July 11, 2023 | Nick Holland
If you were at the Money20/20 Amsterdam in June, one recurring theme in the sessions and in conversations was AI, and specifically the integration of the latest technologies within financial applications. It would be hard to have missed hearing “ChatGPT” and “Generative AI” in some shape or form over the course of the three days.
However, it’s a technology that is fraught with concerns about its ability to upend careers, governments and even the role of humans, as the myriad of countries stepping up to coerce the AI genie back into a somewhat more controlled bottle demonstrates.
For instance, on June 16, a Reuters article broke the news that Southeast Asian countries are drawing up governance and ethics guidelines for artificial intelligence (AI) that will impose "guardrails" on the booming technology. A spokesman for Singapore’s Ministry for Communications and Information said that as 2024 chair of that meeting, the country would be collaborating with other ASEAN states "to develop an ‘ASEAN Guide on AI Governance and Ethics’ that will serve as a practical and implementable step to support the trusted deployment of responsible and innovative AI technologies in ASEAN."
So why is there so much interest and concern about AI now? And just what is the “generative” bit all about? Are we talking Skynet, or something more like Microsoft’s “Clippy” on steroids?
Generative Adversarial Networks (GANs) were first introduced in 2014 by Ian Goodfellow, Yoshua Bengio, and Aaron Courville. The paper presented the GAN framework and its training algorithm, describing how the generator and discriminator networks compete and learn from each other.
You’ve probably seen GANs in action over the past few years, with Google’s AI that would turn ordinary photos into some sort of hallucinogenic monster collage of dogs, or more terrifying, deepfakes where I become ALL of the Spice Girls (Editor: this is why Nick isn’t allowed nice things). At the consumer end it’s been pretty playful, but we’re just starting to see the real power of AI in its usage for text to image or text representations.
For instance, here’s a Haiku that ChatGPT wrote about Money20/20. In milliseconds.
Money20/20 glows, Fintech minds ignite the stage, Wealth's future unfolds.
GANs are a class of machine learning models designed to generate new data that resembles a given training dataset. GANs consist of two main components: a generator and a discriminator, which are neural networks that compete against each other in a game-like manner.
Generator: The generator takes random input (often referred to as noise) and tries to generate synthetic data that resembles the training data. It learns to map the noise to output data by transforming it through multiple layers of neural networks. The objective of the generator is to generate data that can fool the discriminator.
Discriminator: The discriminator acts as a binary (YES / NO) classifier that distinguishes between real data from the training set and the synthetic data generated by the generator. It is trained on labeled data and learns to classify whether a given sample is real or fake. The objective of the discriminator is to correctly identify the real data while flagging the generated data as fake.
During training, the generator and discriminator play a min-max game. The generator aims to generate data that the discriminator cannot differentiate from real data, while the discriminator aims to correctly classify the real and generated data. The training process involves an iterative back-and-forth between the two components, with updates to their respective weights and parameters based on their performance.
As training progresses, the generator improves its ability to generate increasingly realistic data, while the discriminator becomes more adept at distinguishing real and generated data. Ideally, this competition and cooperation lead to a point where the generator generates data that is almost indistinguishable from the real data, resulting in high-quality synthetic samples.
A recent paper from AI Verify Foundation of Singapore outlines five key risk areas surrounding Generative AI —
RISK 1: MISTAKES AND “HALLUCINATIONS”
Like all AI models, generative AI models make mistakes. When generative AI makes mistakes, they are often vivid and take on anthropomorphisation, commonly known as “hallucinations”. For instance, current and past versions of ChatGPT are known to make factual errors. Such models also have a more challenging time doing tasks like logic, mathematics, and common sense. This is because ChatGPT is a model of how people use language. While language often mirrors the world, these systems however do not (yet) have a deep understanding about how the world works. Additionally, these false responses can be deceptively convincing or authentic, meaning that humans are more likely to consider the results as accurate.
RISK 2: PRIVACY AND CONFIDENTIALITY
Generative AI tends to have a property of “memorisation”. Typically, one would expect AI models to generalize from the individual data points used to train the model, so when you use the AI model there is no trace of the underlying training data. As the neural networks underpinning generative AI models expand, these models have a tendency to memorize.
A worrying finding is that parts of sentences like nouns, pronouns and numerals are memorized faster than others – precisely the type of information that may be sensitive. This memorisation property also poses copyright and confidentiality issues for enterprises and companies.Given that ChatGPT utilizes user prompts to further train and improve their model unless users explicitly opt out, that information is now out in the wild.
RISK 3: SCALING DISINFORMATION, TOXICITY AND CYBER-THREATS
Toxic content — profanities, identity attacks, sexually explicit content, demeaning language, language that incites violence — has long been a challenge on social media platforms. Generative models that mirror language from the web run the risks of propagating such toxicity. In addition, impersonation and reputation attacks have become easier, whether it is social engineering attacks using deepfakes to get access to privileged individuals or reputational attacks by offensive image generation. With generative AI being able to generate images in one’s likeness, there is a question of whether this constitutes an intrusion of privacy.
RISK 4: AN ERA OF COPYRIGHT CHALLENGES
AI and machine learning models have always operated on the basis of identifying patterns present in relevant data. Current generative AI models require massive amounts of data. Scraping the web for data at this scale has exacerbated the existing concerns of copyrighted materials used.
RISK 5: EMBEDDED BIASES WHICH ECHO IN DOWNSTREAM APPLICATIONS
AI models capture the inherent biases present in the training dataset (e.g. corpus of the web). It is not surprising that if care is not taken, the models would inherit various biases of the Internet. Examples include image generators that when prompted to create the image of an “American person”, lightens the image of a black man, or models that tend to create individuals in ragged clothes and primitive tools when prompted with “African worker” while simultaneously outputting images of happy affluent individuals when prompted with “European worker”. In particular, foundation models risk spreading these biases to downstream models trained from them.
Generative AI is being utilized in various ways within the field of fintech to enhance and streamline processes. Here are a few examples:
Generative AI models can analyze large volumes of financial data to identify patterns and anomalies indicative of fraudulent activities. By learning from historical data, these models can improve fraud detection accuracy and assist in preventing financial crimes.
Generative AI algorithms can assist in evaluating creditworthiness and assessing risk by analyzing diverse data points. They can help automate the process of assessing loan applications, detecting patterns in customer behavior, and making predictions about default probabilities.
GANs can be employed to simulate and predict market behavior. Traders can use these models to generate synthetic data that reflects market conditions and test trading strategies before applying them in real-time.
Customer Service and Chatbots:
Generative AI can power conversational agents and chatbots in the fintech industry. These AI systems can handle customer inquiries, provide support, and offer personalized recommendations based on customer preferences and transaction history.
Risk Modeling and Forecasting:
Generative AI models can be trained on historical financial data to analyze market trends, identify patterns, and generate predictive models. These models can assist financial institutions in risk modeling, portfolio optimization, and making data-driven investment decisions.
Generative AI is genuinely game changing for all industries, and as fintech is very much at the leading edge of iterative testing and experimentation for new technologies, we can expect to see the technology increasingly embedded in our banking and payment apps, in smart speakers in our cars and homes, at retail checkouts, and probably in all aspects of our lives. Clearly, with great power comes great responsibility, and intelligent regulation of the technology is necessary to prevent some of the risk factors outlined from becoming pervasive.
However, there is a distinct sense that any guardrails will be reactive and overly draconian, potentially stifling innovation. Keeping abreast of key conversations pertaining to Generative AI in fintech is going to be fundamental to anyone in this space over the coming months, and a core topic of conversation and learning for Money20/20 attendees around the world.