The most famous bar candidate in America gave this topic its cautionary tale. After failing the July 2025 California bar exam, Kim Kardashian told interviewers that she'd been photographing practice questions into ChatGPT for answers — and that the chatbot was "always wrong": "It has made me fail tests all the time." Legal-education commentators piled on about a "dangerous trend," and the jokes wrote themselves.
The jokes are cheap, though, because the underlying question is serious and the honest answer is genuinely two-sided. The same technology that fails as an answer machine is quietly excellent at several real prep tasks. Here's the split — what AI is actually good for in bar prep, where it will reliably hurt you, and the workflow rules that keep you on the right side of the line.
The paradox you have to hold
Both of these are true. In 2023, GPT-4 made headlines for passing a simulated uniform bar exam — the famous "top 10%" claim, which a later peer-reviewed re-analysis argued was closer to the middle of the pack against first-time July takers, but a pass either way. And yet candidates who lean on chatbots as answer machines get burned, publicly and repeatedly — including practicing lawyers sanctioned for filing briefs citing cases that never existed.
The resolution: a model passing one exam under lab conditions is not the same product as a model reliably teaching you correct law on demand. Language models generate plausible text; on well-trodden doctrine they're often right, and when they're wrong they are wrong in exactly the same confident voice. An occasionally-wrong authority that sounds identical when wrong is worse for a precision exam than no authority at all — because the whole game on MCQs is distinguishing nearly-right rule statements, and a chatbot's hallucinated version of a rule is a distractor, delivered as truth, memorized before you ever meet the real one.
Where AI genuinely helps
Used as an assistant rather than an authority, a chatbot earns its place in four prep tasks:
- Explaining a concept you already have correct. When the rule is in front of you from a verified source and it isn't clicking, "explain the Erie doctrine like I'm ten, then again like I'm a 2L" is a legitimately great use. The model is elaborating your material, not supplying its own — comprehension help, not content.
- Socratic self-quizzing on your materials. Paste your own outline section and ask to be quizzed on it. Because the source of truth travels with the prompt, the failure mode is capped — and retrieval practice is the mechanism that builds recall, so being interrogated beats rereading by a mile.
- Hypo generation for issue-spotting reps. "Give me a fact pattern that mixes hearsay and Confrontation issues" produces serviceable practice at unlimited volume. The hypos are imperfect and occasionally incoherent; for spotting practice, that's tolerable in a way that wrong rule statements never are.
- Study logistics. Summarizing your own notes, restructuring a week, drafting a review checklist — clerical work where errors are cheap and visible.
Notice the pattern: every safe use has you supplying the law and the model supplying the packaging. Every dangerous use inverts it.
Where it will hurt you
1. Rule statements from the model's memory. The elements come back subtly off — a factor merged, an exception dropped, a majority rule swapped for the minority — in flawless legal register. You can't audit what you haven't learned yet, which is precisely the state you're in as a student. This is the single most damaging use, and it's the default one.
2. The stale-scope problem — the NextGen-specific trap almost nobody flags. Language models learned the bar exam from a decade-plus of internet writing about the legacy exam. Ask one about "the bar exam" and you'll get confident MEE essay templates, 200-question MBE arithmetic, subjects that left the exam, and nothing reliable about select-two scoring, integrated question sets, or the 500–750 scale. Ask about "the NextGen bar exam" and you'll often get legacy answers wearing new labels — or inventions. The exam changed in 2026; the training data mostly didn't. For format questions, deadlines, and state adoption facts, the only sources that count are NCBE's official materials and pages verified against them.
3. Grading your written answers. Chatbots are agreeable by construction. Feed one your performance-task answer and it finds things to praise, invents a rubric, and calibrates to nothing — while real graders score against real rubrics with points attached to specific issues. AI feedback on writing feels like coaching and drifts you toward confident mediocrity. Self-grading against a model answer and issue rubric is slower and actually works.
4. The outsourcing trap — the Kardashian workflow, named precisely. Photographing a question into a chatbot practices asking. The exam grades answering. Every rep you outsource is a retrieval attempt you didn't make, and retrieval attempts are the entire mechanism by which rules become producible under pressure. Do this at volume and you can complete an entire question bank while acquiring almost nothing — with a chat log that looks exactly like studying.
The five workflow rules
- Never accept a rule statement you haven't checked against a verified source. Not "usually check." Never. The confident wrong ones are indistinguishable by design.
- Bring your own law. Paste the outline, the rule text, the explanation — then let the model explain, quiz, and elaborate. Source of truth in the prompt, or no prompt.
- Keep AI out of timed practice entirely. Practice scores only mean something if they measure you. A "diagnostic" taken with an assistant measures the assistant.
- Treat AI grades as vibes, rubrics as grades. Use model answers and issue-by-issue rubrics for written work; let the chatbot rephrase a rubric line you didn't understand, never author the verdict.
- Route every miss to a real rep. If a chat explanation genuinely unlocked a concept, prove it: ten practice questions on that rule, closed-book, this week. Understanding that can't survive a question wasn't understanding.
Where we stand, since you're reading this on a prep site
BARGO's position is published in our AI policy and it's the inverse of the failure modes above: our practice questions, explanations, and flashcards are written by our team — mapped to NCBE's content scope, with jurisdiction facts verified against primary sources — and AI helps us build and run the website, not author the law. On a precision exam, the content layer is exactly where accountability has to live: when a rule statement decides points, someone with a name should have stood behind it.
That's the honest close, too. AI in bar prep is a competent assistant and a catastrophic authority — a tireless explainer of law you supply, and an unreliable narrator of law you don't. Hold that line and it will genuinely help; blur it and you'll be quizzing a very fluent stranger. Start on the right side: ten free diagnostic questions, written and explained by humans, timed at the real pace — and use the machine to help you understand whatever you miss.