My grandmother, towards the end of her life, lived in a care home in a village in Gujarat. She spoke Gujarati and a little Hindi. The care workers were Indian, but many of them were from other states — Kerala, Tamil Nadu, Bengal — and they didn't share her language.
For the last two years of her life, the thing that let her have conversations with the people caring for her was a small app on a tablet that translated in real time. Imperfectly, sometimes comically. But well enough that she could tell them she was cold, ask about their children, and share the opinions of which, in her opinion, the staff were underappreciating her.
That app was built on natural language processing — the field of AI that teaches machines to work with human language. I want to talk about what NLP is, how it got good, and where it's quietly changing lives.
What NLP is actually trying to do
Human language is slippery. The same sentence can mean different things depending on tone, context, who's saying it, and what happened yesterday. For decades, computer scientists tried to tame it with rules — dictionaries, grammar parsers, carefully crafted logic. It sort of worked. Not really.
The breakthrough came when we stopped trying to tell machines the rules of language and started showing them a lot of language and letting them learn the patterns. The Rosetta Stone but with a trillion parallel sentences.
Today's NLP systems don't "understand" language the way humans do. But they've learned statistical regularities deep enough to translate between languages, summarise articles, answer questions, correct grammar, and have a remarkable proportion of the conversations we have with computers every day.
Where NLP is quietly working
Translation. Google Translate serves more than a billion translations a day. The quality is now good enough that it's changing how people travel, how diasporas stay in touch with home, and how small businesses sell across borders. It's imperfect — every translator will tell you about moments where the machine missed something crucial — but it's gotten good enough to be life-changing for people like my grandmother.
Autocomplete and autocorrect. Every time your phone predicts the next word, that's NLP. Modern systems are personalised to your typing patterns, context-aware, and multilingual. They don't feel clever because they're so well-integrated — which is the hallmark of good technology.
Search. When you search for "best Italian restaurant near me," the search engine has to understand that you mean "serving Italian food," not "in Italy." The ranking that puts the right restaurant first is NLP-driven.
Spam filtering. The reason your inbox isn't drowning in Nigerian prince emails is because NLP filters got incredibly good, mostly behind the scenes. This is probably the highest-value NLP deployment in the world, and nobody thanks it.
Voice assistants. "Hey Siri" has to understand what you meant, even when you mumble, even when there's background noise, even when you used an expression the machine hasn't heard exactly before. Remarkable.
Accessibility. Live captioning for deaf users. Text-to-speech for blind users. Plain-language rewriting for people with cognitive differences. These tools have transformed daily life for people who were poorly served by previous generations of technology.
Where it's still hard
Low-resource languages. The big language models are trained on trillions of words — most of them in English. For languages with less digital presence — many African, South Asian, and indigenous languages — the tools work much less well. Closing this gap is one of the most important and least glamorous challenges in the field.
Nuance, sarcasm, code-switching. Humans routinely mix languages mid-sentence, use sarcasm, speak in half-references that require cultural knowledge to parse. NLP systems still stumble on all of this. They're getting better, but honestly — so are humans, and we've had a head start of some millennia.
Context over long stretches. Modern systems are vastly better at context than they were ten years ago, but they still lose track over very long documents, or across sessions, or when context comes from outside the text.
A note on the LLMs
Large language models — GPT, Claude, Llama and their cousins — are NLP turbo-charged. They're trained on essentially the entire readable internet, and they've produced capabilities that were science fiction five years ago.
They're also not magic. They make things up. They have strong opinions that sometimes reflect biases in their training data. They don't know things the way a librarian knows things; they model the patterns of how humans write about things.
Treating them as statistical writers rather than oracles is the mental model that keeps you out of trouble. Used that way, they're extraordinary. Used as search engines, they'll confidently invent plausible lies.
Where I keep looking
The NLP deployments that move me most are the ones at the edges: accessibility, translation across minority languages, literacy tools for kids in underserved schools. The places where language used to be a barrier and now isn't, quite.
My grandmother would have liked those stories. She was a great believer in the idea that language should never be what keeps us from understanding each other.
If you're thinking about language technology for your own product — or wondering whether the tools are good enough for your specific application — we'd love to help you think.
