How Random is RANDOM?

 

How Random is RANDOM?

Understanding AI's Decision-Making

This is a challenging question that gets at the heart of how AI actually works versus how we talk about it. Let me break this down:

What "Random" Means in AI

When people say AI makes "random" decisions, they usually mean one of three things:

1. Training initialization randomness
When an AI model is first created, its parameters (the thousands/millions/billions of numbers that define how it processes information) are set to random values. This is like giving a new-born brain completely random neural connections—they have to be trained into something meaningful. Without this initial randomness, every model trained on the same data would learn exactly the same patterns in exactly the same way, limiting their ability to discover different solutions.

2. Sampling randomness (temperature)
When an AI generates text, it doesn't just pick the single most probable next word every time. That would produce boring, repetitive output. Instead, it uses a probability distribution—it considers the top candidates and randomly selects among them based on their probabilities.

Think of it like this: if the next word should be "happy" with 40% probability, "excited" with 30%, "pleased" with 20%, and "content" with 10%, the AI will randomly choose among these options weighted by those percentages. The "temperature" setting controls how sharp or flat this distribution is:

• Low temperature (0.1): Almost always picks the highest probability word (less random, more predictable)
• High temperature (1.0): Follows the natural probability distribution
• Very high temperature (>1.0): Flattens the distribution, making low-probability words more likely (more random, more creative/chaotic)

3. Dropout randomness (during training)
To prevent overfitting (memorizing rather than learning), AI training randomly "drops" or ignores certain neurons. This forces the network to develop redundant pathways and generalize better.

But Here's the Critical Part: It's Not Really Random

Despite all this randomness in how AI is built and how it generates responses, the outputs are deterministic given the same inputs and settings.

If you give an AI the exact same prompt with the exact same "seed" value (a number that initializes the random generator) and the exact same temperature setting, you'll get the exact same response every time. This is because computers can't actually generate true randomness—they use pseudorandom number generators that produce sequences that look random but are completely determined by that starting seed.

The Statistical Probability Part

AI is fundamentally a statistical probability model. At its core, an AI like GPT is doing something remarkably simple (in concept) and incredibly complex (in execution):

It's calculating: "Given all the words so far, what word is most likely to come next?"

But the magic is in the scale:

• It's not just looking at the last few words—it's considering the entire context up to hundreds of thousands of words
• It's not just looking at word patterns—it's built representations of concepts, relationships, reasoning patterns
• The probability calculations happen across billions of parameters in ways that produce emergent behaviours no single parameter encodes

So How Random Is RANDOM?

Type	How Random?	What It Means
True Random	100%	Like radioactive decay—truly unpredictable, can't be reproduced even with same starting conditions. Computers can't do this without special hardware.
Cryptographic Random	~99.9%	Designed to be unpredictable enough for security. Uses system entropy (mouse movements, keystroke timing).
AI Training Initialization	Pseudorandom	Looks random but reproducible with the same seed. Ensures diverse learning paths.
AI Generation (sampling)	Pseudorandom	Weighted randomness that can be controlled via temperature. Same seed = same output.
AI With Temperature=0	0% random	Greedy decoding—always picks the single most probable token. Completely deterministic.

The Deeper Question: Does This "Randomness" Matter?

The concern behind your question might be: "If AI is just making statistically probable guesses, how can we trust its decisions?"

This is exactly what "AI Snake Oil” addresses. The authors argue that we need to distinguish between:

1. Predictive AI (what most people worry about)

• Makes decisions based on patterns in data
• Can encode biases, make errors, be confidently wrong
• Examples: credit scoring, hiring algorithms, risk assessment

2. Generative AI (what's captured public imagination)

• Creates content based on probability distributions
• Has no understanding, just statistical patterns
• Examples: ChatGPT, image generators

The authors' warning is that we treat both as "magic" when they're actually just sophisticated pattern matchers with all the limitations that implies.

The Bottom Line

AI is about as random as a carefully weighted dice that you can control. You can make it perfectly predictable (temperature 0), moderately exploratory (temperature 0.7), or completely chaotic (temperature 2.0). But given the same settings and starting point, it will always roll the same way.

The real question isn't about randomness—it's about whether statistical probability, even at massive scale, can produce something that deserves to be called "understanding" or "reasoning." That's where "AI Snake Oil" makes its most interesting arguments.