What even is AI?

This blog series by Peter Impey, Consultant General Counsel at InTouch, looks at what AI is, how it’s evolving, and what it means for legal services. It outlines key developments and use cases, touches on risks, regulation and the SRA’s position, and explains InTouch’s practical, low-risk approach to bringing AI into everyday work in a law firm. You can find the full series and related articles on the InTouch blog.

The rising tide of semiconductors, computing power and software development has lifted many fields of computer science. Advances are taking place in many ‘AI’ fields. In turn, this is driving considerable change in society and business, and transforming or opening up areas of the economy.

So, what is ‘artificial intelligence’? Simply, the capability of computational systems to perform tasks typically associated with human intelligence is known as ‘artificial intelligence’. The kind of AI most relevant to legal is ‘generative’ AI. We’ll discuss what that means shortly.

But first, I think it is important to understand that there is a whole range of AI, all of which are becoming increasingly powerful and useful. The scale and impact of the technology is astonishing. Entire industries are being transformed, from the medical world to the military.

Business consulting, chip manufacture, synthetic biology, finance, legal services – all and more are being disrupted.

 Types of ‘artificial intelligence’

Artificial Intelligence has many categories. The recent ‘rise’ of AI includes:

• Image recognition

• Synthetic data creation

• Federated machine learning

• Causal AI

• Voice recognition and NLP (natural language processing)

• Sentiment analysis

• Forecasting, i.e. predictive analytics

• Generative AI (LLMs)

These are distinct areas of development.

When people talk about “AI”, they’re often lumping together lots of quite different technologies. It’s helpful to separate them out a bit and see what each one actually does in practice. Here are some further descriptions:

 Image recognition

Deep learning algorithms can detect faces, determine gender, mood, levels of concentration, and even signs of lying.

• Computer image recognition is thousands of times faster, with a wider field of vision (even 360°), huge powers of zoom and microscopic capabilities, and can view areas in the light spectrum that are invisible to humans.

• Specific advances in image recognition include CNNs – convolutional neural networks. Complicated overlays of filters work by performing a mathematical operation called convolution on the image, which involves sliding the filter over the image and multiplying the values in the filter by the corresponding pixel values in the image. This produces a feature map that highlights specific patterns or features in the image.

• In the medical world, image recognition can help detect breast or lung cancers, for example. Advanced medical imaging modalities such as PET/CT hybrid, three-dimensional ultrasound computed tomography (3D USCT), and simultaneous PET/MRI give high resolution, better reliability, and safety to diagnose, treat, and manage complex patient abnormalities.

 Synthetic data creation

AI that generates synthetic data based on patterns and relationships learned from actual data. This ability to generate synthetic data has numerous applications, from creating realistic virtual environments for training and simulation to generating new data for machine learning models.

 Federated machine learning

Describes a way to train AI models without anyone seeing or touching your data, offering a way to unlock information to feed new AI applications. It is a decentralised approach to training machine learning models. It doesn't require an exchange of data from client devices to global servers. Instead, the raw data on edge devices is used to train the model locally, increasing data privacy.

 Causal AI

Can explain causes and effects, for example the underlying causes and effects of events or behaviours, and can integrate into decision-making. It is all about understanding the underlying relationships between variables in data and using this information to make predictions and decisions.

 Voice recognition + NLP (natural language processing)

Systems that turn spoken language into text and then analyse it for meaning, enabling things like dictation, call transcription, chatbots and voice-driven assistants.

 Sentiment analysis

The process of recognising emotions expressed in text. AI comprehends the tone of a statement, as opposed to merely recognising whether particular words within a group of text have a negative or positive connotation. Applied to emails, for example, this can reveal how employees feel about different dimensions, from work/life balance to compensation and benefits.

 Forecasting i.e. predictive analytics

Weather, retail buying patterns, sports results, insurance, financial modelling, etc.

 Generative (LLM)

‘Large language models’ are known as LLMs. These use a text-based “next word” approach to generating text based on prompts. They are hugely powerful and one of the more interesting types of AI in the context of legal services. Examples include ChatGPT, Google’s Bard, Claude AI. For more information, visit the World Intellectual Property Organization (WIPO).

‘The cat sat on the…’

Did you guess the next word?! A “Generative Pre-trained Transformer” (GPT) is a large language model (LLM) trained to do just that.

As you’ve no doubt seen, an LLM will read your input text, ‘understand’ the prompt and predict the next ‘token’ (word, letter) in a text string answer the model generates. This is known as ‘generative AI’ and ChatGPT from OpenAI is just one example.

There are many LLMs and they all simply generate text based on the model’s best guess of what the next word should be. Their power comes from the model’s training and reasoning capabilities to determine context and decide what token to place next.

Even in early releases the LLMs were trained on large parts of the internet and a huge corpus of other text materials. This vast training set of practically all text-based information provides the LLM with the ability to create human-like text responses, based on the ‘prompt’ you provide.

The models have billions of parameters and can operate in time-scales that appear almost instantaneous to the user. Considering that the law is text-based and most legal services involve text, LLMs are clearly an incredibly useful form of AI in the legal world.

A key risk everyone worries about is that the LLM ‘hallucinates’ and simply invents text that looks compelling and realistic but is unfortunately not grounded in reality.

Several high-profile instances of fictional case law or inappropriate legal arguments have already created problems for lawyers relying on LLM-generated materials. Just Google “lawyer caught using AI generated fake citations”!

GPT models are improving

Huge investment is taking place to enable increasing levels of sophistication in the models and the reasoning behind the answers they provide. Significant investment is also improving the capabilities of the computing hardware.

The improvements are dramatic, with models running on fewer parameters and using less power. In combination, this means that the models are increasingly capable and cost-efficient.

There is also an increasing corpus of specialised training materials, for example all litigation cases with filings containing legal arguments and the actual judgements from courts in the US.

In addition, access to live information on the internet or specific repositories of documentation, such as a company’s policies, results in reduced hallucination, greater contextual accuracy and overall usefulness.

A highly illustrative metric for understanding the improvement and relevance to legal services is the Uniform Bar Exam. In December 2022, GPT-3.5 was already able to pass the exam better than 10% of test takers.

In only three months, by March 2023, the GPT-4 release beat 90% of people taking the exam designed to test knowledge and skills that every American lawyer should be able to demonstrate prior to becoming licensed to practise law.

That’s worth repeating. In Q1 2023, ChatGPT was able to pass the Bar exam better than ninety percent of people aiming to practise law in the United States.

And since then, there have been further huge improvements, becoming yet one more driver in the current paradigm shift in legal services.

Quill to Keyboard

For centuries, contracts, deeds and laws were written on vellum – stretched animal skin – in English, French and Latin, in a long era of quill and ink. A major paradigm shift took place when we moved from the quill to the keyboard, first through manual typewriters.

Later, the interface moved via the computer keyboard onto the screens of personal computers. This second stage brought another huge change: the collapse of the cost of replication.

In digital form, copying a document is instantaneous and virtually cost-free. That has also contributed to the shifting paradigm, particularly as the internet has collapsed much of the cost and friction of distribution.

Keyboard to Cloud

I describe what we are witnessing in the current legal paradigm shift as the move from Keyboard to Cloud. The new normal is cloud-based systems: data, digital workflows and voice-activated AI delivered via the internet “cloud” rather than installed on a single machine.

For legal services, this means matter files that can be accessed securely from anywhere, systems that are updated continuously, and AI tools that sit alongside everyday work. In other words, the keyboard is no longer the end point – it is a gateway into a much larger, connected platform.

 Next in the series, we’ll dig into the “so what?” of generative AI in legal services — and what it might mean in practice for firms and their clients.

This series of blogs about AI is written by Peter Impey, General Counsel for InTouch. If you’re exploring how to use AI in your firm, you can find the full series and related articles on the InTouch blog.