GPT-4: the model that could pass the bar exam

2023

In March 2023, OpenAI released GPT-4, a model so capable that its benchmark scores read like a top student's transcript. Bar exam: 90th percentile. SAT: 93rd percentile. GRE Verbal: 99th percentile.

These weren't just numbers. They represented genuine reasoning ability: understanding complex legal scenarios, analyzing evidence, constructing arguments, and applying precedent. The kinds of cognitive tasks that defined professional competence.

For the legal profession, the implications were seismic. If an AI could pass the bar exam, what did that mean for junior lawyers? For legal research? For access to justice for people who couldn't afford attorneys?

But GPT-4's impact went far beyond any single profession. It was the first model that felt genuinely useful for complex, multi-step reasoning tasks. Write a business plan. Design a software architecture. Analyze a scientific paper. Plan a project.

The model also introduced multimodal capabilities: it could understand images, not just text. Upload a photo of a math problem and it would solve it. Screenshot an error message and it would diagnose the issue.

GPT-4 didn't just improve on GPT-3.5; it qualitatively changed what people believed AI could do. The question shifted from "Can AI be useful?" to "What can't AI do?"