AI Ethics: The 'Black Box' Problem

AI Ethics: The "Black Box" Problem

As I integrate more Agents into my workflow (like the Reconciliation Agent discussed on Jan 5th or the automated underwriting bots), I run into a major philosophical and technical hurdle: Explainability.

In traditional software development, I know exactly why something happened. The logic is explicit.

Code: if (applicant_age < 18) { return "Rejected"; } Result: I rejected the user because they are under 18. It is binary, traceable, and legally defensible.

In modern AI (Deep Learning), it is different. You feed data into a Neural Network, it performs a trillion floating-point math operations across hidden layers, and it spits out an answer.

AI: "Reject this loan application." Human: "Why?" AI: "..."

The AI cannot easily explain its reasoning. It just "feels" right based on the patterns it saw in the training data.

The Black Box

This is the "Black Box" problem. I know the Input, and I see the Output, but the middle is opaque.

If I cannot explain why the AI made a decision, can I trust it with money? Can I trust it with medical diagnosis? What if the AI rejected the loan not because of financial risk, but because its training data contained historical biases against certain zip codes or demographics?

Regulatory bodies (like the EU with the AI Act) are now demanding XAI (Explainable AI). You cannot just say "The computer said no." You must provide a reason.

my Approach: Human-in-the-Loop (HITL)

At g-makris.com, I mitigate this risk by following a strict "Human-in-the-Loop" architecture for critical decisions. I never let the AI have the final say on high-stakes actions without a paper trail.

• Citations: I force the AI to cite the source data. It cannot just hallucinate a fact; it must say: "I recommend rejecting this because of Document A, Page 4, Paragraph 2." This allows a human to verify the claim.
• Confidence Scores: The AI must state how sure it is. "I am 85% confident in this match."
• Thresholds: If the confidence is below 99%, the system automatically flags the item for Human Review. A human must click "Approve" before the action is taken.

AI is a powerful engine, but it shouldn't steer the car without a driver. I use AI to augment human intelligence, not to abdicate human responsibility.

Best,

Gerasimos Makris Founder of g-makris.com AI Web Developer | Double Master's in MBA & FinTech and Blockchain

Tech Glossary & Concepts

• Black Box: A system which can be viewed in terms of its inputs and outputs, without any knowledge of its internal workings.
• Neural Network: A computer system modeled on the human brain and nervous system, designed to recognize patterns. It "learns" by adjusting weights between nodes.
• XAI (Explainable AI): A set of processes and methods that allows human users to comprehend and trust the results and output created by machine learning algorithms.
• Human-in-the-Loop (HITL): A model of interaction where a human is required to interact with the AI system to provide feedback, confirmation, or error correction.
• Hallucination: When an AI model generates false or misleading information but presents it as a fact.