What is the difference between generative AI and traditional AI?

Traditional AI mostly *classifies or predicts* — it takes an input and produces a label or a number. Generative AI *creates* new content — text, images, audio, code. The shift is from 'which category is this' to 'produce something that doesn't exist yet'. Traditional AI, which has been quietly in products for years: - Spam filter: is this email spam or not? (classification) - Recommendation engine: what will this user click? (prediction/ranking) - Fraud detection: is this transaction suspicious? (classification) - Credit scoring, demand forecasting, churn prediction (regression) - Image recognition: is there a cat in this photo? (classification) The output is constrained and usually verifiable. There's a correct answer and you can measure accuracy against it. Generative AI: - Write an essay, a summary, an email, a function - Generate an image from a description - Produce speech, music, video - Translate, rewrite, restructure The output space is effectively infinite and there's no single correct answer, which changes almost everything about how you build with it. What that shift actually means in practice: 1. **Evaluation becomes hard.** With a classifier you compute accuracy. With generation, 'is this summary good?' has no ground truth — you need human judgement, rubrics, or model-based grading, all imperfect. This is why so much of the engineering effort in modern AI products goes into evaluation rather than the model. 2. **Failure looks different.** A classifier fails by outputting the wrong label, which is visible. A generative model fails by producing fluent, confident, wrong content, which is not. 3. **Generality replaced specificity.** Traditional models were trained per task — a spam model did spam and nothing else. One large language model does summarising, translating, coding and analysis without task-specific training. That generality is the genuinely new thing. 4. **The interface changed.** You instruct it in natural language rather than engineering features and retraining. That's why adoption was so fast — the barrier dropped from 'ML team' to 'anyone who can type'. The important nuance for anyone choosing a technique: traditional approaches are not obsolete. For structured tabular data — fraud, forecasting, risk scoring — gradient boosted trees still beat language models comfortably, run in milliseconds, cost almost nothing, and can be explained to a regulator. Using an LLM for a problem that a classifier solves is a common and expensive mistake. Match the tool to the shape of the problem, not to the news cycle.