An ABC of AI

Leon Schumacher, from Digital Ekho, gave the opening talk at the AI & CB (Artificial Intelligence and Central Bank) online conference in December. For those new to AI, it provided a fast overview of what AI is and some of the issues connected to it.

Artificial Intelligence (AI) is a branch of computer science that deals with the creation of computer systems able to perform tasks usually requiring human intelligence.

AI is based on three capabilities. Today the only category of AI is ‘narrow AI’, specialising in one task. General AI, also known as AGI, can do any intellectual task a human can do. Finally, there is Super AI which surpasses human intelligence.

Narrow AI

Within Narrow AI there is Automated, Augmented and Autonomous AI. Automated AI replaces tasks currently done by humans. Automated AI is robotic process automation. Augmented AI pairs humans and machines. It learns continuously to help humans make augmented decisions. Autonomous AI replaces humans because it allows autonomous systems to operate independently of human control.

Narrow AI is already part of everyday life. The sort of tasks that Narrow AI can do is generating content. The likes of Chat GPT, Google Bard, Microsoft CoPilot etc. are focused on this. It can give recommendations, for example Netflix and Amazon suggesting films you may like. Facebook’s photo tagging is image recognition, while Siri/Alexa/Cortana etc. recognise speech. The almost ubiquitous online bots are AI. Self-driving cards use a combination of AI types. IBM’s Deep Blue or AlphaGo by DeepMind that plays Go, Chess etc. are AI. Card companies use AI to identify anomalous behaviour.

Machine learning

Machine learning (ML) is a sub-field of AI in the same way as Expert Systems, Robotics and Search Algorithms are. ML is a programme or system that trains a model from input data. It gives a computer the ability to learn from data without explicit programming.

Within ML there is supervised learning, where computers are given information with labels, unsupervised learning, which is information without labels, reinforced learning, when it is given live data that it has to work out, and deep learning. Deep learning uses layers of algorithms known as Artificial Neural Networks (ANN). These can process much more complex patterns than traditional ML and are inspired by human brain neurons.

Within deep learning there is Discriminative AI, which classifies or predicts labels for data points, and Generative AI (GenAI), that generates new data based on a learned probability distribution of existing data. The new data can include text, audio and video.

But how does all this get us from AI to ChatGPT? The path is AI to ML to Deep Learning to GenAI to Large Language Models to Generative Pre-trained Transformers to Generative Pre-Trained Transformer series to GPT-4 with Chat API to form Chat GPT.

Issues with AI

• Ethical and societal impacts – extinction risks, biases, surveillance, deepfakes, autonomous weapons, bioweapons etc.

• Job displacement and economic inequality – job losses, unemployment, widened economic inequality

• Responsibility and explainability – black boxes, transparency, understandable decisions, liabilities.

Recent Developments in AI and the Central Banking Response

The Head of Information Technology and Services at the Bank for International Settlements (BIS) built on Leon Schuhmacher’s introduction to consider AI and Central Banks.

A mix of neural networks, what he termed hyperscale hardware, the transformers already mentioned in the introduction, instruction tuning, and vast amounts of data form the key ingredients that enable and create a ‘ChatGPT’ cake.

Risks and opportunities

There are, of course, two ways to look at AI, and both can be simultaneously true.

First, that new AI capability could transform the work of central banks, albeit with risks. The ability of AI to enhance human computer interactions, eg. translating, manage document enquiries and summarising, power intelligent searches extracting and interpreting data and creating draft documents, including source code generation, have applications across both how organisations work and what they work on – automating operations, acting as a legal co-pilot, accessing multi-lingual communications, economic forecasting etc.

Major challenges and risks around the adoption of AI include the ability to explain and replicate the answers given, reliability, cloud security, data sovereignty, misinformation, disinformation, bias and discrimination, cyber security, ethical use, copyright, personal data protection and regulatory alignment. This last challenge will be exacerbated as different countries adopt different regulations. Work is starting but there is much to do before the full potential of AI is accessible.

Second, new AI capabilities could destabilise markets and harm consumers. This, of course, is what is driving the regulatory and supervisory work that is now going on. Risks here include the concentration of power in a small number of non-traditional entities which may encourage procyclical dynamics and undermine attempts at regulation.

Automated AI capabilities may lead to systemic risks such as ‘flash crashes’ in financial markets. The long-term impact of AI on the real economy are unknown.

New consumer protection measures will need to address the long list of risks already mentioned, including identity theft and fraud, bias and discrimination, misinformation and disinformation.

A Legal Perspective: Challenges of AI Transaction Monitoring

Prof Dr iur Cornelia Stengel, a partner at the law firm Kellerhals Carrard, talked about AI-supported transaction monitoring, giving some real life examples of the risks discussed by the BIS.

AI offers the ability to reduce the traditional manual effort required to monitor transactions. Traditional rule ‘engines’ can be automated by AI applications, supported by some manual editing.

Traditional rule engines work on a ‘if- then’ model. The decision is binary to the extent that either an alert is generated according to the programmed rules, or not. No information is included about the probability of a transaction with increased risk according to EU’s anti-money laundering authority. An AI application powered by a machine learning algorithm detects patterns independently through the analysis of training data. It makes the decision based on statistical correlations.

The ‘Bias’ and ‘black box’ problem

A major challenge is producing an algorithm which does not take into account problematic data characteristics, for example skin colour or gender. A statistical review process is needed to establish if there is bias in the algorithm.

The goal has to be for the algorithm to produce a correct, robust and reproducible result. AI can often seem like a black box where the result is opaque and/or inconsistent. The selection of the model has to be documented to check the suitability for concrete application, it must use statistical considerations for the model performance and take into account interpretability and verifiability, a major task for complex models.

Relevant calibration details, such as the selection of training data and use of tuning parameters, must be captured and the validation model needs to be carefully described. All of this needs to be detailed and auditable, particularly the learning processes employed.

Legal basis

The European Union (EU) presented a proposal for a regulation on AI in April 2021. It established an advisory group at the Council of Europe level that has prepared a draft binding legal framework for the regulation of AI. On 9 December 2023, Parliament reached a provisional agreement with the Council on the AI act. The agreed text will now have to be formally adopted by both Parliament and Council to become EU law.

AI Discrimination Bias in Consumer Financial Services

While Prof Stengel had presented the risks in a broad brush approach, Nicholas Schmidt of SolasAI gave a more granular example of how difficult balancing machine learning (ML) is, based on consumer financial services. SolasAI creates software that measures and mitigates discrimination bias.

ML is good at finding patterns in data and it can handle more data and more complex data better than traditional algorithms. The problem is that it is hard to know why most ML algorithms arrive at a prediction. The algorithms tend to overfit data and, to compound those problems, people have too much trust in ML and AI. If humans already understand the data-generating process, applying ML does not add value.

This matters a great deal if the decisions AI is making are important. The BIS mentioned bias and discrimination in its presentation. To illustrate this, Nicholas Schmidt shared a New York Times article from February 2018 by Steve Lohr, ‘Facial recognition is accurate, if you’re a white guy’. The article quoted research that found that gender was misidentified in up to 1% of lighter-skinned males by AI software out of 385 photos, 7% of lighter skinned females out of 296 photos, 12% of darker-skinned males out of 318 photos and 35% of darker-skinned females out of 271 photos.

He explained that while addressing algorithmic discrimination is essential, it may well be that the algorithm is playing a small role in the model system. The development of a credit model assessment used to make credit offers may start with an analysis of differences in customer demand. This drives business decisions regarding markets and locations and uses preferences based on pre-existing customers to create marketing materials and targeted outreach campaigns. Only at this stage are there marketing responses and applications that are used in the credit model assessment.

ML models have the potential for flexibility and more than one model may meet all necessary model governance requirements. This allows organisations to optimise their decisions based on multiple metrics, including fairness. It also means that individual feature explanations may not be reliable or tell the full story. Effectively, AI can be used to fix AI problems.