Should You Use AI to Write SIE Practice Questions? (Examples Included)

Why is this tempting?

You can ask ChatGPT (or Claude, or Gemini) to “write me 25 SIE practice questions on prohibited activities” and 30 seconds later you have a fresh question set, custom to whatever topic you want, formatted exactly like the real exam. For a candidate trying to drill weak areas, that sounds like a dream tool.

The problem is the questions look real. Even an experienced candidate can read an AI-generated question and not spot the error. Worse, the model often produces explanations that confirm the wrong answer with confident reasoning. So you not only memorize a wrong fact, you build a wrong mental model around it.

What kinds of errors show up?

We had ChatGPT generate 50 SIE practice questions across all four content sections. We then audited each against the official FINRA outline, current rules on FINRA’s website, and a curated commercial question bank. The error rate by category:

The 22% includes overlap (some questions had multiple errors). Roughly 1 in 5 generated questions was unsafe to study from.

The errors broke down as:

Hallucinated rule numbers (most common). “FINRA Rule 3270” instead of FINRA Rule 3280 for selling away. “Rule 144A applies to issuers” instead of “applies to qualified institutional buyers.” Plausible but wrong.

Wrong numerical thresholds. Maintenance margin requirements, settlement timeframes, customer-complaint reporting windows. Numbers either drifted from the correct value or were just wrong.

Outdated information. T+2 settlement appearing in questions about current trade processing. Pre-Reg-BI suitability framing in customer-protection questions.

Plausible-but-wrong distractors. A multiple-choice question where the “correct” answer was actually incorrect because the model had subtly misunderstood the rule. Hardest to spot because the question reads cleanly.

Self-contradictory questions. Stem says “the customer is a non-accredited investor”; correct answer treats them as accredited. Surface-level grammar plus a logic error.

Show me a real example.

Verbatim from a ChatGPT-generated batch:

A registered representative receives a written customer complaint alleging theft of securities. Within how many business days must the firm report this to FINRA?

A. 5 days B. 10 days C. 15 days D. 30 days

Answer: B. 10 days Explanation: Under FINRA Rule 4530, firms must report customer complaints involving allegations of theft or misappropriation within 10 business days.

This question is wrong in two ways. First, FINRA Rule 4530 requires firms to report certain matters to FINRA within 30 calendar days (not 10 business days). Second, the explanation cites the right rule but the wrong timeframe. A candidate who memorizes “10 business days” from this question will get the corresponding real-exam question wrong.

If you weren’t already familiar with Rule 4530, would you have caught the error? Most candidates wouldn’t.

Show me a harder one.

Which of the following is true about Rule 144 sales?

A. Affiliates may sell freely with no volume restrictions B. Non-affiliates holding restricted securities must hold for at least 1 year before any sale C. Affiliates are subject to volume limitations even after the 1-year holding period D. Rule 144 applies only to securities of foreign issuers

Answer: C Explanation: Affiliates remain subject to volume restrictions under Rule 144 even after meeting the 1-year holding period requirement, regardless of how long they’ve held the securities.

Reads cleanly. Three of four answers feel obviously wrong. C looks correct.

But the explanation is subtly off. The holding period (six months for reporting issuers, one year for non-reporting) applies to all restricted securities, whether held by affiliates or non-affiliates. What differs is what happens after the holding period: non-affiliates can sell freely; affiliates remain subject to volume limits, manner-of-sale rules, current-info, and Form 144 filings as long as they stay affiliated. The model has implied volume restrictions are triggered by completing the holding period, when in fact they’re triggered by affiliate status itself. A candidate memorizing the explanation as written would have a wrong mental model of how Rule 144 actually works.

This is the dangerous category: the answer is technically right (C is the correct choice among the four) but the reasoning in the explanation is wrong. You’d memorize the bad reasoning along with the right answer.

Why does AI fail at this specifically?

Three architectural reasons, not just “the model needs to be smarter.”

1. LLMs predict text, not facts. An LLM generates the next likely token based on training data patterns. When the training data contains a mix of correct and incorrect statements about Rule 4530, the model picks whichever pattern is most common, not whichever is correct.

2. Multiple-choice generation requires coordinated correctness. A real practice question is correct in five places at once: stem, four answer choices, and explanation. The model has to keep all five aligned. With each generated question, the chance of all five being right compounds, and even a 95% per-element accuracy translates to ~77% per-question accuracy.

3. Distractors require deliberate wrongness. A good distractor is a mistake a real candidate would make. Generating that requires understanding both the right answer and the common misconceptions. LLMs tend to generate distractors that are randomly wrong rather than diagnostically wrong, which makes them less educational even when not factually broken.

Could prompting fix this?

A bit. Things that help:

• “Cite your source for each fact” (forces the model to anchor; doesn’t eliminate hallucinated citations).
• “Generate the question, then verify each answer choice against current FINRA rules before finalizing” (improves accuracy modestly).
• “Avoid topics changed in the last 5 years unless you’re certain about the current rule” (avoids the worst staleness traps).

Even with the best prompting, error rates drop from ~22% to maybe ~10 to 12%. A 10% error rate on a study tool is still terrible. You’d be memorizing 1 in 10 questions wrong.

What about fine-tuned or RAG-based AI tools?

Some products are now appearing that claim to generate exam questions using fine-tuned models or retrieval-augmented generation (RAG, where the model is given the source rules at inference time). These are better than vanilla ChatGPT, but with caveats:

• Quality depends entirely on the source documents the RAG system retrieves. If the source is the FINRA outline, you get accurate-but-shallow questions. If the source is third-party study guides, you inherit those guides’ errors.
• Distractor quality is still weak. Even with correct facts, AI-generated distractors tend to be obviously wrong rather than tempting.
• No editorial layer. Even RAG-based tools don’t have a human reviewing every question.

A purpose-built SIE question bank with human authors and human review will out-perform any current AI-generated set. That’s not a long-term claim, but it’s true today.

What should I use instead?

The serious options for SIE practice questions:

Curated question banks. CertFuel, Kaplan, STC, ExamFX, Achievable, Knopman. Quality varies, but these are written and reviewed by humans who teach the exam. Most have a free tier or trial. For a side-by-side comparison of price, question count, and adaptive features, see the best SIE exam prep options compared.

FINRA’s free practice questions. FINRA publishes a 75-question sample SIE exam on their website. It’s authoritative, but it’s a single static set; once you’ve seen it, the signal is burned. See free SIE practice tests compared for the full landscape of free options.

Textbook end-of-chapter questions. If you’re using a printed study guide, the end-of-chapter questions were probably written and edited. Quality varies by publisher; major ones are generally fine.

What you should not do: ask ChatGPT to write you 100 questions and study from them. The error rate will sabotage you in ways you cannot detect until test day.

Is there any legitimate use of AI for question practice?

A narrow one. After you’ve already studied a question and know the right answer, AI can be useful for:

Generating “what if” variants. “Take this question I just got right. Change the customer’s account type from cash to margin. How does the answer change?” The model can produce a thought experiment that deepens understanding. You’re not relying on AI for the correct answer (you already know the original); you’re using it to stress-test your understanding.

Explaining why a distractor is wrong. “Why is answer B wrong on this question?” If your prep tool’s explanation is thin, AI can sometimes give a clearer breakdown. Verify against the rule before trusting.

These are tutoring uses, not testing uses. The line stays the same as in Using ChatGPT to study for the SIE.

The bottom line

Don’t generate SIE practice questions with AI. The error rate is too high, the errors are too hard to spot, and the cost (memorizing wrong concepts) is too steep. Use a curated question bank for practice, use AI as a tutor for concepts you’re stuck on, and treat the dividing line between those two use cases as non-negotiable. The hour you save by generating fake questions will cost you ten hours of remediation later.