What is Generative AI
We explain how generative AI differs from classical algorithms and why an LLM is neither a search engine nor a database.
Generative AI: where the name comes from
The word "generative" means the model creates new content rather than retrieving it from stored records. Unlike classical algorithms (e.g., a spam filter with hand-written rules), generative models train on vast datasets and learn to reproduce patterns — in text, code, images, and audio.
How it differs from classical AI
| Classical AI | Generative AI | |---|---| | Rules written by hand | Patterns learned from data | | Returns yes/no or a number | Generates free-form content | | Brittle on new inputs | Generalizes flexibly | | Transparent logic | Hard to explain individual outputs |
A classic example of classical AI: a spam filter that checks for words like "win," "free," "urgent" and blocks the email. It works — but add one new trigger word and you need to manually update the rules.
A generative model trained on labeled spam/not-spam examples learns contextual signals that a human never would have thought to encode.
What generative AI can do today
- Text: write, rewrite, translate, summarize, answer questions
- Code: generate, explain, debug, write tests
- Images: DALL-E, Midjourney, Stable Diffusion
- Audio: speech synthesis, voice cloning
- Video: short clip generation (Sora, Runway)
- Multimodal: accept images/PDFs as input, return text
Why it's not a search engine
A search engine finds existing pages. An LLM generates a response on the fly — based on statistical patterns absorbed during training. This means:
- The answer may not exist anywhere online — the model invented it
- The answer may be plausible but false (hallucination)
- There's no "taken from here" link by default — if you need sources, you need RAG or a search-enabled model
Practical takeaway
Think of an LLM as a very fast, well-read colleague who has read almost everything but occasionally lies with confidence. Their input is valuable — but requires verification for anything critical.
Correct mental model:
LLM = draft acceleration + idea generation + concept explanation
NOT = authoritative fact source without verification
Open ChatGPT or Claude. Ask one question whose answer you already know. Evaluate accuracy. Then ask about something highly specialized in your field and check the answer against your own knowledge. Note where the model got it wrong.
Copy and adapt to your context. Text in angle brackets should be replaced.
Explain <topic> to me as if I'm a smart professional but a complete beginner in this area. Use analogies from everyday life. After the explanation, give me 3 questions I should ask to verify I understood correctly.
- Taking model responses as facts without verification — especially dangerous with numbers, dates, and names.
- Assuming confident tone = correct answer. Models write both correct and incorrect content with equal confidence.
- Comparing AI to a search engine and feeling disappointed there are no links — they're different tools with different strengths.
- Ask the model to state its confidence: "How confident are you in this answer? What might be inaccurate?"
- Use the model for draft generation; keep verification to yourself — this is the fastest workflow.
- For factual tasks, enable search (Perplexity, ChatGPT with web access) — this addresses the hallucination problem with sources.
When you need to quickly generate a draft, explain a concept, brainstorm, or rework existing text.
When factual accuracy is critical and you have no way to verify the answer — e.g., medical or legal decisions without expert review.