- text
AI writes a marketing email
A model predicts the next phrase over and over until a complete, on-tone email exists.
It writes confidently about offers, dates and prices it has no way of knowing — every factual claim is yours to check.
Explained by: What is AI, really?
- text
Instantly draft 20 headlines
Sampling the same prompt many times produces genuine variety in phrasing, not twenty rewrites of one idea.
Variety is not judgement — it cannot tell you which of the twenty will land with your audience.
Explained by: What is AI, really?
- text
Summarise customer interviews in seconds
The model compresses a transcript by keeping high-signal phrasing and dropping repetition.
It flattens the outlier — the one furious customer whose comment mattered most gets averaged away.
Explained by: What is AI, really?
- research
AI sees and summarises charts
A vision model reads the image, converts what it sees into text, then reasons over that text.
Misread an axis or a legend once and every number downstream is wrong, stated just as smoothly.
Explained by: Why AI makes things up
- research
AI researches a competitor teardown
A search tool fetches live pages and the model organises what it retrieved into a structure.
Retrieval quality caps everything — cited-looking claims can come from a marketing page or from nothing at all.
Explained by: Why AI makes things up
- images
AI makes images
A diffusion model starts from noise and removes it step by step toward your description, including product shots in any scene.
It has no idea what your product actually looks like, and it still fumbles text, hands and repeated details.
Explained by: It sees, hears and speaks too
- video
AI makes video
Frame-by-frame generation with a consistency constraint so the scene holds together over time.
Physics and continuity break over a few seconds — objects morph, and long shots stay out of reach.
Explained by: It sees, hears and speaks too
- audio
AI sings
A model generates audio directly, predicting waveform or spectrogram chunks the way a text model predicts words.
Voice likeness is a consent and rights problem before it is a technical one.
Explained by: It sees, hears and speaks too
- audio
AI talks
Text-to-speech with learned prosody, so the emphasis lands where a human would put it.
It performs emotion it does not have — the wrong tone on a sensitive message reads as careless.
Explained by: It sees, hears and speaks too
- code
AI codes an app from text
It has seen enormous amounts of code, so it predicts plausible working code for a described feature.
Plausible code compiles and still does the wrong thing — nobody was accountable for the edge cases.
Explained by: What is AI, really?
- code
AI ships a SaaS product
Scaffolding, wiring and boilerplate — the most patterned parts of software — get generated end to end.
The last ten percent (auth, billing, data safety, uptime) is exactly where generated code is weakest.
Explained by: What is AI, really?
- code
AI personalises landing pages
Copy and layout variants are generated per segment and served dynamically.
Personalisation drifts into claims you never approved, at a scale nobody is reading.
Explained by: What is AI, really?
- agents
AI automates routine tasks
A loop of plan, act, observe, repeat runs a multi-step task without you in the middle.
Nothing in the loop notices when step two was wrong — it just keeps going, confidently.
Explained by: AI agents — when AI does the work
- agents
AI agents working for you
Several tool-using loops coordinate, each handling a slice of a larger job.
Errors compound across steps, and the cost and the blast radius grow with the autonomy.
Explained by: AI agents — when AI does the work