This article was produced with AI assistance. Editorial standards apply — see our About editorial process.
Key takeaways
- Traditional (discriminative / predictive) models answer “what is this?” or “what happens next?” Generative models answer “make another one like the training distribution.”
- NIST AI 100-1 (January 2023) defines AI systems by outputs such as predictions, recommendations, or decisions. NIST AI 600-1 (July 2024) is the generative-specific risk profile.
- Microsoft’s AI 101 page: generative AI creates new content from prompts; traditional AI automates bounded tasks; predictive AI forecasts from history.
Two jobs for the same word “AI”
How generative AI works vs traditional AI comes down to output type: traditional systems classify or predict; generative systems sample new text, images, or other media from a learned distribution.
NIST AI 100-1 (January 2023) calls an AI system an engineered, machine-based system that can, for given objectives, generate outputs such as predictions, recommendations, or decisions, adapted from the OECD AI Recommendation and ISO/IEC 22989:2022. That umbrella already includes more than chatbots. Daily assistants that wrap either kind of model are listed in AI assistants.
How the pipelines differ
Microsoft’s generative-versus-other-types page is the plain-language split: generative models use deep learning to create new content from natural-language prompts; traditional AI (as Microsoft groups it) automates specific tasks with predefined rules and includes predictive and conversational products; predictive AI forecasts from historical data. Microsoft’s FAQ states machine learning underlies all of these, with traditional ML aimed at classification and prediction. Ignore Microsoft’s “strong AI” label for generative tools—that is vendor wording, not the academic AGI sense.
A fraud score, a next-best-offer rank, and a demand forecast are traditional/predictive jobs. A draft email, a storyboard still, and a synthetic molecule proposal are generative jobs. Shortcut lists that mix both live in productivity hacks.
Why NIST split a generative profile
NIST AI 600-1 (26 July 2024) is a companion profile for generative AI under the same RMF, written to surface risks that generic “AI system” language under-specifies: unconstrained new content, prompt injection, and related socio-technical failure modes. Use 100-1 for the system definition; use 600-1 when the system emits open-ended media.
Prompt playgrounds are not a substitute for that split. Fun experiments belong on fun prompts. Keep the definitional page inside Productivity & Workflows and the AI tools silo.
FAQ
Is ChatGPT traditional AI?
No in the Microsoft create-versus-forecast sense. A chatbot that writes new sentences is generative. A rules engine that routes a ticket to a queue is traditional automation, even if a vendor stamps “AI” on the box.
Where do recommender systems sit?
Usually predictive/traditional: they rank existing items. When a system writes a new item (a unique subject line, a new image), that module is generative.
Do I need two vendors?
Not necessarily. Many products bolt a generative interface onto an older scoring backend. Ask which component produces the artifact you will ship.
