Think Artificial Intelligence (AI), and you might come up with The Matrix, C-3PO and R2-D2 from Star Wars, or Arnold Schwarzenegger in The Terminator. Present-day AI is about search engines like Bing, virtual assistants like Alexa, targeted online adverts and content recommendations. AI is an increasingly common feature of credit control software. Using payment history, credit rating, and spending patterns, AI can automate offering early settlement discounts, dunning letters, and electronic reminders. It can even predict the most suitable time of day to telephone an accounts department. Credit reference agencies use AI to determine risk scores which debt collection agencies like Advocate make good use of.
Avoiding Techspeak…How Does AI Work?
AI is deemed present when a computer’s response (output) is more significant than its input (programming). There are many different disciplines at work within AI, so at the risk of oversimplifying, we will now take a quick look at the main principles.
The Turing Test
Enigma code expert and mathematician Alan Turing is regarded as the father of AI. He came up with the Imitation Game (now called the Turing Test), which tested the ability of AI to imitate human behaviour to the point where one was indistinguishable from the other. It raised the question of whether intelligence is the recollection or interpretation of information.
Automated Decision Making
ADM uses algorithms (sequences of instructions) and data from multiple sources to decide. As with all AI, the output is only as good as its programming. Coupled with Machine Learning, ADM can avoid repeating the same mistakes.
Natural Language Processing
NLP is the ability within AI to interpret human text or speech whilst understanding the context and subject. NLP uses a pre-set list of rules and annotations from which to give an appropriate response.
Generative Pre-trained Transformers
GPT takes NLP one step further. It learns to predict the next word in a passage of text. GPT can learn unsupervised and in the moment. Again, the quality of the text is essential, along with an element of supervised learning to correct misunderstandings and programmed bias.
Combine ADM with NLP or GPT, and you’ve got Chatbots capable of passing the Turing Test. Chatbots are most common in B2C situations, especially out of hours or at times of high demand. The rise of Chatbots has coincided with the ascension of tech-savvy Millennials and Generation Z. The most advanced bots can handle a million conversations a day. NLP chatbots, and less so for GPT, are prone to algorithmic bias. They also have a blind spot of being unable to make an exception to the rules. The lack of empathy a computer provides means debtors are less likely to react emotionally to the severity of their predicament. Conversely, people are less embarrassed because a computer is impersonal, potentially producing a more successful outcome.
Launched in March 2023 by OpenAI, ChatGPT-4 goes beyond just imitating human interaction. It can compose music, author essays, write computer programs, entertain hypothetical questions, and even pass the Bar exam. Google has also launched their own next-generation AI called Bard, which, unlike OpenAI’s offering, has sight of current affairs post-2021. AI typically must prove itself elsewhere before being rolled out for taboo topics like health and wealth. For that reason, it will be some time yet before we see this generation of Chatbots working alongside us, mere mortals.
The Good AI – Algocracy
AI is already embedded within the fabric of society, more so than most people think. It can be anything from public contract tenders to traffic management and facial recognition. The UK is an example of an Algocracy where democracy and society are maintained through AI. Others include Australia, America, Japan and Holland.
In Australia, AI-based Split Up software is used to help judges determine how the assets of a marriage are divided in divorce. In America, COMPAS software is used in several states to determine the risk of criminals reoffending. They also use NarxCare to review prescription registries and calculate risk scores. In 2018 an AI mayoral candidate for Tama in Tokyo came third at the ballot box. In 2022 Danish AI political candidate and leader of the Synthetic Party, Ledar Lars, started to engage with voters as a chatbot. The UK’s Metro Bank uses chatbots with the ability to read facial expressions and emotions like frustration or nervousness so it can change the script accordingly. Shenzhen-based Ping An Insurance reads customer faces to determine the likelihood of incomplete disclosures. China’s AI Judge is a virtual judge conducting hearings for the most repetitive functions of courtrooms.
The Bad AI – Robodebt
Designed to tackle benefit fraud by automatically calculating overpayments and pursuing debts through the courts, Australia’s Robodebt was scrapped in 2020 after four tumultuous years. During that time, it collected $751mn of non-existent or incorrectly calculated debts affecting over 400,000 citizens. Robodebt moved the burden of proof on to the individual who had to show the AI’s calculations were wrong. Subsequent enquiries heard testimony from the victims who lost everything and from the families mourning those who saw no other way out. The fundamental and most basic flaw in Robodebt was calculating debts using average income vs benefits rather than looking at what someone had actually earned at a given time vs the benefits received. It has rightly been branded the ‘unlawful debt collection scheme’.
The Ugly AI – SyRI
In 2021 the Dutch government resigned over the benefit-cheated AI software SyRI (System Risk Indication) used between 2014 and 2020. It analysed seventeen categories of personal data from different government agencies, flagging inconsistencies such as someone receiving housing benefits but not being registered to vote at the address. SyRI assigned risk scores and placed individuals on a list where investigators could prosecute if necessary. A 2020 case brought before The Hague District Court ruled that the use of SyRI breached Article 8 of the European Convention on Human Rights and found the system insufficiently transparent. What we know of the algorithms suggests that they applied the most damning explanation to an inconsistency. For example, assuming lower water usage indicated housing benefit fraud rather than a broken meter or being more frugal.
Rise of The Machines
When you monetise software and technology, it becomes the preserve of the rich until such time economies of scale make it affordable to the masses. AI is expensive to develop. The algorithms and technologies are billion-dollar assets shrouded in NDAs. This pushes AI away from open source and into a black box of secrets, lacking visibility or accountability for flaws and bias. Looking at Robodebt and SyRI, AI, by design or accident, can swing the pendulum towards oppression. Add into the mix a new focus on quantum computing, which, unlike regular tech, is not limited to the processing of ones and zeros, and it’s not a massive leap from the Turing Test to surmise AI could learn beyond its design to the point where humans lose control – think Skynet in the Terminator movie franchise. This century’s most prominent tech minds, Hawking, Gates, Musk, and Zuckerberg, agree AI should be a tool for humanity, not a replacement.
At Advocate, we use AI as a debt collection tool because it helps accelerate the path to judgment day, prompting payment before it’s too late. In the words of John Conner, ‘The future has not been written. There is no fate but what we make for ourselves.’