NLP · July 2026
For many people, AI and language began in November 2022, when ChatGPT appeared. But anyone who looked something up on Google before then was already using a language model. It’s called BERT, it has been inside the search engine since 2019, and it’s related to the models behind ChatGPT and Claude. Their differences say a great deal about which model you should be choosing for which task.
BERT is a language model from Google that does one thing extremely well: understand what a text says. It reads a sentence in both directions at once, and so it grasps context. In the search “travelling to the US without a visa”, BERT understands that the word “without” changes the entire question; older search technology looked mainly at separate keywords. That is why Google built the model into search in 2019, and since then almost every query has passed through it.
What BERT doesn’t do is write text of its own. The model reads, recognises and judges. You give it a piece of text and it tells you what that text is about, what sentiment sits in it, or which category it belongs to. For my master’s thesis I trained a BERT model myself to recognise hate speech in tweets. That could be done on an ordinary computer, because BERT is small enough to run yourself and to steer towards a specific task with a few thousand examples.
ChatGPT and Claude come from the same family. They are built on the same underlying technique, the transformer, and they learned language the same way: by working through enormous quantities of human text. The difference is that their task runs the other way round. Where BERT reads and assesses, these models predict the next word each time, and out of those predictions new text emerges.
That makes them far more broadly usable: they write, summarise, translate and think things through with you. It also makes them far larger, costlier and hungrier for energy, and their output is open. You don’t know in advance exactly what will come out, and a convincing answer isn’t necessarily a correct one. In machine learning or generative AI you can read what that difference means for the risks you run.
The family resemblance runs deeper than the marketing suggests. Both kinds of model learned from text written by people, and so both take on the patterns sitting in that text, the skewed ones included. My thesis research showed that concretely: a model meant to recognise hate speech can itself judge with prejudice against the very groups it ought to protect. That lesson applies just as much to today’s large generative models.
The real difference is in size and predictability. A model like BERT is specialised and compact: you can run it on your own servers, your data stays in-house, and the same message gets the same judgement today and tomorrow. The large generative models run in a provider’s cloud, cost money and energy per use, and phrase things slightly differently every time.
This is where it gets practical for your organisation. Many of the tasks companies now point a large generative model at are really reading tasks: routing incoming email to the right department, flagging complaints that are urgent, filtering comments on a forum. For work of that kind a small, specialised model is often faster, cheaper, leaner and easier to check. You can see how much that saves in climate impact in my AI CO₂ calculator (in Dutch).
The question that really matters: what does your task need? Does something have to be read and assessed, or does something have to be made? Ask that first and you choose a tool that fits. Want help with it? Have a look at tailored AI solutions, or email info@kimberleyvanruiven.nl.