It can, as long as it's grading against a real rubric. Given the grading standard, a model like ChatGPT or Claude is the closest thing to a personal grader most candidates will ever have. That's why FreeFellow grades answers against rubrics built from published solutions and grader commentary, and why it turned every released SOA Exam PA sitting and all 13 released CAS Exam 5 sittings into graded walkthroughs. Without a rubric, the same model gives good scores to answers that sound fluent, whether or not they answer the question.
I think written-answer practice is where AI helps candidates the most, as long as you set it up right.
Why feedback matters so much here
With multiple choice, you know right away whether you got it right. Essays and written-answer exams (CFA Level III essays, SOA Exam PA, CAS written-answer papers, CPA simulations) used to leave you two options: pay a grading service real money for feedback on a few essays, or practice with no feedback at all. Most candidates did the second, which means practicing blind on the format that fails the most people. It's hard to notice on your own when an answer doesn't respond to what was asked, and a grader spots it right away.
Given the question, the rubric and your answer, a model applies the standard consistently. It doesn't get tired, so your twentieth practice essay gets graded as carefully as your first. It's also specific in the way that makes feedback useful: which rubric points you earned, which you missed, and what sentence would have earned the missed ones.
Speed helps more than you'd expect. Feedback that shows up in seconds, while you still remember your reasoning, changes how you write the next answer. Feedback from a grading service two weeks later is much less useful.
Where it goes wrong
Ask a chatbot to "grade my essay answer" with nothing else and it'll grade on how fluent and plausible the answer sounds. Language models are trained to be agreeable, so without a standard to grade against, they go easy on you and give good scores to well-written answers that never do what the command word asked. Exam graders do the opposite. They look for specific items on the rubric and give nothing for good writing.
The fix isn't a clever prompt. Give the model the actual grading standard and tell it to award points only for rubric items that are clearly in your answer, quoting the sentence that earns each point. Done that way, most of the generosity goes away, and what's left is the real gap between what you wrote and what the graders reward.
You can't make up a usable rubric in a chat window. It has to reflect how the exam really awards points, so it has to come from the released material: published solutions, sample graded answers and grader commentary. For the actuarial written-answer exams, the released sittings make this concrete. FreeFellow's walkthroughs of every released Exam PA sitting and all 13 released CAS Exam 5 sittings each include the rubric, the published sample answers and model solutions, and the grader commentary that explains what full credit took. CFA Level III essay practice works the same way, with original questions written to the published grading style.
How to do it for free
FreeFellow's free tier has a copy-to-AI prompt builder on written-answer practice. It puts the question, the rubric and the model solution into one prompt, so you can paste your answer into your own ChatGPT or Claude and get graded against the rubric at no cost. Fellow includes five AI-graded attempts a day in the app, graded by Claude against the same rubric. Fellow Plus removes Fellow's 5-attempt daily allowance on exams with AI grading. Grading rate and usage limits still apply. Every annual plan is Fellow Plus.
Here's the routine I'd use. Write your answer under exam time pressure first, before looking at any solution. Get it graded. Rewrite the answer to pick up the points you missed. Then read the grader commentary and note what the exam rewards that you didn't expect. Most of the value is in the rewrite, so don't skip it.
AI grading helps you practice writing answers, but it can't recreate exam-day time pressure, hand fatigue on a paper exam, or the exam body's actual grading decisions in a given sitting. Treat any practice score, from AI or from a person, as a direction to work in, not a prediction. Only the exam body grades the real thing. You can start with the free constructed-response practice.