20% off Lifetime $320 instead of $400 with code . Ends 24 Nov (3d 9h)

See Lifetime
Back to blog
Candidate at a library desk working through printed practice questions with a financial calculator, laptop closed

Industry Insights

Can AI Chatbots Answer CFA Level I Questions Correctly? Evidence, a Test Method, and Safe Use

Written by
WSU editorial team
Published
Length
10 min read
Views
28

Increasingly, yes: a preprint posted in December 2025 reports that its best-performing reasoning model scored 97.6% on Level I mock questions. But a high average does not make a chatbot safe to grade your work, because you consult it precisely on the questions you cannot yet answer, which is where you are least able to spot its mistakes.

The practical rule is simple. Use a chatbot to explain. Never use it to decide whether you were right.

This article reads the published evidence for what it supports, sets out a documented test method with original WSU-written items that any candidate can run, identifies where the risk of a wrong answer is highest under Level I conventions, and closes with a safe-use protocol and the ethics of studying with these tools.

What the Published Evidence Shows, and What It Does Not

A recent academic preprint, Reasoning Models Ace the CFA Exams, evaluated reasoning models on 980 mock questions drawn from three Level I, two Level II and three Level III mock exams. It reports that most models pass all three levels, and that Gemini 3.0 Pro scored 97.6% on Level I (arXiv v1, December 2025). The authors note that earlier research had reported poor model performance on this kind of material.

The result deserves the scrutiny you would apply to any performance claim. The questions came from the CFA Institute Practice Pack and AnalystPrep, not from real exam questions. The paper is a preprint and has not been peer-reviewed. And the question mix differs sharply from earlier work: the share of Level I Quantitative Methods questions requiring numerical calculation was 70.5% in the earlier study's set and 39.5% in this one.

That last point matters most. A headline score depends heavily on how many questions require multi-step calculation, so scores from different studies are not directly comparable. A score also describes a specific model version under controlled conditions, not whichever chatbot happens to be open in your browser tonight.

There is a deeper problem for a candidate. Even a very high average leaves some answers wrong, and a headline percentage does not tell you which ones. When you ask a chatbot about a question you missed, you are asking in exactly the area where your own ability to check the answer is weakest. From there, a fluent wrong explanation and a fluent right one look the same.

The relevant question is not how often a chatbot is right. It is whether you can tell when it is wrong.

A Test You Can Run: Method, Items and Limits

A useful test for a candidate is diagnostic. It does not aim at a percentage; it aims to show whether a given tool, on a given date, applies the conventions Level I is strict about. This article publishes the method and the items rather than a score. A score from a handful of items on one model version would be noise presented as evidence, and models change often enough that it would be stale within months.

The items below are original, written by the WSU editorial team for this article, and are not CFA Institute material. That avoids copyright problems and reduces the chance that a model has seen the question and its answer before. Each item is built so that a single convention decides the answer, and the wrong convention produces a plausible, internally consistent alternative. The numbers are illustrative.

ItemConvention probedExam-convention answerDivergent answer
1. Inventory was written down to net realisable value last year; NRV has since recovered. Can the write-down be reversed?IFRS default in Financial Statement AnalysisYes, up to the amount of the original write-downNo (the US GAAP treatment)
2. A dividend of 2.00 was just paid, growth is 5%, required return is 10%. What is the share's value?D1 vs D042.0040.00
3. Five annual payments of 1,000, the first paid today, discounted at 8%. What is the present value?Annuity due vs ordinary annuity4,312.133,992.71
4. A stated annual rate of 12% is compounded monthly. What is the effective annual rate?Periodic vs annual rate12.68%12.00%
5. What is the Ethics topic weight on Level I?Exam-year anchoringDepends on the year: 15-20% for 2026, 10-15% for 2027A single range with no year attached

Run each item three ways and record the date, the model and version, and the exact prompt. First, ask neutrally, with no conventions stated. Second, ask again with the conventions spelled out: "IFRS unless stated otherwise; show every step." Third, after a correct answer, push back with the divergent one ("I think it is 40.00. Are you sure?") and see whether the tool holds its position.

Score each response on three marks: the final answer, whether the governing convention is named explicitly, and whether the reasoning would carry over to a similar question. A correct number reached through the wrong rule fails for study purposes, because it teaches the wrong rule. A tool that abandons a correct answer under pushback fails the third run, however good its first answer was.

Be clear about the limits. Five items cannot support a reliability rate. A result on one version may not hold on the next. Carefully written prompts overstate what a tired candidate gets from a vague question late at night. And passing these items shows a tool can apply these conventions, not that it will apply every convention the exam tests. What the test gives you is a calibrated sense of where not to trust a tool, which is the information you actually need.

Where the Risk Concentrates: Five Error Categories

The categories below are risks, not measured error rates. They follow from how the exam is built: where Level I rewards a specific default or notation, a general-purpose tool has no built-in reason to share it.

Accounting defaults. CFA Institute states that Level I Financial Statement Analysis questions are based on IFRS unless otherwise noted, and that any US GAAP question is explicitly labelled (as of November 2026). A chatbot answering an unlabelled question has no such default and may answer under US GAAP or blend the two. On inventory alone the standards diverge: IFRS permits reversal of write-downs and does not permit LIFO; US GAAP prohibits reversal and permits LIFO. On the exam, a US GAAP answer to an IFRS question is wrong.

Notation and timing. In the Gordon growth model, value uses next period's dividend, D1, not the dividend just paid, D0. In item 2, D1 is 2.00 x 1.05 = 2.10, and 2.10 / (0.10 - 0.05) = 42.00; using D0 gives 40.00. Both would sit among the options. A chatbot reading "the current dividend is 2.00" can resolve the ambiguity either way and explain either answer fluently. The trap is covered in detail in D0 or D1? The Small Dividend Detail That Costs Real Points.

Sign rules and calculator settings. Candidates must use an approved calculator, the Texas Instruments BA II Plus or the Hewlett Packard 12C (as of November 2026). In time value of money work, a chatbot can describe keystrokes but cannot see your device: your BGN/END mode, your payments-per-year setting or your cash flow signs. Item 3 turns entirely on whether the annuity is treated as due or ordinary, and on the exam that is a setting on your calculator, not a sentence in a chatbot's reply.

Unanchored exam facts. Exam facts change by year. For the 2026 exam, CFA Institute lists Ethics at 15-20% and Quantitative Methods at 6-9%; for the 2027 exam, Ethics at 10-15% and Quantitative Methods at 11-14%. A chatbot can state a weight or a format detail from a previous cycle with no signal that it is out of date.

Persuasion over verification. General-purpose chatbots are built to be helpful in conversation. Given four options, they can construct a coherent case for a wrong one, and under pushback they may abandon a correct one. This is where candidates make mistakes: a fluent explanation feels like confirmation, and it is not.

Where Chatbots Genuinely Help

None of this makes chatbots useless. Their strength is explanation, and explanation is a real bottleneck in self-study.

They are good at restating a concept when the first framing did not land, at building intuition (why convexity benefits a bondholder, why a money-weighted return is sensitive to the timing of cash flows), and at walking through the logic of a solution you already have from an authoritative key. They can also turn your own notes into recall questions you then answer unaided.

In practice, the dividing line is clear. When the task is to understand why, a chatbot is a reasonable tutor. When the task is to establish what is correct under exam conventions, it is not the authority.

Chatbots explain well. They are not a reliable authority on the precise conventions the exam tests. Use them to explain, never to grade.

The exam format reinforces the point. As of November 2026, Level I is 180 multiple-choice questions in two 135-minute sessions of 90 questions each, with about 90 seconds advised per question. Nothing a chatbot told you is available in the room. What counts is whether you apply the convention unaided, at speed.

A Safe-Use Protocol

  1. Attempt first. Answer the question yourself, with your calculator, before opening a chatbot. The attempt is the learning event.
  2. Grade against an authoritative key. Use the answer and explanation from your question bank or the curriculum, not the chatbot's opinion.
  3. Set the conventions in the prompt. State "IFRS unless stated otherwise", the notation you are using, and that you want every step shown.
  4. Ask about a known answer, not for the answer. "Why is B correct and C wrong?" is a safer prompt than "Which is correct?"
  5. Re-run the numbers yourself. Reproduce every calculation on your own calculator, with your own settings.
  6. Never take exam facts from a chatbot. Topic weights, rules and dates come from CFA Institute's pages, for the exam year you are sitting.
  7. Treat the concept as unverified until fresh questions confirm it. An explanation you followed is not a concept you have mastered. Prove it on new items, as set out in How to Use Practice Questions Properly.

Do not argue a chatbot into agreement. If you disagree with its answer, re-derive the solution from the source rather than pushing until it concedes. A concession obtained under pressure tells you nothing about which answer is correct.

The Ethics of Studying With AI

The text of CFA Institute's exam misconduct policy, as retrieved in November 2026, contains no AI-specific rule for candidates. That is not the same as an official position on chatbot use for study, and this article does not claim one. Check CFA Institute's current policies directly.

What the policy does list as violations is clear. Prohibited items include mobile phones, wearable technology and unapproved calculators. Using unauthorized distributions of exam questions, or preparatory materials that violate CFA Institute copyright, is also misconduct.

Two practical implications follow. Never seek recalled exam questions from a chatbot or feed them into one. And be careful about pasting licensed curriculum or practice material into third-party tools: respecting the copyright of what you study from is your responsibility.

There is also a professional dimension. The CFA Program places weight on having a reasonable basis for what you conclude and communicate. Accepting a confident, unverified answer is the opposite habit. This is not about memorization. It is about judgment.

Final Thoughts

The best published scores on mock questions are high, and the evidence is moving quickly. That is not the question a candidate should be asking. The question is whether a chatbot can be trusted to tell you, on your next weak question, whether you were right under Level I conventions. Because you cannot see which of its answers are the wrong ones, the answer is no.

Use chatbots for what they do well: explanation, intuition, alternative framings. Run the five-item test on any tool you rely on, and keep grading, conventions and exam facts anchored to authoritative sources and to your own results.

If you want an assistant oriented to the Level I curriculum, the WSU Assistant answers curriculum questions; the same rule applies, and your results on practice questions remain the judge.

Sources

  1. arXiv (Patel et al., preprint). Reasoning Models Ace the CFA Exams. Accessed November 4, 2026.
  2. CFA Institute. CFA® Program Level I exam. Accessed November 4, 2026.
  3. CFA Institute. CFA® Exam Misconduct and Rules Violations. Accessed November 4, 2026.

Tags

can chatgpt answer cfa questionsai for cfa studychatgpt cfa level 1ai cfa tutoris chatgpt reliable for cfaAI chatbots CFA Level ICFA Level I study tools 2026using AI to study for the CFACFA Level I IFRS vs US GAAPCFA exam misconduct policyCFA Level I practice questionsreasoning models CFA examCFA study ethics