AI Insights · Foundation

From Algorithms to AI: The Honest History of What Changed

For most of the past decade, "AI" was a marketing term applied to things that barely deserved it. This article is a plainly honest account of what AI actually was, what changed, and why the shift since 2022 is genuinely different from what came before.

If you have been sceptical of AI claims from vendors over the years, your scepticism was largely justified. Recommendation engines, spam filters, and fraud detection systems were all labelled AI at various points. Most of them were algorithms: structured sets of rules, or statistical models trained on labelled data, doing a narrow job well. Useful, certainly. But not what anyone would recognise as intelligent.

The reason this history matters is practical. Understanding what changed — and when, and why — helps you separate genuine AI capability from rebranded software. It also explains why the AI you can deploy today behaves in ways that earlier "AI" did not, and why governing it requires a different approach.

The Algorithm Era: Rules and Decision Trees

Before machine learning became widespread, most AI was rule-based. A system was given a set of conditions — if this, then that — and it followed them. Spam filters checked for known patterns. Fraud detection flagged transactions that matched certain criteria. Recommendation engines used relatively simple scoring models.

These systems worked, but they were brittle. They could only handle situations their designers anticipated. Add a new type of spam and the filter missed it until a human added a new rule. They did not learn. They did not generalise. They were sophisticated if-else statements, and calling them AI was generous.

Machine Learning: When Systems Started Learning from Data

Machine learning changed the approach. Instead of writing rules by hand, you gave the system examples and let it find the patterns itself. Feed it thousands of labelled images of cats and dogs, and it would learn to distinguish between them. Feed it historical loan repayment data, and it would learn to score credit risk.

This was a genuine advance. But it came with significant constraints that are still widely misunderstood:

• Narrow scope: A model trained to classify images could not write text. A model trained to detect fraud could not answer questions. Each model did one thing, and only that thing.
• Labelled data dependency: You needed large volumes of correctly labelled training data. This was expensive and time-consuming to produce.
• Feature engineering: Data scientists had to decide which variables to feed the model. The model's performance was only as good as the features humans thought to include.
• Interpretability: As models became more complex — particularly with deep neural networks — understanding why they reached a particular conclusion became increasingly difficult.

Machine learning underpins a huge number of genuinely valuable products: image recognition in smartphones, personalised content feeds, medical imaging analysis. But it did not produce systems that could understand language, hold a conversation, or reason across topics. The gap between ML and what people imagined AI would be remained large.

The Timeline: What Happened and When

What Actually Changed with Large Language Models

The question worth asking is: what specifically is different about LLMs compared to everything that preceded them? There are three things that matter for business.

Generalisation across tasks. A single LLM can draft a contract summary, answer a customer query, classify a support ticket, write a function in Python, and translate a document — without being retrained between tasks. No earlier approach came close to this level of generalisation. This is not an incremental improvement. It is a different category of capability.

Language as the interface. Before LLMs, interacting with software required learning its interface: its menus, its fields, its query syntax. LLMs take natural language as input. This removes a barrier that has existed since the first computer. A business user can now instruct an AI system in plain English.

Emergent reasoning. At sufficient scale, LLMs exhibit reasoning behaviours that were not explicitly trained. They can work through multi-step problems, identify logical inconsistencies, and produce structured analysis. This is not fully understood even by the researchers who build these models — which is one reason governance matters as much as capability.

The Comparison in Plain Terms

Why the Buzzword History Matters Today

Many businesses are still treating AI the way they treated earlier "AI" — as a feature to be switched on, producing reliable, rule-like outputs. That mental model does not fit LLMs.

An LLM is not a rule engine. It is a probabilistic system. Given the same input twice, it may produce slightly different outputs. Given an ambiguous input, it will fill the gaps with plausible-sounding inference rather than flagging uncertainty. It can be wrong while sounding exactly as confident as when it is right.

This does not make LLMs unsuitable for business. It makes the way you deploy them — the prompt design, the validation steps, the human oversight at key decision points — more important than the model itself. Platform knowledge (understanding how the AI connects to your data and processes) and governance (knowing what the AI is doing and checking its work) are not optional extras. They are where the value is.