Identifying AI
Guidance to help you determine whether your system contains artificial intelligence (AI).
Purpose
To provide general guidance to NSW government public servants involved in the lifecycle of digital solutions—from design and sourcing to procurement, selection, and management—seeking to understand if these solutions may contain artificial intelligence (AI). It is not an exhaustive source of considerations upon which to determine the existence of AI or the applicability of the AIAF.
By gaining confidence in this area, you are able to better identify risks and ensure compliance with assurance frameworks in NSW Government, including the NSW AI Assessment Framework (AIAF).
How to use
After reviewing the ‘What is AI?’ section to become familiar with AI and its characteristics, undertake the questionnaire to help identify whether you are using AI.
This information serves as a guide only. Ultimately, NSW Government agencies are responsible for determining whether AI is being utilised, and therefor whether the AIAF must be applied. Any decision not to apply the AIAF must be documented in your agency's record management system.
What is AI?
“Artificial Intelligence (AI) is the ability of a computer system to perform tasks that would normally require human intelligence, such as learning, reasoning, and making decisions.”
NSW AI Assessment Framework 2024
There has yet to be a universally accepted technical or legal definition of AI. For this reason, it can be difficult to objectively identify and agree that AI is being used within a system. It is suggested that where there is the possibility of AI use, the mandatory AIAF be applied.
Characteristics
While not all AI systems demonstrate the same characteristics, it is common to see the following exhibited.
AI systems make decisions with uncertainty and are data-driven, whereas traditional systems follow predetermined rules and produce the same result for identical inputs.
AI systems learn and improve over time, while traditional systems need manual updates to change behaviour.
AI systems rely on data to make predictions or decisions, while traditional systems execute predefined rules and workflows without adapting.
AI systems recognise complex patterns within data, while traditional systems perform tasks based on straightforward logic.
AI systems adapt their decisions based on new data and changing conditions, while traditional systems follow fixed workflows that do not evolve.
Questionnaire
Identifying AI can be complex, especially when AI is integrated as a feature within applications. This guide provides a foundational approach, but professional advice should be sought within your Agency.
Questions are divided into two main sections: "Non-AI use cases" and "AI use cases". Starting with the "Non-AI use cases" section can help identify systems that don’t involve AI. If your system doesn’t meet all the criteria in that section, proceed to "AI use cases" to further assess whether AI is involved.
Non-AI use cases
You’re unlikely to be using AI if you answer "Yes" to all of the following questions:
| Question | Example |
|---|---|
| 1. Does your system operate on predefined rules without learning or adapting from data? | Non-AI example: A basic FAQ chatbot that always provides the same response to specific questions without changing or learning over time |
| 2. Does your system only perform simple data storage and retrieval without analysing patterns, making predictions or generating insights? | Non-AI example: A customer database that stores and retrieves information without producing new insights or analysis. |
| 3. Does your system always produce the same result for the same input? | Non-AI example: An automated payroll system that calculates employee salaries based on fixed formulas, always producing the same output for identical inputs each pay cycle. |
| 4. Does your system perform tasks that don’t require understanding complex data formats such as text, images, or audio? | Non-AI example: A file management system that only stores and organises files without analysing or interpreting their content. |
| 5. Does your system follow predefined workflows without adjusting automatically based on context or data? | Non-AI example: An automated email responder that sends the same response to every inquiry, based on a fixed template. |
| 6. Does your system only allow interaction through standard forms, menus and workflows, without the ability to use natural language as a question / command? | Non-AI example: An online form that collects user information through a series of fixed steps without conversational capabilities. |
| 7. Does your system require manual updates to change behaviour, rather than automatically adapting? | Non-AI example: Accounting software that needs manual updates for new tax laws, instead of automatically adjusting based on changes. |
| 8. Does your system handle straightforward tasks that don’t involve pattern recognition, prediction, or complex analysis? | Non-AI example: An electronic voting system that records votes but does not analyse voting patterns or predict outcomes. |
AI use cases
If you answer "Yes" to any of the following questions, it is likely that you are using AI. Please note that the system must go beyond simple rule-based processing or statistical models:
| Question | Example |
|---|---|
| 1. Does the system make predictions, suggestions, or decisions based on data without explicit programming and/or through deep learning techniques? | A fraud detection system that identifies suspicious transactions based on historical data patterns and refines its predictions over time. |
| 2. Does the system change its performance or accuracy over time without manual updates? | A benefits eligibility system that learns from past cases and gets better at determining who qualifies for assistance through ongoing analysis of past interactions and outcomes. |
| 3. Can the system analyse and generate insights from complex data formats like images, audio, or documents? | A medical system that transcribes voice recordings into text and identifies potential health risks based on the audio data over time |
| 4. Does the system automatically make decisions or recommendations that go beyond rule-based processes and incorporate dynamic learning or pattern recognition? | A traffic management system that adjusts signals in real-time based on evolving traffic patterns, learning from past data and adjusting itself to new conditions. |
| 5. Can the system engage in human-like interactions, such as conversations via voice or text, that adapt over time? | A voice assistant on a government website that continuously refines its responses based on user interactions and feedback, improving its accuracy in answering questions. |
| 6. Does the system adapt its responses based on user interactions or feedback, improving its functionality over time? | A government portal that suggests services based on past user interactions and improves its recommendations as more data is gathered over time. |
| 7. Is the system capable of performing tasks that typically require human judgment or decision-making, and does it improve its judgment based on new data? | A healthcare system that assists doctors by analysing large datasets and refining its diagnostic capabilities based on evolving medical trends and patient outcomes. |
What’s next?
If you believe your system may utilise AI, the AIAF must be applied to ensure safe and responsible use of AI.
If you're still unsure whether you are using AI or what your responsibilities may be, seek advice from your Agency Chief Information Officer, Digital Officer, or Data Officer.
Resources
Explore further AIAF Guidance
Feedback
Agencies are encouraged to provide feedback for any suggested improvements to this guidance to the AI Secretariat.
