Section 4.5 AI Algorithm Development

Many tools people use every day include both predictable steps and best-guess behavior. A calculator performing arithmetic produces the same output every time for the same input. Statistical AI systems (like spam filters and chatbots) use probability to choose what output is most likely under uncertainty, which means they can still make mistakes even when they behave consistently. Understanding these differences helps users decide when results should be checked or verified.

A deterministic algorithm produces the same output every time it is given the same input, so its behavior is predictable and repeatable.

A probabilistic algorithm makes decisions based on likelihoods rather than fixed rules. Because it models uncertainty, it may produce outputs that include a confidence score or probability estimate. For a given input, the system will often produce the same result when the model and decision threshold stay the same, but the outcome can change if the threshold is adjusted or the model is updated.

Probabilistic algorithms are often used for problems where inputs are incomplete, variable, or ambiguous and there is no single clear rule-based solution, such as recognizing objects in images, interpreting speech, filtering spam, or generating text. Because these systems make best-guess decisions, programmers evaluate them by testing performance across many examples and checking how often outputs are accurate, useful, and reliable for the intended context.

Testing deterministic and probabilistic algorithms requires different approaches. Deterministic algorithms can be checked with fixed expected outputs, while probabilistic algorithms are evaluated by testing performance across many examples and checking whether outputs meet success criteria for the intended context.

AI systems can produce outputs using different approaches, such as rule-based systems and models trained on data. Each approach has different strengths and limitations, so the best choice depends on the task, available data, and the need for transparency.

Rule-based systems use explicit rules created by people. They are predictable and easier to test, but can struggle with complex real-world situations that are variable, ambiguous, or hard to capture with rules, such as identifying spam messages that change over time. A spam filter that relies only on fixed keyword rules may miss new scams or incorrectly flag legitimate messages because the same words can appear in both spam and non-spam messages.

Some AI models use mathematical relationships to estimate likelihoods and make a best guess when information is uncertain or incomplete. They can be transparent and consistent, but their accuracy depends on whether the model’s assumptions fit the real situation. A weather forecast model might estimate the chance of rain based on pressure, humidity, and past patterns.

Machine learning models learn patterns from training data to make predictions, classify inputs, or generate outputs. They can perform well on complex pattern-based tasks, but results depend heavily on the quality and representativeness of the training data, and it may be difficult to explain why a particular output was produced. For example, an image classifier trained mostly on bright, high-quality photos may perform poorly on blurry or low-light images.

Choosing between rule-based systems and AI models trained on data depends on the task and constraints, such as the type of output needed, acceptable error rate, speed, cost, and whether people need a clear explanation of how the result was produced.

People interact with AI-enabled features every day, such as autocorrect, recommendations, spam filters, and face unlock. These tools produce outputs in different ways: some follow fixed rules, and others use models trained on data to estimate the most likely result. Rule-driven tools usually produce more predictable outputs. Model-driven tools may produce outputs that are uncertain, context-dependent, or affected by changes to data, settings, or the model itself.