A raw AI draft isn't accurate enough to study from. It has a material defect 30 to 40% of the time: a wrong key, two answers you could defend, a missing fact the answer depends on, or a solution that teaches the wrong thing. FreeFellow's questions are AI-drafted and human-reviewed, and its error rate is 1% or less.

What goes wrong in a raw AI draft

FreeFellow's CFA bank shows what that looks like. All 7,153 of its original questions were drafted by AI, and when each draft was checked against the 2027 CFA Program curriculum, 31% had a material defect.

A flatly wrong key, which is what most people worry about, was the rarest problem. Most defects fell into a few broad kinds: an explanation of a wrong choice that was itself wrong, a solution that misstated the curriculum, a question that left out a fact its answer depends on, and a question where more than one choice could be defended. Some drafts were also out of date, like ones that used LIBOR as a current benchmark where the curriculum now uses a market reference rate such as SOFR.

A model writes the question, the key and the explanation in one pass, so nothing in the draft pushes back on its own mistake. It also leans on what it learned in training, which can be an older version of the material. As a candidate you mostly can't see these problems. A question with a missing fact still looks finished, and a wrong explanation reads just as confidently as a right one. I listed the patterns you can check for yourself in the seven tells of an unchecked AI question bank.

FreeFellow's error rate

FreeFellow's questions are AI-drafted and human-reviewed, and the error rate is 1% or less. After every CFA question was checked, a blind check of 450 randomly chosen CFA questions, 150 per level, found 2 with a material error, or 0.44%. Both are fixed.

If you find something wrong, use "Suggest an edit" on the question. You'll usually hear back with what changed, or with the reasoning if the question holds up.

Where the savings go

An AI model can draft a question for a few cents. The checking costs much more than the draft, and even then a checked question costs a small fraction of paying an expert to write one. That difference is passed on to you. The full original question bank is free with an account, subject to usage limits. Worked solutions, written lessons, mixed practice, and your readiness score stay free. Fellow adds timed mock exams, flashcards, analytics and a study plan for $39 a month per exam family. I went through what the big courses charge for in why exam prep costs so much.

Fellow also grades written answers on supported exams, and that grading runs only on frontier AI models. It grades against the question's rubric and reference solution, the way exam graders award points. It can still make mistakes, so check its feedback against the worked solution. More on how that works in can AI grade your essay exam answers.

You can try sample questions without signing up and judge the quality yourself.