On 14 March 2023, OpenAI released GPT-4, the successor to the model behind ChatGPT. OpenAI reported that it performed far better on professional and academic exams than its predecessor, handled longer documents, followed instructions more reliably and could accept images as input in some settings.
What improved
- Accuracy and reasoning on many tasks, although mistakes remained.
- Longer context, so it could work with more text at once.
- Better writing and code, which made it useful for drafting and debugging.
What did not change
It still sometimes invents information. It does not have a view of the world beyond its training and the tools it is connected to. It cannot replace a teacher's judgement about a particular child.
What it means for learners
AI that can pass exams makes it more important, not less, to test understanding in other ways: asking pupils to explain, to build something, to discuss. Coding is a good example. A learner who can describe a problem, break it into steps and test the result can use AI as an assistant. A learner who only copies cannot tell when the answer is wrong.
That is why we teach the fundamentals first and bring in AI tools after pupils understand what they are looking at.