Daily Tech Wizard Newsletter

Blog /

Generative AI Explained: How It Works in 2026

A sunlit home-office desk in an Australian brick-veneer house, framed low across the timber surface: an open laptop angled

Last reviewed 16 November 2026

Generative AI is a class of machine learning system that produces new content - text, images, code, audio or video - by learning statistical patterns from very large training datasets and then predicting plausible continuations. Instead of sorting or scoring data that already exists, it makes a fresh artefact each time you ask.

Example: you type "write a two-line thank-you note to a customer called Priya", and the system writes a note that has never existed before.

Key takeaways

  • Generative AI predicts likely output, token by token. It does not look answers up in a database.
  • That one fact explains almost every strength and every failure, including hallucination.
  • Different jobs use different model families: transformers for text and code, diffusion for images, video and audio.
  • It pays off fastest on a task you already repeat weekly, and only when a human checks the output.

On this page

Last updated: 16 October 2026.

What Is Generative AI? (The 40-Second Answer)

Generative AI, often shortened to genAI, is software that has been shown enormous quantities of examples and has learned the patterns inside them well enough to produce convincing new examples of its own. Text in, text out. Text in, picture out. Picture in, description out.

The important word is plausible. The system is built to produce something that looks like a good answer, which is usually the same thing as a correct answer, but not always.

Underneath, it is a neural network with billions of adjustable numbers, trained until it can guess the next piece of a sequence with uncanny accuracy. Everything else - the chat window, the voice mode, the agent that books your flights - is scaffolding around that one trick.

Generative AI vs Traditional AI vs Machine Learning

These terms get used as if they are interchangeable. They are not, and the difference is easy to hold onto.

Predictive AI: Sorting What Already Exists

Traditional or predictive AI looks at data you give it and puts a label or a number on it. Is this email spam? Will this customer churn in the next 90 days? Is this transaction fraudulent?

The output is a judgement about existing data. Your bank has been running this kind of model for decades, and it is still the workhorse of most commercial AI.

Generative AI: Producing Something New

Generative AI produces an artefact. Not "this email is spam" but "here is an email". The output did not exist before you asked, and asking twice may well get you two different versions.

That is the whole distinction. Predictive AI describes the world; generative AI adds to it.

Where Deep Learning and LLMs Fit

Think of it as nested boxes. Machine learning is the broad method - systems that improve from data rather than from hand-written rules. Deep learning is the subset that uses neural networks with many layers. Generative AI is an application built on deep learning.

A large language model, or LLM, is one specific kind of generative model: a very large neural network trained on text. A foundation model is simply a big general-purpose model that lots of narrower products get built on top of.

And a plain-speaking note for 2026: when a vendor's marketing page says "AI", it almost always means generative AI. The word has quietly narrowed.

How Generative AI Works (One Prompt, Traced End to End)

Most explainers stop at "it predicts the next word". Let us follow one prompt all the way through, because once you have seen the machinery, the failure modes stop being mysterious.

Our prompt: write a two-line thank-you note to a customer called Priya.

Step 1 - Your Words Become Tokens

The model cannot read letters. Before anything happens, your sentence is chopped into tokens - short chunks of text, roughly three-quarters of a word on average in English - and each token is swapped for a number from the model's fixed vocabulary.

Our prompt comes out at somewhere around fifteen tokens. Common words like "write" and "to" are single tokens. A less common name like "Priya" may well be split into two or three pieces, which is exactly why models sometimes mangle unusual names and spellings.

Those numbers then become vectors - long lists of decimals that encode how each token relates to every other token. This is the only form in which the model ever "sees" your request.

Step 2 - The Model Predicts the Next Token

The network runs your token sequence through its layers and produces one thing: a probability score for every single token in its vocabulary, as the candidate for what comes next.

Not a sentence. Not an answer. A ranked list, perhaps 100,000 entries long. After our prompt, the top candidates might be "Dear", "Hi", "Hello" and "Thank", each with a different weight.

One token is chosen. It gets appended to the sequence, and the whole process runs again from the top - now with sixteen tokens of input instead of fifteen. Then seventeen. The note you read was built one chunk at a time, left to right, with no plan for the ending when the first word was picked.

There is no database lookup in that loop. There is no internal fact-checker, and no concept of truth. There is a probability distribution, and the text that falls out of it. Hold on to that, because it is the whole explanation for hallucination later.

Step 3 - Sampling, Temperature and Why Answers Vary

If the model always picked the single highest-probability token, it would be repetitive and oddly flat. So the choice is sampled, and two dials control how adventurously.

Temperature flattens or sharpens the distribution. Low temperature means it nearly always takes the safest token, which is what you want for data extraction. Higher temperature lets less likely tokens through, which is what you want for brainstorming.

Top-p (nucleus sampling) trims the list to the smallest set of candidates whose probabilities add up to a threshold, say 0.9, then samples from that. Tail nonsense gets cut; genuine variety survives.

This is why the same prompt gives different answers on Tuesday and Thursday. It is not a bug, it is sampling. It is also why "it worked when I tested it" is a weak guarantee for anything that matters.

Where the "Knowledge" Came From: Training and Fine-Tuning

Three stages, in plain terms.

Pre-training. The model is shown a colossal corpus of text or images and repeatedly asked to predict missing pieces. Every wrong guess nudges its internal numbers. Do that at scale and grammar, facts, style and a surprising amount of reasoning get absorbed as a side effect.

Supervised fine-tuning. A much smaller, curated set of example prompts and good answers teaches the raw model to behave like an assistant rather than an autocomplete.

Reinforcement learning from human feedback (RLHF). Humans rank competing answers, and those rankings train a reward model that shapes tone, helpfulness and refusals. This is the stage that makes a model feel polite and safe, and also the stage that makes it sometimes sycophantic.

One consequence matters commercially: the model's knowledge is frozen at its training cut-off. Unless it has been given live tools, web search or document retrieval, it genuinely does not know what happened last month.

The Main Model Families and What Each Is Good At

Transformers and Large Language Models (Text and Code)

The transformer architecture was introduced by Vaswani and colleagues at Google in the 2017 paper "Attention Is All You Need". Its key idea, self-attention, lets the model weigh every token against every other token in parallel rather than reading strictly in order.

That parallelism is what made training at today's scale affordable. Every mainstream text and code model is a descendant.

Diffusion Models (Images, Video, Audio)

Diffusion models learn to reverse noise. During training, images are progressively corrupted into static; the model learns to undo each step. To generate, you start from pure noise and have the model denoise it, guided by your text prompt, until an image appears.

The approach was set out by Sohl-Dickstein and colleagues in 2015 and made practical by Ho and colleagues' denoising diffusion work in 2020. It now underpins essentially all serious text-to-image and much text-to-video generation.

GANs and VAEs: The Older Generation

A generative adversarial network, introduced by Ian Goodfellow and colleagues in 2014, pits two networks against each other: a generator making fakes and a discriminator trying to spot them. The results were startling for their time and GANs remain fast, but training is notoriously unstable and the variety of output is narrow.

A variational autoencoder, from Kingma and Welling's 2013 paper, squeezes data into a compact "latent space" and reconstructs it. VAEs rarely generate finished work today, but they are not obsolete - a VAE sits inside latent diffusion image models doing the compression step.

Generative Deep Learning Models in One Comparison Table

FamilyIntroducedBest forWeak atExample products
Transformer LLMs2017Text, code, reasoning, summarising, translationExact arithmetic, anything after its cut-offGPT, Claude, Gemini, Llama
Diffusion models2015 theory, 2021-22 mainstreamImages, video, audio from textText inside images, precise counts, slow samplingStable Diffusion, Midjourney, DALL-E
GANs2014Fast image synthesis, upscaling, face generationUnstable training, limited output diversityStyleGAN-family tools, some upscalers
VAEs2013Compression, latent-space editing, anomaly detectionBlurry standalone outputEncoder stage inside latent diffusion
Multimodal models2023 onwardMixed input and output: image in, text out; voice both waysCost, latency, uneven quality across modesCurrent flagship assistants

A Short History: From ELIZA to ChatGPT to Agents

1966 - ELIZA, the First Generative Chatbot

Joseph Weizenbaum at MIT published ELIZA in 1966. It worked by pattern-matching your sentence and reflecting it back, in the manner of a psychotherapist. Crude by any modern measure, yet people confided in it, which unsettled Weizenbaum enough that he spent years warning about it.

1987-2014 - Neural Nets, AI Winters and the GAN Breakthrough

Through the 1980s, Markov chain and n-gram generators produced the first statistical text machines: pick the next word based on the last one or two. Funny, shallow, and the direct ancestor of what your phone keyboard does.

Twice - in the mid-1970s and again from the late 1980s - grand promises outran results and research funding collapsed. Those are the AI winters, and they are worth remembering when hype arrives.

Progress continued quietly. Hochreiter and Schmidhuber published the LSTM in 1997, giving neural networks a workable memory for sequences. Kingma and Welling's VAE landed in 2013, Goodfellow's GANs in 2014.

2017 - "Attention Is All You Need" Changes Everything

The transformer paper removed recurrence and made training parallel. Within two years every leading language model used the architecture. OpenAI's GPT-2 arrived in 2019 and was initially released only in reduced form, with the organisation citing concerns about misuse for mass-produced text - a decision much argued over at the time.

2022 - When Generative AI Went Public

Images went first. DALL-E was shown in January 2021, Midjourney opened to the public in July 2022, and Stable Diffusion was released openly in August 2022, which put image generation on ordinary home computers.

Then on 30 November 2022, ChatGPT launched as a free research preview, and the whole thing stopped being a specialist topic within weeks.

2023-2026 - Multimodal Models, Agents and On-Device AI

Three shifts since. Models became multimodal, handling images, audio and documents in the same conversation. Reasoning modes arrived, where the model is trained to work through intermediate steps before answering, which measurably helps on maths and logic and costs more to run.

And models learned to use tools. An agent is a model in a loop with permissions: it can search, run code, call your systems, check the result and try again. That is where most 2026 product effort has gone, and where most of the new risk sits too.

Quietly alongside all that, small models now run on phones and laptops. Slower and less capable than the flagships, but private, offline and free to run - which for a lot of small-business tasks is the better trade.

What People Use Generative AI For

Writing, Summarising and Translating

This is where the return is most reliable. Turning a messy meeting transcript into actions, drafting a quote follow-up, rewriting a technical paragraph for a customer, translating a product description.

The pattern to notice: it is strongest when the source material is in the prompt. Summarising a document you supplied is low-risk. Asking it to recall facts from memory is not.

Can Generative AI Write Code? (Yes, With Caveats)

Yes, and it is genuinely good at a specific band of work: boilerplate, unit tests, regular expressions, translating between languages, and explaining unfamiliar code you have inherited. Developers consistently report real time savings on that routine layer.

It is weak on large multi-file changes, on security-sensitive logic such as authentication and permissions, and on anything that depends on parts of a codebase it cannot see. It will also invent a library function that sounds exactly right and does not exist.

Unreviewed generated code is a known source of introduced bugs and insecure patterns. Treat it like work from a fast, confident junior: useful volume, mandatory review.

Images, Video and Audio - and When Generative AI Art Started

Mainstream generative art dates from 2021-22, with DALL-E, Midjourney and Stable Diffusion. Text-to-video became usable for short clips through 2024-26, and voice cloning from short samples arrived alongside it.

For small businesses this is mostly good for concepting, backgrounds and social filler. Anything where a specific real product or a real premises must be accurate is still a job for a camera.

What Is Generative Design? (Engineering's Version)

Generative design is a CAD technique, and it is a different thing to generative AI despite the shared word. An engineer specifies goals and constraints - load paths, material, weight limit, manufacturing method, cost - and the software generates and evaluates many candidate geometries against them.

The output is often the organic, bone-like lightweight bracket you have seen in aerospace and cycling parts. It predates modern generative AI and rests on optimisation rather than language modelling, though the two toolsets increasingly borrow the same learning methods.

Customer Support, Search and Retrieval (RAG)

Retrieval-augmented generation, or RAG, is the practical fix for both stale knowledge and invention. Rather than trusting the model's memory, you first search your own documents for passages relevant to the question.

Those passages go into the prompt, and the model is instructed to answer only from them and to cite which one it used. If nothing relevant is found, it should say so.

This is how nearly every credible business deployment now works - the support bot that knows your warranty terms, the internal search that quotes the real policy. It does not eliminate error, but it changes the question from "what does the model remember?" to "did we retrieve the right document?", which is a question you can audit.

What Generative AI Still Gets Wrong

Hallucination: Confident and Incorrect

Hallucination is when the model produces fluent, confident, false output. It is not a malfunction. It is the prediction loop working exactly as designed on a question where plausible and true came apart.

The best-known example is a 2023 US federal case, Mata v. Avianca, where lawyers filed a brief citing judicial decisions that did not exist. ChatGPT had generated the citations, and when asked whether they were real, affirmed that they were. The court sanctioned the lawyers involved.

The lesson is sharper than "check your work". In a generative model, tone and accuracy are uncorrelated. A fabricated case citation is produced with exactly the same breezy confidence as a correct one, because the model has no signal that distinguishes them.

Add sampling to that, and the same prompt can give you a right answer today and a wrong one tomorrow.

Copyright, Training Data and Who Owns the Output

Two separate questions, both unsettled.

Inputs. Whether training a commercial model on copyrighted text and images without a licence is lawful is being litigated across multiple jurisdictions and has not been conclusively resolved. If your business depends on the answer, get advice rather than a blog post.

Outputs. Here there is more clarity, and it surprises people. Copyright protection generally requires a human author. The US Copyright Office has stated that purely machine-generated material is not registrable, and registration has been granted only for the human-authored arrangement and editing around AI output. Australian law takes a similarly human-centred view of authorship.

Practical read: your AI-generated logo or ad copy may well sit outside copyright protection entirely. Anyone can reuse it.

Energy, Cost and the Price Per Million Tokens

Commercial models are priced per million tokens, charged separately for input and output, with output usually dearer. The spread across the market is wide - small fast models cost a tiny fraction of flagship reasoning models for the same job, and prices have fallen steeply year on year. Check the vendor's current pricing page before budgeting; any number printed here would be stale within months.

The practical lesson is to match the model to the task. Running a flagship reasoning model on routine text cleanup is like running a tip truck to collect the post.

On energy, per-query consumption for text prompts is small - Google reported a median of about 0.24 watt-hours per text prompt to its Gemini app in 2025 - but the aggregate across billions of queries, plus training runs and image and video generation, is where the real footprint sits.

Tasks It Is Still Bad At

  • Precise arithmetic without a calculator tool attached. It is predicting digits, not computing.
  • Counting things in long inputs. Ask how many times a word appears in a 40-page document and you will get a confident guess.
  • Anything after its training cut-off, unless it has live search or retrieval.
  • Legal, medical and financial advice. Fluent, unaccountable, occasionally wrong in ways that cost money.
  • Long multi-step constraint lists. Give it eleven rules and two will quietly be dropped.
  • Knowing what it does not know. There is no reliable internal "I am unsure" signal, which is the root of all of the above.

How to Start Using Generative AI This Week (Three Steps)

Step 1: pick one task you already repeat weekly. Not a moonshot. The quote follow-up email, the job notes you type up every Friday, the product description for each new line. Repetition is where the payoff compounds.

Step 2: write a proper prompt. Four parts: role ("you are writing as a plumber in Geelong"), context (paste the real job notes), format ("under 120 words, no bullet points, plain English"), and one example of a past version you were happy with. That example does most of the work. Our prompt-writing guide goes deeper.

Step 3: verify before it leaves your hands. Every name, number, date, price and claim checked against a real source. Build the habit on day one - our notes on AI content accuracy checks cover a quick routine for it.

You can test the whole idea for nothing. The major assistants all run free tiers that are perfectly adequate for a trial, and if you want help choosing, see our AI tool comparison. Only pay once a task has earned it.

Generative AI FAQs

What is the definition of generative AI?

Generative AI is a type of machine learning system that creates new content - text, images, code, audio or video - by learning statistical patterns from large training datasets and predicting plausible continuations. Unlike predictive AI, which labels or scores existing data, it produces an original artefact in response to a prompt.

When was generative AI introduced to the public?

Generative image tools reached the public first, with DALL-E demonstrated in January 2021 and Midjourney and Stable Diffusion opening up in mid-2022. The breakthrough moment for the general public was ChatGPT's free launch on 30 November 2022, which took the technology mainstream within weeks.

What was the first generative AI chatbot?

ELIZA, written by Joseph Weizenbaum at MIT and published in 1966. It generated replies by pattern-matching the user's sentence and reflecting it back as a question. It had no learning and no model of meaning, yet users still formed real attachments to it, which troubled its creator for decades.

Is generative AI the same as ChatGPT?

No. ChatGPT is one product, built on one family of large language models. Generative AI is the whole field, covering image, video, audio, code and text systems from many organisations, including open models you can run on your own laptop. ChatGPT is a brand, not a category.

Does generative AI learn from my prompts?

It depends entirely on the product and plan. Consumer free tiers have often used conversations to improve models by default, while business and API tiers typically exclude customer data from training by contract. Check the specific terms and the data controls in your settings, and never paste client data into a tool you have not checked.

Is generative AI going to replace my job?

It is replacing tasks faster than jobs. Drafting, summarising, first-pass translation and routine code are being automated; judgement, accountability, physical work and client relationships are not. The realistic risk is being out-competed by someone in your trade who uses the tools well, rather than by the tools themselves.

The Short Version

Generative AI is a prediction machine of extraordinary fluency and no understanding. It builds output one token at a time from patterns in its training data, which makes it brilliant at anything resembling a pattern and unreliable at anything requiring certainty.

Know where it came from - 1966 for the first chatbot, 2017 for the architecture, 2022 for the public - and know its shape: strong on text you supply, weak on facts it must recall, dangerous when nobody checks. Give it your documents, give it a clear format, and give the output a read before you send it.

Generative AI only pays off when you point it at a task you already repeat. Grab our free prompt template pack and subscribe to the Daily Tech Wizard newsletter for one practical AI workflow each week.