Publication
Artificial Intelligence (AI)
From Department of Public Expenditure, NDP Delivery and Reform
Published on
Last updated on
From Department of Public Expenditure, NDP Delivery and Reform
Published on
Last updated on
In 2019 the European Commission’s High-Level Expert group on Artificial Intelligence defined AI as “systems that display intelligent behaviour by analysing their environment and taking actions – with some degree of autonomy – to achieve specific goals."
Broadly speaking there are three different types of AI:
1. Predictive systems are generally trained on a specific dataset and learn associations between the variables used in the training data and how these relate to a target variable. Examples could be systems that are trained to predict the price of a second hand car based on vehicle details such as make, model, and year of manufacture, and the typical price second hand cars sell for.
2. Classification systems are also trained on specific datasets with the aim of being able to categorise new items presented to the model. An example could be a system trained on photographs of different types of birds being able to categorise a newly presented image of a bird.
3. Generative systems are different from both predictive and classification systems in that they generate new content from the training data. These models can be used to generate first drafts of written material, images, etc.
General AI also known as “strong AI” usually refers to systems that mimic human like intelligence.
In contrast, Narrow AI focuses on performing a very specific tasks intelligently. Examples of Narrow AI used in public service are available via the links below.
• Computer vision is the ability to process and synthesize visual data. This includes being able to detect and classify objects based on image analysis. For example, the Imperial College London in the UK is leading on a project using AI to detect cancer from MRI scans.
• Another example is facial recognition. Security forces around the world are using this to detect people of interest from images captured from CCTV or public surveillance cameras. See INTERPOL's information page about the use of facial recognition.
• Natural Language Processing (NLP) refers to the ability interpret human language and can include text analysis or translation. Coupled with the field of Speech Recognition which focuses on the ability to understand speech and if required, translate this into text, NLP systems can be linked to speech recognition systems to deliver intelligent chatbots. An example from Smart Dublin describes how comments made by citizens about Dublin City on social media sites can give insights about their views in civic issues.
• Anomaly detection is an example of Machine Learning. One method of model development named Supervised Learning uses labelled data to learn what valid and fraudulent claims looks like and then applies this knowledge to future cases in an effort to detect fraud. Another method of model development - Unsupervised Learning - applies algorithms to large amounts of data in order to detect any underlying structures that may exist. Data items that seem to be outliers can be referred for further examination. See Using Data Mining to Detect Health Care Fraud and Abuse: A Review of Literature from the National Library of Medicine in the US.
While there is great potential for AI to deliver on the new technologies outlined above, it is not without ethical and legal issues. Essentially AI is a combination of algorithms and data, and there is potential for bias and human error in both.
For example, a data set that is inaccurately labelled or does not fully represent the domain it is supposed to represent, can result in a biased solution being rolled out. Likewise a bug in a human developed Machine Learning (ML) algorithm can cause sub-optimal operation.
In addition, another aspect of AI known as Deep Learning, uses Artificial Neural Networks (ANN) to come to its own understanding of sometimes unstructured data. This can make determining where a fault lies very difficult especially in AI applications where there is an integrated suite of ANN’s.
EU bodies are actively trying to come to terms with the above in order to protect citizens from the adverse effects of errant AI. This is especially important in high-risk critical systems such as medical applications or autonomous vehicles.
With the publication of the Government’s National AI strategy AI - Here for Good there is an expectation from Government that Civil and Public Service organisations will embrace AI in both an innovative and responsible way to enhance the delivery of current public services and deliver new ones.
The Government has instructed that all AI tools used by the Irish Public Service should comply with seven requirements for ethical AI that have been developed by the European Commission’s High Level Expert Group on AI in their Ethics Guidelines for Trustworthy AI document.
The seven requirements for Ethical AI, as set out in these Guidelines, fall under the headings of:
• Human agency and oversight
• Technical Robustness and safety
• Privacy and data governance
• Transparency
• Diversity, non-discrimination and fairness
• Societal and environmental well-being
• Accountability
Interim Guidelines for Use of AI have also been developed by a cross-Department Working Group on Trustworthy AI in the Public Service. This document sets out interim guidelines and issues for consideration for Public Service organisations when considering the use of AI tools. It outlines Government’s commitment to the ethical use of AI. It also encourages risk assessments be carried out and outlines the safeguards and considerations that are relevant when exploring the use of AI tools.
Other Government supports available to Public Service organisations include a direct drawdown procurement framework for Robotic Process Automation (RPA), funding support for AI projects (including via the recently launched Public Service Innovation Fund 2024 ), as well as learning and development interventions for public servants at no cost to public service bodies.
In early 2018, the Irish Revenue Commissioners initiated a pilot project to examine if AI-based Natural Language Processing (NLP) technologies could be used to deliver an improved customer service, reduce costs and increase efficiencies.
This resulted in the implementation of a Virtual Digital Agent (VDA) or voicebot designed to focus on a subset of calls from the Irish taxpayer relating to tax clearance. A high proportion of these calls are reasonably repetitive and the knowledge required to provide suitable responses and positive experiences for customers is well understood.
A suite of integrated technologies were used to convert customer speech to text, understand the text using NLP so that a response could be formulated, and then convert this response back to speech so the customer could hear the answer.
The Irish Revenue demonstrated that voicebot technology can offer a fully automated service, providing an efficient, effective experience for customers. Below is a summary of some of the results:
• Up to 50% of calls were handled from start to finish by the voicebot.
• 70% of first-time applicants engaged with the voicebot when submitting their application.
• 75% of tax clearance holders were able to retrieve their tax clearance access number (TCAN).
• Only 10% of calls were transferred due to failure to understand.
The success of the project has encouraged the Irish Revenue Commissioners to deploy this technology in future service offerings.