Think of it this way: a master chef, a master architect, and a master surgeon are all "experts" — but you would not ask your surgeon to design your kitchen. AI systems are similar. Each is highly capable within its domain and essentially useless outside it.
Trained on hundreds of billions of words — books, articles, code, websites, conversations. They learn the statistical relationships between words and ideas well enough to generate fluent, contextually appropriate language. They can answer questions, summarize documents, write essays, translate, and write working code.
⚠️ Critical nuance: language models don't "look things up." They generate responses based on patterns learned during training. This means they can produce fluent, plausible text about things that aren't true — because fluency and accuracy are separate properties.
ChatGPTClaudeGeminiCopilotLlama
Trained on millions of labeled photographs. These systems can classify objects, faces, scenes, and subtle visual features with accuracy that often exceeds trained humans — especially in narrow domains like reading medical scans or detecting manufacturing defects.
⚠️ Critical nuance: these systems can be brittle in unexpected ways. A small deliberate change to an image — invisible to humans — can completely fool a recognition system. They're pattern-matching in ways that don't always align with how humans see.
Face IDGoogle PhotosMedical imaging AIQuality inspection
Trained on enormous collections of images paired with text descriptions, these systems learn the relationship between words and visual content. Given a description, they generate a new image — not by retrieving an existing one, but by constructing something new from learned patterns.
⚠️ Critical nuance: the same technology that generates creative images can generate convincing fake images of real people doing things they never did. Understanding this is part of being a thoughtful consumer of visual media.
MidjourneyDALL-EStable DiffusionAdobe Firefly
The oldest and most commercially widespread form of AI. Trained on historical data to predict future outcomes — credit scoring, fraud detection, medical risk, demand forecasting, targeted advertising. Running quietly in the background of most major industries.
⚠️ Critical nuance: these systems are particularly susceptible to inheriting historical biases, because they're literally trained to replicate past patterns — including patterns that reflect historical inequities in lending, hiring, and criminal justice.
Credit scoringFraud detectionMedical risk modelsAd targeting
Rather than learning from labeled examples, these systems learn by doing — taking actions, receiving feedback about whether those actions were good or bad, and improving. This produced AI that mastered chess, Go, and complex video games at superhuman levels — without being told the right moves.
💡 Key insight: this approach excels at problems where you can define what "winning" looks like, even when the path to winning is completely unknown. It's increasingly used in robotics and complex optimization.
AlphaGoRoboticsData center optimization
Trained on recordings of human speech across languages, accents, and acoustic conditions. They transcribe speech to text with accuracy that matches human transcriptionists. They also convert text to speech with natural cadence — and can now clone a specific person's voice from just a few seconds of audio.
⚠️ Critical nuance: voice cloning technology makes audio fraud dramatically easier. A convincing phone call from a "family member" asking for help can now be synthesized in seconds. This is a real and growing threat.
SiriAlexaWhisperElevenLabs
Narrow vs. general — the most important distinction
The spectrum from narrow to general AI
General AI
(doesn't exist yet)
Narrow specialistGeneral intelligence
Every system deployed today sits firmly on the left side of that spectrum. A language model that writes beautifully has no idea how to identify a tumor in an X-ray. An image recognition system has no ability to hold a conversation. When people speculate about AI that could match human capabilities across all domains, they're describing something that doesn't exist — and nobody knows when or whether it will.
"The word 'AI' covers wildly different systems. Understanding which type you're dealing with tells you most of what you need to know about when to trust it — and when to be skeptical."