Is an algorithm the same as a model?

No. The algorithm is the training or decision procedure; the model is what the training procedure produced.

Are algorithms always fair?

Procedures are exact, not fair. Fairness lives in data, objectives, thresholds and review.

Why do people fear the word?

Because it is used for whole systems whose data and incentives are invisible, and those parts deserve the scrutiny.

AI Glossary / AI and machine learning / Beginner

Algorithm

Definition

An algorithm is a precise, finite set of instructions for solving a problem or completing a task: the same steps, in the same order, producing a result from a given input. A recipe is an algorithm; so is every search, sort and training procedure in AI.

What it means

The word carries mystique it has not earned. An algorithm is simply a procedure written precisely enough that anyone, or any machine, following it gets a reliable result. Long division is an algorithm. So is a recipe, a checklist and the routing logic in your satnav.

In AI, the word covers two different things that are worth separating. Training algorithms, like gradient descent or Adam, are procedures that adjust a model until it performs well. The trained model itself is not an algorithm in the same sense; it is the product the algorithm produced.

Headlines saying the algorithm decided usually refer to the whole system: data, model, rules and incentives. The distinction matters for accountability, because each part has a different owner and a different fix.

Algorithms have measurable properties: correctness, speed and how their cost grows with input size. These properties are why some problems are easy at small scale and impossible at large scale, even with a perfect algorithm.

For a business reader, the useful instinct is to ask which algorithm means which part: the procedure that trained the system, the model it produced, or the rules wrapped around it. Each fails differently.

How it works

An algorithm takes an input, applies its steps deterministically, and produces an output or a decision. Because the steps are explicit, the same input always yields the same result, which is what makes algorithms testable.

Training algorithms work by iteration: make a prediction, measure the error, adjust the model's parameters to reduce it, repeat millions of times. The procedure is simple; the scale is what feels magical.

Efficiency analysis asks how the work grows with the input: doubling the data might double the work, quadruple it, or barely increase it, depending on the algorithm's design. This growth behavior often matters more than raw cleverness.

Around the core procedure sits the system: the data pipeline feeding it, the thresholds translating its output into decisions, and the people reviewing exceptions. Most real-world behavior lives in the wrapping, not the core.

Three practical examples

Sorting: arranging ten thousand invoices by date in a fraction of a second, using a procedure invented decades ago.
Training: adjusting a model's millions of parameters over many passes through labeled examples until error stops falling.
Matching: pairing riders and drivers by proximity and destination using a defined scoring procedure.

Think of it this way

Think of a master baker's recipe card. Follow it exactly and you get the bread, whoever you are. The card is the algorithm; the bread is the model; the bakery, with its suppliers, ovens and staff, is the system.

The limit: the card assumes the ingredients match its specification. Flour from a different mill changes the bread without any change to the card, which is exactly how data shifts change model behavior.

Worked example

untangling a blamed algorithm

Suppose a retailer's prices surge oddly one weekend and the report says the pricing algorithm raised them. The investigation separates the parts.

The optimization procedure worked correctly. The data feed had double-counted a flash sale, the model inferred falling stock, and the threshold rules allowed a larger jump than policy intended. Four different parts, four different fixes, only one of them the algorithm proper.

The post-incident review assigns an owner per part: data feed, model, thresholds, monitoring. The scenario is invented; the separation of procedure, model, data and rules is the diagnostic habit.

Next time the algorithm is blamed, the first meeting starts with the four-part question, and the fix lands in days instead of weeks.

Illustrative case study

A fictional hiring platform is criticized for biased shortlists, and the vendor defends its algorithm as neutral. An audit finds the ranking procedure indeed neutral, but the training data came from ten years of one department's biased hires.

The fix is not algorithmic at all: relabel the target, resample the data, and add an outcome review per cohort. The procedure stays identical.

The invented lesson: the algorithm is often the least biased part of the system. Data, objectives and review design carry the fairness question.

Common misconceptions and mistakes

An algorithm is an AI brain. It is a procedure; the model it trains is a product of the procedure, and the system around both decides real outcomes.
Algorithms are neutral. Procedures are exact, but their inputs, objectives and thresholds embed choices, so systems behave as designed in ways designers did not intend.
Newer algorithms are the main progress. Much practical progress comes from better data, better evaluation and better wrapping, not new procedures.

Business use

When evaluating AI claims, ask for the separation: what procedure trained it, what data fed it, what rules wrap it, and who owns each part.

In incident reviews, resist the algorithm did it framing. Walk the four parts, data, procedure, model, rules, and assign each an owner and a fix.

Check your understanding

A pricing system surges prices and the algorithm is blamed. Name the four parts the review should separate and why.

Answer: Data feed, training procedure, model and threshold rules, because each has a different owner and fix, and only one is the algorithm proper.

Related terms

Learn next

AI for Business Owners builds exactly this habit of separating system parts when judging AI: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary. [S3] NIST, AI glossary (AIRC). Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Algebraic connectivity Next: Algorithm BSTW