In This Handbook
What is Generative AI?
Generative AI represents a fundamental shift in how we build software, moving from strictly deterministic programs to probabilistic and creative systems. In traditional deterministic software, the same input always produces the same output. If you press a button or run a function with identical data, you can expect an identical result every time. This predictability is a core principle of classical computing; it makes programs testable and reliable. Generative AI flips this on its head. These systems generate new content, and may produce a different valid output each time you run them.
From deterministic to probabilistic software
Think of how a simple tax calculator works: you input your income, it applies fixed rules and spits out the same tax due every time. That’s determinism. Generative AI, by contrast, is probabilistic software. It doesn’t have one hard-coded answer for a given prompt, instead, it looks at your input and predicts an appropriate response based on patterns learned from vast data. Because it’s driven by probabilities, a generative model might answer the same question in multiple valid ways. For example, if three people ask a generative AI assistant the same question about a policy, it might explain it in three different yet correct ways (perhaps one answer is a step-by-step list, another a conversational explanation). All those answers could be accurate and helpful, just phrased differently. This variability is by design. It’s a feature that allows AI to be more adaptable and personalized.
One immediate implication of generative AI’s probabilistic nature is that there is often no single “correct” output. Traditional programs had a rigid notion of correctness, i.e. an exact expected result. But generative AI models return many acceptable answers, each shaped by learned probabilities. When an AI writes an email or answers a question, there isn’t just one “right” way. It’s essentially choosing from a weighted menu of possibilities, maybe there’s a 60% chance of one phrasing, 20% chance of another, and so on. All of them could convey the same factual content. This means quality is judged more on reasonableness, usefulness, and adherence to guidelines than on matching a single expected output. As one AI product director put it, we have to redefine what “correct” means when an AI can give multiple valid answers to the same query.
ⓘ How generative models produce outputs
At a high level, generative models learn a probability distribution over data during training. Generation happens by sampling from that learned distribution rather than following fixed rules. This is why outputs can vary even for the same input: each response is one possible sample among many plausible ones. Different generative models implement this idea in different ways, but the core concept remains the same.
What “Generative” really means
Generative AI systems create new content, unlike systems that classify or retrieve existing data. They learn patterns from training data (language, images, etc.) and synthesize original, statistically similar outputs. This “generative” ability does not imply human consciousness, but the production of novel combinations. For example, an image model can generate a realistic, non-existent cat, and a language model can write a statistically valid, new sentence.
ⓘ Training Time vs. Usage Time
Generative models operate in two very different phases. During training, the model learns patterns from massive datasets in an expensive, offline process. During inference (normal usage), the trained model is fixed and does not update its knowledge. When you interact with a generative AI system, it is not learning from you in real time, rather, it is applying what it learned during training to generate a response.
Generative AI differs from earlier, discriminative AI by producing, rather than recognizing, patterns. They are inherently non-deterministic, meaning multiple runs can yield different results, especially with “creativity settings” applied. This introduces beneficial creativity and diversity into tasks like writing and design, where many possibilities exist rather than a single correct answer.
ⓘ Generative AI Is Bigger Than Language Models
Generative AI is a broad category of models that can create new data across many modalities, including text, images, audio, and video. Large Language Models (LLMs) are one important subclass of generative AI focused specifically on language. While LLMs power chatbots and code assistants, other generative models, such as diffusion models for images or audio synthesis models for speech, use different architectures and training objectives. In the upcoming lesson, we’ll zoom in on LLMs, but it’s important to remember that they represent just one branch of the broader generative AI landscape.
Generative AI vs. traditional AI
AI is broad. Traditional AI is often deterministic or rule-based (like sorting algorithms or expert systems) and excels at well-defined problems, typically predicting a fixed output or label (e.g., “spam” vs. “not spam”). Generative AI, a subset of probabilistic, learning-based AI, focuses on creating new data (text, images, etc.) that didn’t exist before.
Traditional AI answered multiple-choice questions; Generative AI answers open-ended, free-form questions, making it far more flexible (e.g., writing a poem, explaining physics, brainstorming recipes). This flexibility, however, introduces challenges like unpredictability, evaluation difficulty, and the need for massive data and compute resources. In short: classical AI predicted/prescribed; generative AI composes/creates.
Examples
Generative AI spans many media. For a clear mental model, consider a few examples:
- Text generation (ChatGPT, etc.): You give a prompt, like “Explain the water cycle to a 5th grader,” and the AI writes a unique explanation in paragraph form. Ask again, it might phrase it differently or use a different analogy, but still correctly explain evaporation and rain. These large language models generate essays, answers, stories, dialogues, code documentation.
- Code generation: Models like GitHub Copilot or CodeGPT can generate programming code. You might start typing a function, and the AI suggests the next lines. For the same specification, it could write the code in a slightly different way each time (different variable names or structure), yet still solve the problem. It’s synthesizing new code based on patterns learned from millions of code examples.
- Image generation: Tools like DALL·E, Midjourney, or Stable Diffusion take a text description (e.g. “a castle on a floating island, in watercolor style”) and generate a new image that fits that description. Every image is unique. If you run the same prompt twice, you’ll get variations. The AI isn’t pulling from a database of existing pictures; it constructs a new image by sampling from the visual patterns it learned. The results can be startlingly creative and different each time.
- Audio and video: Though slightly behind text and images, generative models can create music (e.g. composing a melody in the style of Mozart), synthesize voices (text-to-speech that sounds like a specific person, or completely fictitious voices), and even generate video snippets or animations from descriptions. These are emerging capabilities showing that generative techniques apply broadly.
In all these cases, generative AI outputs are not predetermined. They are produced on the fly and can exhibit variation. That’s powerful as it enables personalization (the AI can phrase something the way you might like) and creativity (coming up with designs or ideas humans didn’t explicitly program). But it also means we don’t judge these systems by a single right/wrong output, rather by a range of criteria like whether the output is plausible, useful, and follows the instructions.
Limitations and misconceptions
With all the excitement, it’s important to clarify what generative AI is not. They are given in the table below:
| Area | What the Model Can Do | Core Limitation |
|---|---|---|
| Understanding | Generate fluent, well-structured language | Lacks true understanding of meaning, intent, or truth |
| Knowledge | Appear knowledgeable across many domains | Does not “know” facts; relies on statistical patterns |
| Reasoning | Mimic reasoning steps and explanations | No grounded world model; brittle outside seen patterns |
| Accuracy | Produce confident, detailed answers | Can hallucinate plausible but false information |
| Consistency | Maintain coherence within short contexts | Can contradict itself; no persistent beliefs or memory |
| Judgment | Sound authoritative or decisive | No awareness of correctness or uncertainty |
| Agency | Follow instructions convincingly | No goals, intent, or common sense unless explicitly engineered |
These limitations explain why generative AI must be treated as an assistive system. Without verification, guardrails, or grounding mechanisms, models can mislead users, propagate errors, or fail unpredictably in novel situations.
In short: generative AI simulates intelligence, but it does not possess understanding, awareness, or intent. Any system that depends on it must account for these constraints by design.
It’s also worth clarifying some misconceptions:
- “Generative AI is just copy-pasting from training data.” This is inaccurate. While a well-trained model resembles the style of its training data and can regurgitate specifics (like quotes or formulas) if prompted, it typically generates new, original combinations rather than verbatim copying large chunks.
- “Bigger models = always better.” Larger models generally show better capabilities, but improvement isn’t linear. Huge models can still make mistakes if poorly trained or aligned, and are expensive to run and hard to control. For specific tasks, a smaller, fine-tuned model or one with tools (like a calculator) may outperform a larger, raw model. Size isn’t everything; training and usage are key.
- “AI will just figure it out if we throw it at a problem.” Generative AI is powerful, but not magic. It requires clear prompts and extensive, relevant training data to avoid irrelevant output and know about a topic. It lacks inherent knowledge of your business policies or post-cutoff current events. Useful outcomes demand thoughtful integration, not just plugging in a model.
Which brings us to a key point: achieving robust, reliable performance from generative AI requires thinking about systems, not just models.
Generative models are just one component in a larger system. Like an engine needs a car, a generative model needs surrounding apparatus to be truly useful. Successful AI products, such as an AI tax assistant, are hybrid systems. The user-facing chatbot part uses an LLM for flexible language, but behind it, other modules retrieve factual information (like tax codes) and a deterministic calculator performs reliable computations. The generative AI adds a friendly explanation layer, while the final, critical numbers come from a verified, non-generative source. This orchestration balances the generative model’s flexibility with deterministic accuracy. Building with generative AI means carefully designing a system with checks, balances, databases, and specialized models alongside the main generative model, not blindly trusting one model to do everything.
Generative AI is a probabilistic, pattern-based, creative software, not a truly intelligent magic box. It’s a learned model that produces diverse outputs with the right inputs. Next, we’ll explore Large Language Models (LLMs), the crucial generative AI behind chatbots and many text-based AI applications.
Key takeaways:
- Generative AI produces novel outputs (text, images, etc.) by learning patterns from data, unlike deterministic software which follows fixed rules and always gives the same result.
- These systems are probabilistic: the same input can yield multiple valid answers, reflecting different ways to respond. There isn’t one “correct” output – instead, quality is judged by usefulness, accuracy, and context.
- Generative models do not truly understand the content they generate. They predict likely words or patterns without grounding in actual facts or comprehension, which can lead to confident-sounding but false answers (AI “hallucinations”).
- Generative AI differs from traditional AI: instead of just classifying or making yes/no decisions, it creates new content. This enables flexibility and creativity but introduces challenges in control and consistency.
- Systems, not just models, matter: To use generative AI reliably, it must be part of a larger system (with prompts, retrieval of real data, rule-based components, etc.). The model is a powerful component, but human oversight and complementary tools are needed to ensure accuracy and safety.
I’m Fahim, a software engineer specializing in generative AI, system design, and responsible AI adoption.