This is the fifth installment of AI Fluency Corner, a 16-part weekly series that builds connected mental models of artificial intelligence (AI) in plain language.
Unlock your phone with your face in the dark. A banking app reads your handwritten receipt. The chatbot routes voice complaints to the appropriate department. None of this is guesswork, magic, or particularly luck. This is deep learning, and the same underlying architecture powers all three.
Last week, we looked at machine learning, or finding patterns in data to improve systems. This week, we take a step inside the architecture behind some of the most capable AIs. Neural networks are not new. What has changed is what happens when you make them very deep.
What are neural networks and what they actually are?
Although this name is borrowed from biology, it should not be romanticized. Neural networks are not digital brains. There is no thinking, reasoning, or feeling. This is a mathematical structure. Each layer of connected processing nodes receives input, applies weights, and passes the results forward.
Here’s an example to click on: Imagine a talent show with 50 rounds of judging. Round 1: The performer moves to the left or right, whether the pitch is high or low. Scores were passed before. Round 2 uses these scores to focus more on rhythm, timing, contrast, and more. By round 50, the final judges had never seen the performer. Although they receive only increasingly sophisticated summaries, they still pinpoint the act.
That is a neural network. Intelligence emerges from sequences, not a single layer. If there are many layers, the system becomes a deep neural network and this approach is called “deep learning.” “Deep” means many layers, not deep wisdom. This is an important difference.
How your face becomes a mathematical certainty
Facial recognition shows this most clearly. When your phone scans your face, the image is fed into a convolutional neural network as a grid of pixel values. The first layer only detects basic features such as edges, curves, and contrast. They don’t know what a face is. They do some light reading.
The middle layer joins these edges into recognizable structures (eye sockets, bridge of the nose, jawline). The deepest layer assembles these into geometric signatures (distances and angles between features, i.e., user-specific) and compares them to stored templates. A match within the learned tolerance will unlock the phone.
This is why photography cannot fool modern systems. A flat image produces a two-dimensional edge pattern. Real faces have curvature, shadows, and depth, and the deeper layers learn these differences from thousands of examples. The system did not remember your face. It’s learned what makes your face mathematically yours.
If deep learning is already working
FNB, Standard Bank, and Nedbank have introduced models for fraud detection that combine transaction amounts, device signals, timing, location, and behavioral history to create probability scores that cannot be replicated by rules-based systems. That power is not one dramatic flag. It takes a few dozen weak signals to conclude with confidence.
The voice notes your phone converts to text, the email summaries your clients generate, and the recommendations you make at your retail store are all Transformer-based deep learning. ChatGPT, Copilot, and Gemini learned which word sequences follow other word sequences from billions of examples. They sound consistent because they are good at linguistic pattern matching, not because they understand your question the same way you do.
Limits that never appear on slide decks
Neural networks have one built-in structural weakness. That is, neural networks cannot explain themselves. Deep learning credit models can accurately deny 1 million applications. If you ask why it rejected this particular person, architecture has no answer. This decision resulted from millions of calculations across dozens of layers. There are no rules to surface, no chains of reasoning to read. This is a black box problem, not a technical inconvenience in a regulated industry. It’s a responsibility.
South Africa’s Privacy Act grants individuals the right to human review of automated decisions. The National Credit Act requires that adverse credit outcomes be explainable. Vendors who can’t explain why their model says no may fail not only the technical test, but the legal test as well.
Second limitation: The model knows the world in which it was trained, not the world as it is. Language models built on English-dominated texts may perform poorly in Zulu or Sotho. Not because these languages are inferior, but because they were undervalued in training. A model cannot know what it is never shown.
A model that no one is paying attention to is not a system. This is an assumption and will run unattended.
3 questions that separate fluency from faith
- What was the model trained on? And does it cover your context? A system built on American or European faces, accents, and financial behavior is likely to underperform where you need it most.
- Can the output be explained for each decision? Precision across the population does not satisfy the obligation to explain the outcome to the individual. Brochures are characterized by accuracy. Ask about the explanation.
- How will performance be monitored after deployment? Neural networks do not self-correct. As behavior changes and markets move, unmonitored models drift confidently, quietly, in the wrong direction. A model that no one is paying attention to is not a system. This is an assumption and will run unattended.
our mission this week
The next time an AI feature starts to see, hear, and understand context, stop and ask yourself: What was this model trained on, and perhaps something we haven’t seen before? This question, if applied consistently, is more valuable than any product demonstration. The most dangerous thing about neural networks is not that they are completely wrong. It means we stop asking questions because we often get things wrong.
• Mafinyani is a senior partner in financial engineering and AI at Intellica Analytics, a firm specializing in finance, risk and applied technology. Next week: Natural language processing — Natural language processing is already built into the tools you use and can replace manual text work.
