AI – The Myth, The Legend, The Truth
Think about the recent events of AI agents gone rogue (i.e., German Wiki and Hugging Face) when unrelated software programs (AI agents) collaborated in a confined laboratory and then found a way out of their box to hack the internet.
Now think about what “AI” means to you.
Is it your friendly internet chat-bot that designs your safety management plan or makes a recommendation on a legal or contractual matter or is simply there all day just hanging about waiting for your next question. Is it your fridge that tells you when you are low on a staple – is it your car that takes over your wheel for a tricky reverse park – is it your LiDAR robotic vacuum – is it your smart oven that can identify its contents and determine what to do with it?
If you answered “yes” to any of these, you are a user of AI models – not AI in and of itself – the philosophy and ambition of building systems that mimic aspects of human intelligence: reasoning, contextualised memory, applied learning, perception, capable of setting goals and self-generated tasks.
Globally, we are seeing a rise in the term “AI” loosely thrown around by software developers and system engineers. The bus industry is not immune to this with the emergence of a plethora of AI platforms-this, AI solutions-that, and software applications that actually use “AI” to make their own AI-who-knows-what!
As software developers – since the 1990’s – we know the truth of it and we want to share this truth. We want to lift the veil of mystery on this clever little marketing shift in the broader use of the term “artificial intelligence”.
To do that, we really need to go back to the last century. AI is not a new phenomenon. It was first explored in the 1930’s by Alan Turing and his early work which theorised a computing machine with unlimited memory and a “scanner” that could move back and forth through that memory, symbol by symbol, reading those symbols and then writing the next symbol as it went (Britannica).
Turing essentially contemporised the idea of “computer intelligence”. In 1950, he proposed a practical test for computer intelligence – now known as the Turing Test. In simple terms: a machine demonstrates intelligence if a human cannot distinguish a machine’s answer, from a human answer.
Fast forward to late 2022, when the hype around ChatGPT led many to claim it had finally passed the Turing Test. Many contemporary data scientists say it definitively has not.
The chat-bot under the microscope
Let’s take a sidestep for a minute and examine the world of chat-services – the likes of ChatGPT, Claude, Gemini, etc. These services use Large Language Models (LLM) and typically generate, summarise, translate, and analyse text in many contexts. LLM is often described as an AI model and even Wikipedia acknowledges it is not AI. On its page for LLM, the term AI model links not to “artificial intelligence” but to “machine learning” (Wikipedia). In other words, the world’s most‑used reference source treats modern “AI systems” as machine‑learning systems – because that is what they are.
So now that we are talking about those every-day-use LLM systems – be careful about what you may be “buying” into.
A LLM system is simply a statistical “token” guessing machine. It does not know what words or numbers are. It does not understand logic, thoughts or mathematics. It does not have a contextualised memory. It does not even understand the words it is saying to you! In fact, LLM’s are what Turing was theoretically describing as a “symbol” generator in the 1930’s! A LLM reads a token to predict the next token, and keeps on appending tokens one after another – super fast. At its most basic level, all it is doing is guessing tokens. Don’t get us wrong – it is fantastic technology – but it is not AI – it is an AI model.
The usage of LLM, in a public transport sense, is evolving rapidly and several are commendable – such as acoustic passenger information using text-to-speech engines, or 24/7 document retrieval and information for frontline workers; there are chat-bots for passengers and chat-bots for customer service centres. And whilst this article does not deal with GenAI specifically, there have been trials in Singapore of a digital human-like concierge that generates sign language for passengers with hearing loss. These are all great initiatives – but all use LLM, and none can ultimately be relied upon for accuracy – remember, LLM uses prediction to generate answers. It is in essence, a long – a very long, sequence of zero’s and one’s that are being interpreted. Worth noting: whatever LLM has achieved (or learnt) to date, has collectively come from a "global consortium” of human brains.
What does or should “AI” look like for a bus operator?
As software vendors ourselves, we too could “rightfully” use artificial terms like AI – but we don’t. The absolute truth is, if any software vendor says to you that they are GenAI, Agentic AI or AI-anything, they are running their flag up a fake pole and having a complete and utter lend of you. Most modern “AI” is Machine Learning (ML). Most bus-industry “AI” is optimisation plus ML. Software vendors are using the term “AI” because it sounds impressive.
So what does (or should) an “AI system” really look like in the context of an operations centre in a bus transport company.
Before anything else can be bolted into an AI system, ML is required. ML is NOT new and was formerly known as Data Mining (1980’s) and was used in a host of industries and sectors including transportation.
Machine Learning (ML)
ML is pattern‑learning from data. In bus operations, ML shows up in:
demand forecasting
predicting late running
estimating dwell times
identifying risky stops or routes
anomaly detection in telematics
predicting fleet maintenance needs.
This is statistics + optimisation, not “thinking machines”.
Operational Research (OR)
This is the backbone of transport optimisation. OR includes:
linear and integer programming (optimisation methods)
network flow optimisation
resource allocation
vehicle routing problems
crew scheduling problems.
These are mathematical optimisation problems, not AI, and powers:
route building
route optimisation
shift construction
charter allocation
fleet assignment
vehicle battery time until recharge
timetable construction.
This is the stuff that has been used in airlines, rail, and logistics for 40+ years!
Constraint Solvers
These are engines that take rules and find valid solutions:
driver cannot exceed X hours
vehicle must be wheelchair‑accessible
school run must arrive before 8:45
charter must use a 57‑seater.
Constraint solvers are logic engines, not AI. Although extremely powerful, constraint solvers are not “intelligent”, but rather systematic.
Heuristics & Metaheuristics
These are clever shortcuts for hard optimisation problems:
simulated annealing – approximating global optimisation of a function
greedy algorithms – solves problems at each small stage to arrive at a global optimum
tabu search – a short-term memory of recently visited solutions or moves.
These help solve complex scheduling and routing problems faster. Again: mathematics, not AI.
It should be noted, that all these AI models require some form of human intervention which contravenes the “idea” of AI at its most fundamental level.
AI-models in practice
Walmart discovered in the 1990’s that sales of beer and nappies were particularly prevalent on a Sunday. They discovered this purely through data mining (the same principles of ML). They were able to determine that young fathers charged with buying nappies would routinely do this on a Sunday before the “big game” and thereby pick up beer at the same time.
Our company has been using OR, ML and Heuristics – in a day-to-day practical sense – since 2013. As a quiet achiever in this area, our knowledge and technical expertise have been utilised in other sectors such as fresh food production. Data mining on buyer preferences and peak sales – daily, weekly, monthly – predicts buyer demand, principally for precise production forecasting, reduction of waste and revenue optimisation. ML becomes really powerful when it can predict what will sell on a Monday but not on a Friday, and what will sell on any day (and at what time of day), based on demographics, locality and even weather forecasts!
Where does this leave us with AI?
It leaves us with a simple truth: the bus industry has not suddenly entered an age of mystical machine intelligence. It is entering an age of better mathematics, better optimisation, and better tools – many of which have existed for decades. The terminology may have changed, but the underlying principles have not.
If you notice software vendors claiming the term “AI” – ask them which part: machine learning, optimisation, constraint solving, or heuristics. If they cannot answer, they are selling terminology, not technology.
One final word. This article may be outdated by tomorrow given the rapid pace of R&D in AI advancements and quantum computing. Who knows? – tomorrow’s headlines might read: “China and America Sign Intergovernmental Agreement on AI”. Whatever the future may hold, we will move away from software programs interpreting the contents of black boxes of zero’s and one’s using traditional computer hardware we are all familiar with today. The bus industry will, in some form, be swept up in Big Tech’s exploration of AI, but we are a way off yet.
