top of page

AI’s Hidden Decisions: Why We Don’t Know

Why don’t we KNOW what AI systems will decide to do? Or how they make decisions? Didn’t programmers build these systems? These questions are completely reasonable to anyone unfamiliar with how machine learning works. After all, we’ve been working and living alongside traditional programming outcomes for many years so that paradigm is well understood, even if you’re not a technologist.

Do you remember a time when traditional programming wasn’t a thing? Of course not. Programming languages were created beginning in the 1950s, BASIC was created for non-techies to code with in 1964, and Microsoft introduced drag-and-drop capability in 1991. So we’ve all been surrounded by classically programmed systems for a long time.

Everyone knows that traditional programming involves explicitly giving step-by-step instructions to computers to output very specific results. But machine learning and generative AI are ushering in a new paradigm where the outputs are not fully planned or transparent ahead of time. This change is so significant that it’s confusing and uncomfortable. Like any major paradigm shift - think about life before and after electricity - it’s worth the effort to embrace the learning discomfort.

Game Changing Concept

The core concept driving this shift is that we can build algorithms that learn from immense amounts of data to carry out tasks without being explicitly programmed for every scenario. Rather than programmers directly scripting every command, machine learning systems infer patterns from examples to make predictions, create art, generate text and more. This move toward emergent, self-supervised learning from data represents an immense expansion of what was previously possible.

Let's say you want a computer to identify pictures of cats. Instead of writing a detailed program that tells the computer exactly what a cat looks like, you can show the computer thousands of pictures of cats and let it learn on its own. It will start recognizing common patterns in those pictures, like pointy ears and whiskers, and use that knowledge to identify cats in new pictures it hasn't seen before.

However, there are some drawbacks to this shift. One of them is that we lose a certain level of control and transparency. Traditional code is highly directive, spelling out what the computer is doing at each step. But with complex neural networks, the inner workings are more like a black box. We can't easily see how the computer arrived at a particular decision or prediction.

Imagine you have a magic box that can solve complex math problems, but you don't know how it does it. You put in an equation, and it gives you the answer, but you can't peek inside to understand the steps it took. That lack of transparency can make people uneasy, especially those who are used to computers following explicit instructions.

During this transition period, it's natural to have concerns about transparency. We're moving away from a world where computers simply follow our commands to one where machines learn and make decisions on their own. It's a new paradigm that brings both exciting possibilities and challenges that we need to navigate.

Real World Uses Cases

The value we gain in return for making this mental adjustment from classic programming is evident. Machine learning is already powering applications that would have been impossible just a few years ago. Here are some daily life impacting examples that we already take for granted:

  1. Personalized product recommendations on Amazon: We all have our own intelligent shopping assistant that suggests products tailored specifically to our tastes and preferences. Machine learning algorithms analyze our past purchases, browsing history, and interests to generate personalized recommendations. This level of customization and accuracy would have been unthinkable just a few years ago.

  2. Photo recognition abilities in apps like Google Photos: Thanks to machine learning, photo organization has reached a whole new level. Apps like Google Photos can analyze and identify scenes, objects, and even individuals within your photos. Without the need for explicit programming for each image, these algorithms learn from vast datasets, allowing them to recognize common visual patterns and attributes. This sophisticated capability has transformed the way we manage and search through our ever-increasing collection of digital images.

  3. Language translation models like Microsoft Translator: Breaking down language barriers has become increasingly seamless with machine learning. Sophisticated translation models such as Microsoft Translator achieve a remarkable level of accuracy, comparable to human-level understanding, across a wide range of languages. Gone are the days of relying solely on rule-based dictionaries for translation; these models learn from extensive multilingual datasets to grasp the intricacies and nuances of different languages, enabling smooth and accurate communication across linguistic boundaries.

These examples illustrate the remarkable advancements made possible by machine learning. The ability to extract meaningful insights from huge amounts of data and learn complex patterns is altering our daily lives. With machine learning at the helm, we are witnessing the realization of applications that were once considered impossible.

Generative AI represents the next step by creating novel text, images, videos and other outputs rather than just analyzing existing content. This ability to create entirely new content opens up even more previously unimagined applications.

Pros and Cons

Embracing new machine learning and generative AI systems comes with its own set of pros and cons:


  • Enhanced Capabilities: These systems can perform complex tasks and make predictions that were previously challenging or impossible for traditional programming.

  • Efficiency and Productivity: AI can automate repetitive tasks, freeing up human resources for more creative and strategic work.

  • Personalization: AI can tailor experiences and recommendations to individual users, enhancing user satisfaction and engagement.

  • Data Insights: Machine learning can extract valuable insights from large datasets, enabling better decision-making.

  • Innovation: Generative AI can create novel content, leading to innovative products, services, and artistic creations.

  • Accessibility: Machine learning can make technology more accessible, enabling non-experts to develop AI-driven solutions.


  • Loss of Transparency: AI systems can be complex black boxes, making it challenging to understand how they arrive at decisions.

  • Bias and Fairness: If not carefully designed, AI systems can perpetuate biases present in training data, leading to unfair outcomes.

  • Data Dependence: AI's performance heavily relies on the quality and diversity of training data. Biased or inadequate data can lead to inaccurate results.

  • Job Displacement: Automation could lead to job losses in certain industries, raising concerns about unemployment.

  • Ethical Concerns: AI raises ethical questions about accountability, privacy, and the responsible use of technology.

  • Complexity: Developing and maintaining AI systems requires specialized skills and resources, potentially creating barriers for smaller organizations.

  • Unintended Consequences: AI systems might produce outcomes that were not intended or anticipated by their developers.

Balancing these pros and cons requires careful consideration, ethical guidelines, ongoing monitoring, and a commitment to responsible AI development and deployment.

The Future of Emergent AI

As machine learning systems continue to advance and are applied to more complex tasks, I predict some key developments in how we utilize and oversee emergent artificial intelligence:

  • Growing Adoption but also Increased Scrutiny: Machine learning and generative AI will see widespread integration across industries. But this will also bring increased attention to issues like bias, fairness, safety and explanations.

  • Multi-pronged Explainability Efforts: Both internal model explainability and external system auditing efforts will receive increased focus. Greater explainability will help address transparency but won’t eliminate all concerns.

  • Self-supervised Learning Advances: Future systems will learn from interacting with the world rather than relying solely on labeled training data from humans. This will reduce labeling demands while increasing algorithmic independence.

  • Creative Human Partnership with AI: As outputs become more complex, humans and AI will collaborate as creative partners with each bringing their own strengths to projects. Deep generative models will inspire novel ideas that are then developed by humans.

  • New Governance Models: Governments and international organizations will continue enacting competing and complex regulatory frameworks and auditing models for emergent systems. Public-private partnerships will inform guidance.


We've become familiar with the straightforward instructions of traditional programming, making the transition to machine learning and generative AI perplexing and unsettling. It's natural and prudent to question the unclear decisions made by AI systems. However, similar to past shifts like the adoption of electricity, there’s value in thoughtfully and carefully embracing this learning curve, as the benefits these systems offer make the journey worthwhile. I truly believe that we will cure cancer in partnership with AI.

41 views0 comments


bottom of page