AI as a Prediction Machine
Definition
AI does not execute logic or look up facts — it predicts the most statistically likely continuation of whatever input it receives. Understanding this is the foundation for understanding both AI’s power and its failure modes.
How It Actually Works
The workshop uses a simple sequence to demonstrate:
“Better” → “Better late” → “Better late than” → “Better late than never.”
Each word is a prediction: given everything before it, what is the most likely next word? The model does not “know” the phrase. It has seen millions of text sequences and learned which words follow which, with what probability.
This is not a metaphor. It is the literal mechanism. Large language models generate every word, sentence, and paragraph the same way: statistically likely continuation of the context.
Programs vs AI
| Programs | AI | |
|---|---|---|
| How they work | Execute based on logic and rules | Predict based on statistics and probability |
| 2 + 2 = | 4 (deterministic) | “Based on training data, 99.7% probability: 4” |
| How you deploy them | Install / program | Onboard / nurture / train |
| Failure mode | Crashes or returns an error | Produces confident, plausible, wrong output |
The last row is the critical one. A program that cannot answer a question fails visibly. AI that cannot answer a question produces a confident-sounding answer anyway — because its job is to generate a plausible continuation, not to flag uncertainty.
Why This Is Not a Bug
Plausibility optimization is why AI is useful. It is also why it hallucinates. The same mechanism that produces fluent, coherent, contextually appropriate text also produces fluent, coherent, plausible text that is factually wrong. These are not two different systems — they are the same system in different situations.
This means the leadership question is not “can we stop hallucination?” It is “where does the cost of a plausible-but-wrong output exceed the benefit of speed and scale — and how do we put a human in that gap?”
The Scrambled Text Insight
Slide 46 uses a well-known perceptual trick: scrambled words where only the first and last letters are in the correct position. Humans read them easily because our brains predict meaning from pattern and context, not from letter-by-letter decoding. AI reads text similarly — statistical pattern matching at enormous scale. The difference is that humans have grounded experience of the world. AI has statistical co-occurrence of tokens. Both are powerful. Only one knows what a sandwich actually is.
What to Pay Attention To
- Where AI output is being used without a human verification step for factual claims
- Where the fluency of AI writing is being mistaken for accuracy
- Where the model’s confidence is being treated as a signal of correctness
- Where people have stopped questioning AI output because it “sounds right”
Connections
Hallucination as Plausibility Optimization Programs vs AI Jagged Frontier LLMs as a Language Revolution
Sources
- Mollick and Mollick - Assigning AI
- [inferred from workshop teaching — prediction mechanism is well-established in LLM literature]
Tags: prediction, probability, LLM basics, how AI works, statistical learning