Photo Shutterstock
Artificial intelligence has the potential to impact almost every area of life, and in this second article we explore the technology behind generative AI and how it can be used.
You may have heard about generative AI in the past year or so since the free version of ChatGPT was made available to the public.
Generative AI (GenAI) is at the forefront of today's attention not only as an advancement in artificial intelligence, but also as a transformative technology with the power to reshape how businesses, governments, and humans innovate, operate, create, and deliver value.
GenAI is an AI technique that learns from real-world data to generate new content – text, images, audio, code, video, tabular data, and more – that has similar characteristics to the data used for training.
The three main applications of generative AI are Large Scale Language Models (LLMs), Synthetic Data, and Digital Twins. In this article, we'll explore these different models and how they work, building on previous articles that covered other types of artificial intelligence.
Large-scale language models
Large-scale language models (LLMs) analyze huge amounts of text (millions or even billions of words) and train them to recognize word relationships, allowing them to generate human-like text.
For example, how would you finish the sentence, “After he messed up, the boy was in the doghouse”? Did you say “dog treats” or did you say “hot summer days”? No, of course not. I said “doghouse” because it's a common phrase you've heard many times. “He was in the doghouse” is commonly associated with making a mistake, failing, or behaving badly.
The inventors of LLM took the concept of natural language processing, which scans each word in a sentence and translates it sequentially, and instead reads the entire sentence at once and analyzes all of its parts instead of individual words, which provides better context.
So LLMs track and learn relationships within sequential texts, texts where placement matters, and respond to prompts using similar texts.
When you ask an LLM like ChatGPT or Copilot to write a sonnet about your favorite dog and its love for slippers, the system appears to get creative and generate entirely new ideas. In reality, the system uses millions or billions of pieces of data to sequence the next best word or group of words. With unlimited computing power and incredible speed, tasks that would take humans hours, weeks, or even years can be performed in seconds.
One of the most common uses of LLM is as an advanced search engine – not only does it pull information from every corner of the internet and internal company databases, but it also brings information together in ways that are often relevant and easy to consume. But keep in mind that LLM alone cannot solve a business task; the key is to integrate LLM into your decision-making process. Combining LLM with other computer and human systems – layering it on top of other AI algorithms or making it part of your investigators' research process – accelerates value for your organization.
A great example of this is using NLP and LLM to process large volumes of public comments: what would take hours to read and synthesize public comments can be done in a fraction of a second, as the NLP model ingests the comments and organizes them into groups, and the LLM interprets the results for easy processing.
Synthetic Data
Synthetic data is man-made data that accurately mimics real-world data. This on-demand, automated data is generated by algorithms or rules, as opposed to traditional data sets collected from the real world.
Synthetic data reproduces the same statistical properties, probabilities, patterns, and characteristics as the real data sets it is trained on, and has been shown to be as much as 99% statistically valid.
Governments can use synthetic data for a variety of purposes, including research, testing, and analysis, without violating privacy regulations or revealing sensitive information.
There are three main reasons why governments might want to use synthetic data.
- Synthetic data can supplement your data set when real data is insufficient. For example, say you want to test a road improvement concept but you only have a few months of traffic data. Creating synthetic data, such as simulated traffic flows, can help you test potential road improvements. In such cases, creating artificial data can help you train and test models, run what-if scenarios, and identify optimizations.
- Synthetic data can also be used to protect sensitive data. Synthetic data can mimic real datasets without including any personally identifiable information. Therefore, synthetic data can be created to train and test systems that process health records, student records, or tax information. This allows governments to leverage the benefits of data-driven decision-making while respecting individuals' privacy and data protection rights.
- When real-world data is dirty or has gaps, synthetic data can be used to supplement a dataset, improving its usefulness.
Digital Twin
A digital twin is a virtual model of a real-world physical object or system – for example, a government might build a digital twin of its road network, supply chain, or financial system.
Digital twins can be used to predict real-world impacts such as highway accidents, supply shortages, economic disasters, etc. What-if analysis allows you to virtually test the real-world impact of certain decisions.
Digital twins use a combination of different data as inputs: historical data, real-time data, synthetic data, system feedback loop data, etc. These inputs can be processed in batches or in real time.
Digital twins of the physical environment may be the most obvious, but digital twins can also be used to test impacts within policy frameworks – for example, governments could use digital twins to test the impact of tax changes before making them.
The rise of AI: What to do with it
Many people wonder why AI is only starting to take hold now, when it has been around since the 1950s. It comes down to the maturity of three things: computing power, data, and analytical models. If you have ever programmed, you remember how long it took to run a program that used a lot of data. In the 1980s, it was common to run a program before finishing work, hoping it would run error-free overnight. Today, computers are so capable of processing data that the same program runs in seconds.
Now, with so much data, especially in the public sector, computers have much more to process. All of this data is the fuel that AI needs to deliver results. And finally, there are advanced analytical models that emulate tasks previously performed by humans.
Despite advances in machine learning, deep learning, and generative AI, there are still clear differences between what humans and machines excel at. Humans use common sense, intuition, creativity, empathy, and versatility. Machines automate tasks that humans can perform when they have all the time in the world and can't tire themselves: not just crunching large amounts of data, but learning from it, performing complex calculations, and performing tasks without human intervention or assistance.
Advances in generative AI suggest that machines will soon be able to rule the world, but let’s remember how machines generate output. In large-scale language models, given a prompt, a machine selects the most likely next word, line of code, or image. Some marvel at the creativity of LLMs. But the creativity came when humans fed information into the dataset. A machine doesn’t understand the meaning of heartbreak, but it finishes the sentence “When he left me, he broke my…” with “heart” because millions of people have written it that way. Large-scale language models are operating without common sense and true intuition. Because this approach is effectively guesswork, LLMs simply assemble and present what humans have previously rendered.
Many people wonder if AI will take over their jobs. Of course, we can't predict the future, but it's clear that AI is being deployed to complement human jobs. Humans understand patterns and trends that emerge from large data sets. Humans interpret the results of complex calculations. Humans evaluate the impact of simulated scenarios. Humans handle exceptional cases that can't be automated.
In other words, we expect machines to take over the tedious and time-consuming tasks. It's too boring To help humans optimize how they spend their time and improve their results.
AI technology is very powerful and will become even more powerful in the future. But like any great power, it also comes with great responsibility. Advances in AI technology have far-reaching impacts and implications. Therefore, a reliable and ethical approach is needed when setting strategies and guidelines for the use of AI and generative AI. Human-centrism, the interests of our people, and doing the right thing must come first.
Consider these six principles when adopting an AI or GenAI solution:
- Human-centricity – Remember that AI solutions are developed and implemented to advance people’s well-being.
- Inclusion – Ensuring that systems are created by and for people from different backgrounds, with diverse perspectives and experiences.
- Accountability – We proactively work to identify and stop negative impacts.
- Transparency – Provide a “clear box” rather than a “black box.” In other words, be open about how your AI system works, why it produces certain results, and what data it uses.
- Robustness (also known as stable or resilient AI) – Implementing AI systems that can function effectively even when operating in unexpected or changing environments.
- Privacy and security – Protecting the identities of people who are the subjects of AI systems.
These six principles can be difficult to understand when working with AI or generative AI. When using AI or generative AI, it is critical that humans make good decisions about inputs, prompts, and outputs. “Keeping humans in the loop” allows governments to leverage the strengths of AI while limiting risks. It doesn't diminish the role of humans in government work, it transforms and improves it.
But with the potential for AI to take over many jobs, it's no wonder people worry that AI will take over their jobs. Will AI take over your job? It's more likely that someone who knows how to use AI will take over your job.
This is the second of a two-part article explaining how generative AI works, enabling civil servants to understand how the system works and how to deploy it. Read the first article here: Introduction to Artificial Intelligence – From Machine Learning to Computer Vision.
To learn more about how AI can benefit your government organization, contact Jennifer Robinson, Global Public Sector Strategy Advisor, SAS. [email protected] Or visit our website at sas.com/public-sector.
