Deterministic (algorithm/model)

An algorithm/model that, given the same inputs, always produces the same outputs.

Source: AIEOG AI Lexicon (Feb 2026), adapted from CSRC Glossary

A deterministic algorithm or model is one that is perfectly predictable: if you give it the same inputs today, tomorrow, and a year from now, it will produce the exact same output every time. There is no randomness, no variation, and no ambiguity in the process.

This stands in contrast to many AI models, particularly those used in deep learning and generative AI, which may incorporate elements of randomness in their operation. A generative AI model asked the same question twice might produce slightly different responses each time because it samples from probability distributions during its output generation process.

The distinction between deterministic and non-deterministic behavior has significant implications for governance, testing, and compliance. Deterministic systems are easier to validate because their behavior is fully reproducible. Non-deterministic systems require different validation approaches because you cannot guarantee the same output for the same input.

Traditional rule-based systems are typically deterministic. Most machine learning models, once trained and deployed with fixed parameters, are also deterministic in their inference (prediction) phase, even if the training process involved randomness. However, some deployment configurations (such as using temperature settings in LLMs) introduce non-determinism at inference time.

Determinism relates directly to several regulatory expectations:

• Reproducibility. Regulators and auditors expect institutions to be able to reproduce model outputs for examination and investigation purposes. Deterministic models satisfy this requirement by definition.
• Testing and validation. Deterministic models can be validated through direct comparison of expected vs. actual outputs. Non-deterministic models require statistical validation approaches.
• Audit trails. When an institution needs to demonstrate why a specific decision was made (a loan denial, a SAR filing), deterministic models provide a clear, reproducible answer.
• Adverse action compliance. If a model produces different outputs for the same applicant at different times, providing consistent adverse action reasons becomes challenging.

1. Understand the determinism of each deployed model. Document whether each AI system produces deterministic or non-deterministic outputs and what factors influence output variability.
2. Require determinism where possible. For use cases that require reproducible decisions (lending, regulatory reporting), configure AI systems for deterministic behavior.
3. Adjust validation approaches accordingly. Non-deterministic models require statistical validation that accounts for output variability.
4. Document temperature and randomness settings. For LLMs and generative AI systems, document configuration parameters that affect output determinism.
5. Consider reproducibility in audit planning. Ensure your ability to reproduce model decisions for examination and audit purposes.
6. Log outputs at decision time. For non-deterministic models, capture and store the specific output produced at the time of each decision.

Algorithm, AI model, Black box, Explainability, Validation, Output validation