Kling, Runway).
• Producing music, voice clones, and sound effects.
• Editing photos and videos — removing backgrounds, changing styles, upscaling.

Software

• Code assistants that suggest, complete, and refactor code in real time.
• Automated bug detection, vulnerability scanning, and test generation.
• "Agentic" workflows that string together tool calls to complete multi-step tasks.

Science and medicine

• Protein-structure prediction (AlphaFold) — a Nobel-Prize-winning result that reshaped molecular biology.
• Reading medical images for radiology, pathology, and ophthalmology, often matching specialist accuracy on narrow tasks.
• Discovering candidate drug molecules and materials.

Background infrastructure you already use

• Spam filtering, fraud detection, and recommendation feeds.
• Autocomplete, spell-correct, and predictive text on your phone.
• Speech recognition (every voice assistant, every dictation feature, every captioning system).
• Search ranking — modern search is impossible without machine learning.

What AI cannot do (yet, or at all)

The honest list is just as important as the wins. Today's AI:

Hallucinations: the failure mode that defines today's AI

The single most important limitation to understand is hallucination — when a model produces fluent, confident, false output. This is not a bug that will be patched out in the next release; it falls out of how generative models work. The model is sampling from a probability distribution over plausible next tokens, not consulting a database of facts. When the prompt sits inside the part of that distribution the training data covered well, the output tends to be accurate. When the prompt drifts into territory the training data didn't cover, or the model has to combine specific details (names, dates, citations, code identifiers, legal citations), the output drifts toward whatever was statistically nearby — which may or may not be true. Hallucinations get worse with longer outputs, with topics outside the training data, and with the model trying too hard to be helpful when the honest answer is "I don't know." Practical defenses: keep the model's claims inside its training scope, ask for sources you can check, use retrieval-augmented generation to ground answers in real documents, and verify any specific fact before acting on it. Treat the model as a fluent first draft, not a final source.

Beyond hallucination, today's AI:

• Doesn't know what is true. It produces fluent output. Verifying that output against reality is your job, not the model's.
• Doesn't have a stable memory by default. Most LLMs forget your conversation between sessions unless the product wraps them in some kind of long-term memory layer.
• Doesn't continuously learn from new events. Models are trained, frozen, and shipped. They have a training cut-off date and don't know what happened after it unless they're given fresh data at inference time.
• Doesn't reliably do exact arithmetic, multi-step logic, or anything requiring perfect precision. They approximate. Use a calculator or a code interpreter for the parts that need to be exact.
• Doesn't reason from first principles. It pattern-matches. The patterns can be very good — sometimes almost indistinguishable from reasoning — but novel problems outside the training distribution still trip it up.
• Doesn't experience anything. No senses, no preferences, no agenda. Treating model output as if it expressed feelings is a category error.

Ethics, bias, and safety — without the moralizing

Three concerns are worth taking seriously, separated from the louder rhetoric on either side:

• Bias inherited from training data. Models trained on human content reflect human biases. Image models trained on the public web initially over-produced certain demographics for "doctor" and other professions. Text models inherit the slant of their training corpus. Mitigations exist but they are imperfect, and "the model is unbiased" is never a true claim — only "we measured it on these benchmarks and it scored this way."
• Misuse. Capable generative models can produce convincing fake images, voices, and text. That capability is genuinely useful for legitimate creative work and genuinely usable for fraud, harassment, and misinformation. Watermarking, detection tools, and platform-level policies help; they do not eliminate the problem.
• Concentration of capability. Training frontier models requires hundreds of millions to billions of dollars of compute. That naturally concentrates the technology in a few labs and a few countries. Open-weight models partially counterbalance this and are an active and contested policy question.

The deeper "alignment" debate — how to make sure increasingly capable AI systems pursue goals that are actually good for people — is a real research area, and it isn't science fiction. It's also not something any individual user needs to solve before opening ChatGPT.

How to start using AI today

The fastest way to understand AI is to use it for something you actually want done. A pragmatic starter path:

1. Pick one task. Drafting an email. Summarizing a long article. Generating a hero image for a project. Outlining an essay. Writing a unit test.
2. Pick one tool that fits the task. Text and conversation: ChatGPT, Claude, or Gemini. Image generation: an AI image generator like Wurt.app's Picwurt, Midjourney, or DALL-E. Video: an AI video generator like Wurt.app's Vidwurt, Sora, Kling, or Runway. Code: GitHub Copilot, Cursor, Claude Code.
3. Write a clear prompt. Describe the goal, the format, the audience, and the constraints. Most "the AI is bad at this" complaints turn out to be prompts that didn't specify what was actually wanted. The prompt engineering guide covers the full mental model.
4. Iterate. The first output is rarely the right one. Tell the model what to change. Modern chatbots are fine to argue with — that's literally how they're designed to work.
5. Verify anything that needs to be true. AI is great at producing plausible content and bad at guaranteeing it's correct. The verification step is yours.

The most useful single mental shift is to stop treating AI like a search engine and start treating it like a confident, fast, slightly overconfident intern. You give it work, you review the work, you send back changes, you use what's good.

If creative generation is the part you're curious about, the easiest hands-on starting point is image generation, because the feedback loop is instant and visual. From there, image-to-video models extend the same workflow into motion. The AI recommendation guide walks through which generative tool fits which job, and the AI art styles guide shows the visual vocabulary that prompts rely on.

Frequently asked questions