Editorial illustration.
Originally published 16 June 2023; rewritten and reviewed on 15 October 2026.
A conversation about AI can move from a spam filter to machine learning, then to a supposedly self-aware system, without acknowledging that these are different categories. I would separate the problem being solved from the technique used and the capability being claimed.
Start with the task
Common application areas include language processing, computer vision, prediction, recommendations and planning. A product may combine several of them. Naming an area helps describe the work, but it does not tell us how reliably the system performs it.
For a document classifier, ask which documents and categories matter. For an image system, ask what objects or conditions it must recognise. Those questions give evaluation a concrete target.
Distinguish techniques from capabilities
Machine learning is one broad approach within AI, and deep learning is a family of machine-learning methods. Rule-based systems and search algorithms can also contribute to an AI application. These labels do not form a single mandatory development sequence.
The original article discussed “levels” of AI, including general and self-aware concepts. Such labels should not be presented as a verified roadmap that every current model is climbing. Useful performance on several tasks does not by itself prove human-like understanding, consciousness or unrestricted competence.
Evaluate the system you actually deploy
Define inputs, expected outputs and unacceptable failures. Check the data and context required to reproduce a result. Compare the system with a baseline and inspect errors on examples that resemble actual use.
An application can be valuable while having narrow capabilities and clear limitations. My preference is to describe those limits plainly. It gives the team a stronger basis for deciding where the system belongs than an ambitious label that cannot be tested.
Related reading
Images from the original article
The following images are retained as historical context; interfaces and configuration options may have changed.
