Artificial Intelligence
Understanding generative AI: Capabilities, risks & more
7 July 2023
Navigate the risks of using Generative AI in your reports. From unverified content to unintended biases, here's your guide to responsible AI use at work.
You're facing a tight deadline for a crucial report. You've heard whispers about ChatGPT's ability to generate instant reports, even passing professional exams. Naturally, you're curious: "Can I leverage ChatGPT to meet this deadline?" You're certainly not alone.
With ChatGPT processing over a billion queries daily, professionals and students worldwide are exploring its potential to streamline their workflows. These AI models, known as 'Generative AI', have garnered significant attention for their capacity to produce diverse content, from articles and images to code and artwork.
What is Generative AI?
"AI" stands for Artificial Intelligence, referring to computer systems designed to perform tasks typically requiring human intelligence. This involves creating algorithms and models that enable machines to understand, reason, learn, and make decisions.
Generative AI, however, marks a pivotal shift. It's a subset of AI that doesn't just analyse existing data, but actively creates new content—images, text, audio—by identifying patterns and generating novel outputs. This distinction represents a significant expansion in AI capabilities, opening doors to previously unimaginable applications.
The ability to develop sophisticated models that can understand and replicate complex data structures is truly remarkable. With its ability to create stunningly realistic images, remarkably coherent text, and even intricate musical compositions, not to mention the rapid pace of these advancements, it’s easy to understand the growing interest and applications of generative AI across various industries.
How Generative AI works
Imagine an art student diligently studying thousands of paintings to grasp the techniques of a master. This student is akin to generative AI, learning through a process similar to deep learning, where massive datasets are used to train interconnected layers of algorithms (complex neural networks) to recognise patterns in data. Just as the student uses the learned techniques to create their own original artwork, generative AI leverages these learned patterns to produce novel outputs. The student might refine their brushstrokes and experiment with colour combinations to achieve a particular effect, ensuring their work is both diverse and coherent. Similarly, generative AI employs technical concepts like latent space manipulation and probabilistic modelling —adjusting internal representations of data and predicting likely outcomes—to ensure its generated outputs are varied and consistent.
Generative AI model types
Generative AI model types refer to the specific architectures and algorithms that enable the creation of new content. Just as an art student uses various techniques and tools to create different types of art, each model type has its own unique way of learning and generating content. Here's a look at some key examples:
Generative Adversarial Networks (GANs):
GANs consists of two neural networks, a generator and a discriminator.
The generator creates new data, while the discriminator evaluates its authenticity.
This back-and-forth process refines the generator's output, leading to increasingly realistic content.
Variational Autoencoders (VAEs):
VAEs learn a compressed representation of data, allowing them to generate variations of existing content.
This is particularly useful for creating similar images or audio with slight differences.
Transformers (GPT, T5, BERT):
These models excel at understanding context. They pay close attention to different parts of the input, making them ideal for natural language tasks.
GPT models, for instance, can generate coherent and contextually relevant text by focusing on the relationships between words..
Diffusion Models:
This approach gradually adds noise to data, then learns to reverse the process, refining a noisy input into a realistic image.
This method is gaining popularity for its ability to produce high-quality images and videos.
Why are we only hearing about this now?
While the seeds of generative AI were sown in the mid-20th century, its recent explosion in popularity is a result of several converging factors.
Firstly, advancements in hardware and cloud computing have provided the immense computational power needed to train these complex models.
Secondly, open-source technologies, like Hadoop, have made scalable AI development more accessible.
Thirdly, strategic investments and heightened public awareness have fuelled a surge in research and innovation.
Furthermore, breakthroughs in machine learning and neural networks, coupled with the availability of vast datasets, have accelerated progress.
What can Generative AI do?
Generative AI offers innovative solutions and enhances productivity across a multitude of sectors. Generative AI can generate text, images, audio, and even code, making it a versatile tool for various applications and industries. Here are some industries that currently use AI:
Content Creation:
Generative AI is revolutionising content creation by generating articles, marketing copy, and creative writing. For instance, companies like Jasper AI are empowering businesses to produce compelling marketing materials, blog posts, and engaging social media content. Moreover, platforms like D-ID are enabling the creation of dynamic talking head videos, enhancing training and marketing initiatives.
Software Development:
In software development, generative AI streamlines coding processes by generating code snippets and assisting with debugging. A prime example is GitHub Copilot, developed by GitHub and OpenAI. This tool uses AI to suggest code completions and even entire functions, significantly accelerating the development workflow.
Healthcare:
Generative AI is also making strides in healthcare, generating synthetic medical images for training AI models and assisting in drug discovery. Insilico Medicine, for example, leverages AI to discover and design novel drug candidates, accelerating the process of bringing new treatments to patients.
Design:
The design industry is witnessing a transformation with generative AI, enabling the creation of product designs, architectural renderings, and personalised fashion designs. Tools like Adobe Firefly are being used to generate and manipulate images and create text effects, while companies like Worn Again Technologies are using AI to design innovative textile recycling solutions.
Finance:
In the financial sector, generative AI is enhancing financial analysis and fraud detection. Institutions like JP Morgan Chase utilise AI models to analyse vast datasets, generating insightful market trend reports and detecting anomalies in transaction patterns, thus bolstering fraud prevention efforts.
Limitations of AI
Despite its impressive capabilities, generative AI has several limitations. For example, here’s what ChatGPT told us when asked for the location of GovTech Headquarters:
This information isn’t exactly wrong, but it is inaccurate. Our HQ is actually at 10 Pasir Panjang Road, while the GovTech Hive is at the Sandcrawler instead. It's crucial to remember that AI models like ChatGPT have a knowledge cutoff, in this case, September 2021, and therefore cannot guarantee real-time accuracy. Therefore, always verify AI-generated content with human expertise and reliable sources.
This instance highlights a broader limitations of generative AI:
Reliance on Data:
Generative AI depends heavily on existing data and patterns, limiting its ability to handle novel or underrepresented scenarios.
Pattern Recognition vs. Understanding:
AI generates content by recognising patterns, lacking true comprehension of the information it processes.
Potential for Inaccuracy:
AI may produce inaccurate or outdated information, as shown in our example above.
Dangers & Risks of Generative AI
While generative AI can offer significant benefits, it's crucial to acknowledge the potential risks of AI and consequences associated with its use.
Let’s have a look:
1. Misinformation & Deepfakes
Generative AI's ability to create highly convincing fake images, videos, and audio—known as 'deepfakes'—poses a significant risk. These deepfakes can be exploited to spread disinformation, perpetrate fraud and scams, manipulate public opinion, and severely damage reputations. Alarmingly, deepfakes have contributed to a rise in scams in Singapore, with our nation ranking joint second with Cambodia for the increase in deepfake attacks in 2024. This trend underscores the heightened challenge of scam prevention in our digital landscape, particularly as deepfake technology becomes increasingly accessible.
2. Plagiarism & infringement
The use of AI raises concerns about copyright infringement and plagiarism. When trained on copyrighted material, AI can potentially reproduce substantial portions or even entire works, leading to direct plagiarism. AI-generated content can also be considered a derivative work if it is based on or substantially similar to a copyrighted original. The line between unintentional reproduction and deliberate plagiarism is blurred by AI, creating legal complexities. While there are no current Singaporean court cases that directly involve generative AI and copyright, this is an evolving legal landscape that must be monitored.
3. Unintended biases
This happens because generative AI models learn to identify and replicate patterns present in the data they are trained on. If that data reflects existing societal biases—whether conscious or unconscious—the AI will inevitably learn and amplify those biases in its output. For example, DALL-E, an AI model that generates images from text prompts, has demonstrated racist and sexist biases. Prompts for ‘CEO’ tended to generate images predominantly of white men, while prompts for ‘personal assistant’ or ‘nurse’ often generated images of women. Similarly, in another instance, robots programmed with a popular artificial intelligence algorithm associated words like ‘homemaker’, ‘janitor’, or ‘criminal’ with particular ethnicities. These unintended biases are a serious concern because they can have far-reaching consequences. They can reinforce harmful stereotypes, leading to discriminatory practices in areas such as hiring, lending, and even criminal justice. Therefore, it’s crucial to be aware of how biases can propagate in AI systems and to take steps to mitigate them.
Protect yourself with responsible Generative AI use
To protect yourself from the unintended consequences of using generative AI, here are three simple pointers to keep in mind:
Always verify and review the outputs generated by AI tools before utilising or sharing them. Remember that human judgement remains essential for quality control. Supplement AI with your own "sense-check," cross-referencing facts with reliable sources such as official websites, reputable publications, or human experts.
Establish ethical frameworks and guidelines for the use of generative AI, espeically if you work in a team. Promoting awareness about the potential risks and consequences will help everyone exercise caution and responsibility when using AI-generated content.
Stay updated on the latest developments and best practices in generative AI. Engage in ongoing education and training to understand the nuances of the technology and its associated risks, ensuring that you are equipped to make informed decisions.
Generative AI offers remarkable possibilities for enhancing productivity and creativity, but it also poses some real risks in the workplace. Ultimately, it’s a tool, and how skillfully it’s wielded depends on its user - you!