Definitions of AI

What is AI

Search the web for ‘the definition of Artificial Intelligence’ and you will be presented with numerous slight variations of the definition, with the common words being "technology or computing" and "human intelligence". ICAEW’s Artificial intelligence and the future of accountancy defines AI as "machines or computers that can simulate human intelligence", implying that computers can perform some tasks without human involvement.

To understand how AI works, it helps to compare AI to a simple algorithm – an algorithm takes an input, adds a set of logical pre-defined rules and generates an output. This is a typical rules-based system. These types of data analytics tools, built using simple algorithms, require exact information, plus humans capable of interpreting the datasets.

Unlike a simple algorithm, AI uses an iterative process considering both inputs and outputs to "learn" patterns in the data and produce conclusions when given new inputs, without the need for additional specific instructions. It simulates learning through trial and error until it arrives at an acceptable level of precision and accuracy.  AI systems can learn how to do tasks (such as reading, writing, generating content and analysing) by processing massive amounts of data and identifying patterns to make predictions and decisions. When there’s vast amounts of complex data, AI tools can produce meaningful conclusions or outputs faster than humans, but the quality of the prompts are essential to maximise the results from the AI tools. It is important to note that the outputs still require human review to test for accuracy. 

However, it’s important to remember that Generative AI (Gen AI) tools are more likely to make meaningful conclusions when the input and output is rule based. They are prone to hallucinations when it comes to opinion-based outputs that involve reasoning and emotion. For more information on risks when utilising Gen AI tools, see the managing the risk of using AI section. 

This ICAEW article delves deeper into the definition of AI with simple to understand examples.

Gen AI is developing at pace and is the AI technology most used by the wider public. However, to fully understand the concept of Gen AI it helps to first understand the difference between a simple AI tool as well as machine and deep learning. 

Artificial Intelligence

AI is the broad term which covers any type of program that can carry out tasks that require human intelligence such as reasoning, acting, and adapting. Machine learning and deep learning are both sub-fields of AI and are often used interchangeably, but it is important to recognise the difference between them.

Machine Learning

Machine Learning (ML) is a subset of AI based on algorithms (set of rules) that identify patterns in data sets, which then present a possible interpretation. The tools’ performance improves as the algorithms are exposed to more data and a feedback loop. They are able to learn, predict and make recommendations. There are four main techniques for simulating learning:

• supervised;
• unsupervised;
• semi-supervised;
• and reinforcement.

Each technique has its own unique strengths when it comes to using it during an M&A project. One example of machine learning being used during an M&A transaction is the processing of legal documents in a data room to identify specific contact clauses, or potential legal risks that may have a financial cost implication.

• More information on machine learning

Deep Learning

Deep Learning (DL) is an advanced form of machine learning which replicates how the human brain works using artificial neural networks (containing hidden layers where data is processed) that adapt and learn from vast amounts of complex data. It is often referred to as a ‘black box’ because unlike algorithms it is not always possible to see their workings. This is something to be considered when reviewing their outputs.

• More information on deep learning

An example of a deep learning tool is predictive analytics, which attempts to predict and forecast future outcomes by mining large data sets. Businesses can identify patterns in the data using these tools and these insights can be used to exploit potential opportunities, as well as identify potential risks that need to be managed or further mitigated (eg, client churn predictions can support sales teams to identify dissatisfied clients sooner, enabling them to initiate timely conversations to retain them). In a deals context, predictive analytics tools can be used to assess the reliability of revenue and cost assumptions applied by the management team to their projections.  
However, it is important to remember that the technology can only generate content with human interaction, in contrast to the creation of content which implies the use of initiative and unique deep creativity. 

Generative AI

While machine learning focuses on learning from data to make predictions or suggest a decision, Gen AI can generate original, new content—such as text, images, video, audio or software code—in response to a user’s prompt or request using natural language rather than computer code. It relies on deep learning models. Well known examples of natural language models include Open AI’s ChatGPT and Microsoft’s Copilot. 

However, it’s important to remember that Generative AI (Gen AI) tools are more likely to make meaningful conclusions when the input and output is rule based. They are prone to hallucinations when it comes to opinion-based outputs that involve reasoning and emotion. For more information on risks when utilising Gen AI tools, see the managing the risk of using AI section. 

• Visit ICAEW’s Generative AI hub for more details and relevant content, including case studies 

One example of Gen AI technology is language models. 

Language models

Natural language processing (a form of Deep Learning) involves teaching computers to understand and produce written and spoken language in a similar manner to humans. They are used in practical applications that require interaction with humans in natural language, for example a virtual assistant or chatbot. A subfield of NLP is the area of large language models (LLMs), which relies on more complex, deep learning models and these are trained on extensive datasets. 

Large language models (LLMs) are specialised Gen AI tools that generate human-like text, such as writing articles, composing poetry, or partaking in extended conversations. 

Notable LLMs include ChatGPT from OpenAI, LLaMA by Meta Technology Corporation, Gemini developed by Google and Claude from Anthropic. 

In a deals context, large language models could be used to assist in the drafting of due diligence reports, Information Memorandums and sales and purchase agreements. They can also be used to generate multi-media content such as visualisations and infographics to be used in deliverables such as Information Memorandums and due diligence reports.

Ensemble models

Ensemble models are the models we are most likely to interact with, where a series of different models are “stacked” onto one another generally to improve the performance, or reduce the risk of the core model being used. For example, if you were to use a Deep Learning model to identify dissatisfied clients, you may also want a classification algorithm to assign a confidence or risk scoring to prioritise the most dissatisfied clients.

The widely available large language models are an example of ensemble models (where you have one model for processing written language, and another for generating the output).