This is an overview of all analogies that I have found so far. It includes references to the blog in which the analogy is elaborated. Every main analogy usually includes some additional comparisons on a more detailed level.
| AI Chatbot | Car |
| The various layers of technology and other components that make the whole system work Explain AI like you explain the concept “automobile” to someone from the Middle Ages | |
| The AI Act | Rules of the road, car safety regulations |
| Internet and data | Roads and fuel stations |
| A website or app into which you can type questions | A car that you can drive with steering wheels, accelerator, brakes, etc |
| A transformer (large language model) that converts questions into answers | The engine of a car that converts fuel into motion |
| Operation of the model: arithmetic with words | Operation of the engine: cylinders that creat motion out of fuel |
| Types of AI | Types of vehicles |
| Different ways AI / vehicles are powered. AI is not one thing, but a fleet | |
| Rule-based AI systems such as ELIZA or chess programs | Human-powered vehicles such as bicycles |
| AI based on machine learning (data) | Mopeds with a combustion engine |
| AI based on Large Language Models (chatbots) | Cars |
| AI agents that perform tasks | Work vehicles such as asphalt machines |
| Data for AI | Oil for vehicles |
| How data / oil are used and what effects that has. If data is “The New Oil,” then what is AI? | |
| Slop: new data created by AI that pollutes the internet | Microplastics that arise as plastic breaks down, polluting our environment |
| Explainability of AI | A car repair |
| How to understand the causes (and meaning) of an AI outcome / a car repair Headaches from explainable AI? | |
| AI models that are a ‘black box’ | The mechanism of paracetamol that is not fully understood |
| Reliability of AI | Reliability of Wikipedia |
| To what extent can you trust something that is not produced by a ‘reliable’ process? Trusting AI is dangerous. Not trusting AI, too. | |
| An AI system that does not always give reliable results but is usually (?) better than a human | A self-driving car that causes fewer (?) traffic accidents than a human driver |
| Can AI think? | Can submarines swim? |
| Which words do we use to describe how a machine works Can submarines swim? | |
| How the operation of the human brain has previously been compared with that of machines. | How the operation of machines is described using words that are ‘reserved’ for human traits. |
| Outsourcing human thinking to AI | Outsourcing food production to industry |
| Activities we like to see as ‘typically human’ are still gladly outsourced to technology. Outsourced humanity | |
| Books as technology | Telling stories from memory |
| Chess computer | Chess (which was seen as the ‘ultimate human skill’) |
| Word processors and spell check | Learning to write flawlessly yourself |
| Building an AI language model from training data | Drawing a map based on listening to travel stories |
| How can AI / a cartographer still form a picture of human language / an unknown country even though a computer can never truly understand language / the cartographer has never been there? Back to school: how grammar and topography relate to AI | |
| The co-ccurrence of certain words in the same sentence | The distances between places based on travel time from the stories. |
| A lot of training data. (A whole lot.) | Extensive travel stories. |
| The ‘transformer model’ (inside GPT) that strings words together into sentences. | The navigation app that strings places on the map into a route |
| Energy consumption of AI | Energy consumption of cars |
| Both AI and cars use (a lot of) energy and have a CO2 footprint, but in practice we usually do not know how much, actually. On your bike to go shopping, but still use ChatGPT? | |
| Saying ‘thank you’ to a chatbot costs 0.25 Wh of electricity, so 0.1 grams of CO2 | Roughly 70 centimeters of driving in an average (fuel) car |
| The most nonsensical applications of AI – AI is being put into everything | The most nonsensical applications of engines – engines are being put into everything |
| One ChatGPT question is 7 grams of CO2 | Roughly 50 meters of driving in an average (fuel) car |
| Having AI write a story that would normally take you an hour | Writing a story yourself without AI (but with a laptop) that takes you an hour. |
| The crazy investments in AI | The 2001 internet bubble |
| There are many parallels between the current AI hype and the similarly overheated internet hype a quarter-century ago We are back in the internet bubble of 25 years ago | |
| AI that is built by techies into invisible business applications | The internet being something for nerds used only at universities and big companies |
| ChatGPT turning AI into a major consumer product – you no longer need to be a Python programmer to use AI | Google, broadband and social media turning the internet into a major consumer product – you no longer need HTML to put something online |
| Nonsensical and far-fetched applications of AI, many of which will never succeed | Nonsensical and far-fetched internet services nobody hears about anymore |
| The AI Regulation | Food safety laws |
| For information/food that we put in our heads/bodies there must be rules to protect our mental/physical health What the AI Act has to do with food poisoning | |
| Large supply of cheap or free information services | Large supply of safe and cheap food |
| Mandatory transparency about the training process of AI services | Ingredient declarations |
| AI literacy | Basic knowledge of hygiene |
AI Office and Data Protection Authority and the Netherlands Radiocommunications Agency | European Food Safety Agency en Nederlandse Voedsel en Waren Autoriteit |
| Biased AI | Everyday technology that is inaccessible or designed in a biased way |
| A lot of technologies exclude groups of people, consciously or unconsciously, and sometimes technology is also based on (implicit) prejudice. AI is also a form of technology, and therefore faces the same issues. From crash test dummies to chatbots: old problems in a new guise | |
| AI that helps make technology accessible, by recognizing speech or reading text aloud (positive example) | Technology that is inaccessible to people who are blind, deaf, who stutter, or who have other disabilities |
| Photo software that interprets images of dark-skinned people as gorillas | Blood oxygen meters that do not work properly for people with darker skin |
| An AI model that only hires male applicants because all previous applicants were male as well | Bulletproof vests designed only for men |
| AI that classifies people into categories, such as the COMPAS judicial system in the US | Everyday examples of classifying people into groups: school placement advice, assignment to sports teams |
| Monoculture caused by standardized AI output | Mass media |
| There is fear that abundant use of AI will cause a standardization of language and reasoning. Spoken language becomes written language becomes internet language – let’s pass on AI language | |
| Impact of AI on language | Impact of books on language: writing language has become a separate type of language, as opposed to speaking language |
| An AI Language Model that is pretrained on an unknown selection of texts and that is subsequently finetuned to make it behave | Journalists in mass media selecting facts from the real world, and editorial boards shaping their output and putting those articles together |
| Different AI Language Models from different suppliers | Different channels on TV and different newspapers, from different companies |
AI gedachten | thoughts 