An AI system, in the wording used by the OECD, is a machine-based system that infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions. Software that only follows rules a person wrote out in full, such as a calculator or a spreadsheet formula, infers nothing from data, so it sits outside that definition even when a product page calls it "smart." The test is whether the behavior comes from learned patterns or from instructions that a developer spelled out line by line.
Key takeaways
- The OECD definition, revised in 2023, centers on inference: a system works out from its input how to produce its output.
- Fixed rules written by a person (a keyword list, a formula, a schedule) are ordinary software, not AI.
- A model trained on examples can handle cases its author never listed, which is the practical difference.
- Some systems sit on the border, and experts disagree about them, so the label alone settles little.
- Four plain questions about data, outputs, and updates usually reveal whether a product really uses AI.
What the official definition says
The OECD definition of an AI system was revised by member countries in 2023. It reads: a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. It adds that AI systems differ in their levels of autonomy and adaptiveness after deployment.
Three parts of that sentence do most of the work. The word "infers" means the system derives its behavior from input and does not simply look it up. The outputs are specific kinds of results (a prediction, a piece of content, a recommendation, a decision). And the results must be able to affect something, whether a physical device or a digital process. The U.S. National Institute of Standards and Technology uses a closely related wording in its AI Risk Management Framework 1.0, published in January 2023. It describes an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments, and it credits the OECD as an adapted source.
A worked example: two spam filters
Consider two email filters that both move junk to a separate folder.
The first one is a list of rules. A developer writes: if the subject contains a listed phrase, move the message to junk. Every decision it makes can be traced to a line the developer typed. If spammers change one letter in the phrase, the filter misses the message until a person edits the list. It follows instructions and infers nothing.
The second filter is built differently. Its developers collect a large set of messages that people have labeled as spam or not spam, and a learning algorithm adjusts a set of numeric weights until the model separates the two groups well. Nobody writes the phrase list. When a new message arrives, the model computes a score from patterns it picked up in training and routes the message by that score. It can flag a message built from wording nobody listed, and it can also make mistakes that no rule explains, because its behavior comes from the data and not from a readable instruction.
Only the second filter matches the definition above. Both are called "spam filters," and from the outside they can look identical, which is exactly why the label is a poor guide.
A sorting table for everyday software
The table applies the definition to common examples. It is a simplified orientation, and the last column reflects a plain reading of the OECD wording, not a legal ruling.
| Example | Who defines the behavior | Learns from data | Fits the AI definition |
|---|---|---|---|
| Pocket calculator | Fixed arithmetic rules | No | No |
| Spreadsheet formula | The person who wrote it | No | No |
| Thermostat with a set weekly schedule | The homeowner | No | No |
| Email filter using a fixed phrase list | The developer | No | No |
| Email filter trained on labeled messages | Patterns in the training data | Yes | Yes |
| Speech-to-text on a phone | A model trained on recorded speech | Yes | Yes |
| Chatbot that writes new text | A model trained on large text collections | Yes | Yes |
| Simple trend line fitted to sales figures | A statistical method with few parameters | Yes, in a narrow sense | Borderline |
Where the edges blur
The last row shows why disputes about the label never fully disappear. Fitting a line to data is statistics, taught long before the term "artificial intelligence" became common, yet it does estimate outputs from input. Whether it counts depends on how broadly a reader reads "infers." Rule-based expert systems and search programs from the early decades of AI research raise a similar question: they were studied as AI, and they do not learn from data.
The OECD wording covers a lot of ground, and regulators and standards bodies choose definitions for their own purposes. For a reader, the useful move is to ask what a specific product does, not to rely on whether the word is attached to it.
Four questions that reveal what a product does
- What data was the system built or tuned on, and who labeled it?
- What exactly does it output: a score, a ranking, generated text, a decision?
- Does its behavior change after deployment, through retraining or feedback, and who approves those changes?
- How was it tested, and on which kinds of cases is it known to fail?
A vendor that cannot answer the first two usually has either a plain rule engine or a model it does not control well. Neither is automatically bad. A rule list is predictable and easy to audit, which suits many tasks better than a learned model does.
Common mix-ups
Automation is not the same as AI. A scheduled job that sends a report every Monday is automated, and it is fully specified in advance. An algorithm is also a broader idea: a recipe of steps, which can be fixed or learned. A robot is a body; it may or may not contain a learned model. And an AI system producing fluent output does not show that it understands the content in the way a person does, a point covered in more detail on the page about what AI is not.
FAQ
Is a program with if-then rules artificial intelligence?
Under the OECD wording, a system whose behavior is fully fixed by rules a person wrote does not infer anything from input, so it is ordinary software. Historians of the field do count some rule-based systems in AI research, so the answer depends on which definition is used.
Does AI always need a lot of data?
Systems that learn from data need examples, and the amount varies widely by task. Small statistical models can work with modest datasets, while large language models are trained on very large text collections.
Can the same product mix AI and ordinary software?
Yes, and most do. A banking app may use fixed rules to check a password and a learned model to flag unusual card activity. Each component should be judged on its own.
Why does the definition matter?
Standards and rules often apply only to systems that meet a stated definition, and buyers pay differently for something marketed as AI. A clear test keeps the label from standing in for an explanation.
Related reading
- Common misconceptions about AI
- A beginner's learning path for artificial intelligence
- How computer vision learns from images
- Where Canada stands on AI regulation
More in AI Basics.
Last verified: October 2026
Sources
- OECD definition of an AI system (revised 2023), OECD AI Policy Observatory, 2023
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, National Institute of Standards and Technology, January 2023