Why not just ask ChatGPT for questions from my lecture?

• ·You want a concept explained another way. Ask for a simpler version, an analogy, or a follow-up. Then confirm anything you will be examined on in your lecture.
• ·You want a quick warm-up on a topic you already know. If you can spot a wrong answer key on sight, a few chatbot questions make a cheap warm-up.
• ·Speed matters. Two of the studies credit chatbots with speed and then list the limits. One found that "ChatGPT-4o demonstrates the potential for efficiently generating MCQs but lacks the depth needed for complex assessments." Another wrote: "While ChatGPT can efficiently generate MCQs for ophthalmology residents, it has notable limitations".
• ·Your budget is zero today. Claude's plans page (checked on 28 Oct 2026) lists a Free plan at $0 and says "Free covers everyday questions." Gemini's plans page (checked on 28 Oct 2026) lists a no-cost tier with a Google account.
• ·You want to attach a file. Claude's help page (checked on 28 Oct 2026) lists PDF among the documents it accepts. It allows "Up to 20 files per chat" and "500MB per file". It says Claude "analyzes both text and visual elements" in PDFs of 100 pages or fewer. Gemini's help page (checked on 28 Oct 2026) says "Gemini Apps support most file types" with a 100 MB limit for files other than video.

A chatbot hands you a question and an answer key. Nobody has checked either one.

Five studies in medical journals had experts or faculty review MCQs written by ChatGPT, Claude, Gemini, and Bing. Here is what each one reported in the authors' own figures.

ChatGPT 3.5, immunology (Ngo and colleagues, Academic Pathology, 2023)

The authors pasted the text of their own immunology study guides into ChatGPT and asked for questions: 60 in total. "ChatGPT only generated correct questions and answers with explanations in 32 % of the questions (19 out of 60)." They rated 25% (15 of 60) as completely wrong. The lesson: giving a chatbot your source text did not stop it from writing wrong answers.

Claude Sonnet 3.7, thoracic surgery (Dal, Journal of Health Sciences and Medicine, 2025)

Claude was given the core topics of a fifth-year thoracic surgery internship and asked for a 20-question exam with an answer key. "Among AI-generated questions, 8 (40%) contained major errors, while 1 (5%) had a minor error." Three questions had two correct answers. Three had the wrong answer marked in the key. Two contained medically inaccurate information and left the correct answer out of the options. The specialist-written exam had no errors.

ChatGPT-4o, emergency medicine (Law and colleagues, BMC Medical Education, 2025)

The study compared 100 AI-written questions with 100 human-written questions from a high-stakes emergency medicine exam. Expert reviewers found "more factual inaccuracies (6% vs. 4%), irrelevance (6% vs. 0%), and inappropriate difficulty levels (14% vs. 1%) in AI MCQs." Eleven AI questions had multiple correct answers or no correct answer against one in the human set. To be fair to the AI: the human-written set had 24 grammatical or spelling errors and the AI set had none. The reviewers also knew which questions the AI wrote. The authors concluded: "Human review remains essential to ensure quality."

ChatGPT 4, ophthalmology (Feizi and colleagues, British Journal of Ophthalmology, 2026)

The study compared 100 ChatGPT questions with 100 written by faculty. "Item-writing flaws were more frequent in ChatGPT-generated MCQs (56%) compared with human-authored MCQs (27%, p<0.001)." The ChatGPT questions were also weaker at separating strong candidates from weak ones.

ChatGPT-3.5, ChatGPT-4, Gemini, and Bing (Abouzeid and colleagues, JMIR Medical Education, 2025)

Researchers gave four chatbots the same prompt to write SBA questions for undergraduate medical students. One finding is worth knowing before you drill a chatbot's questions: "Gemini consistently demonstrated a preference for the correct answer to be listed as option B." The authors concluded that all four chatbots "require expert evaluation to ensure their suitability for use."

Read these fairly. Each study tested specific model versions at one point in time on a small sample. Newer versions may do better. The studies used different subjects and methods. They cannot rank one chatbot above another. What they share is the conclusion: chatbot questions need expert review before anyone learns from them.

Recall Engine has two parts. Each one handles the review problem differently.

The prepared question bank. It held 31,840 questions on 28 Oct 2026. Library content follows its source lectures and is reviewed under clinician oversight. AI models help draft library questions. 100% of bank questions have five options. You can report a problem from any question in the app. A correction is published as a new version. A withdrawn item leaves the bank. Every figure is defined and dated on our facts page.

The generator for your own lecture. Upload a lecture PDF or PowerPoint up to 100 MB and pick the pages. SBAs have five options from A to E. Every option is explained. Every answer points to the lecture pages it came from. AI models write these questions on demand from your pages. A second AI model then checks them. That check can fail or miss the errors the studies describe. No clinician sees them before you do. The page reference is there to make your check fast.

What we do not claim. Recall Engine can still get a question wrong. The difference is where your checking starts. Library content follows its source lectures and is reviewed under clinician oversight. A generated question points you to the pages to check.

Privacy. To make a set, the lecture's pages are sent to our AI providers for processing. Only you can see your lectures and study sets.

These checks come from the errors the studies found. They apply to every tool. That includes Recall Engine. For a worked example, read how to check AI-generated medical questions.

1. 1.Find the answer in your source. A wrong answer key was one of the major errors in the Claude study. Do not trust the marked answer until your lecture supports it.
2. 2.Make sure exactly one option is defensible. Two correct answers appeared in 3 of 20 questions in the Claude study. In the ChatGPT-4o study 11 of 100 questions had more than one correct answer or none.
3. 3.Make sure the correct answer is actually there. Some questions in the Claude study left it out of the options.
4. 4.Cover the options and answer from the stem. The JMIR study calls this the "cover test". A good question can be answered from the vignette and lead-in alone.
5. 5.Watch the answer letters. If the key keeps landing on the same letter, you may be learning a pattern instead of the content.
6. 6.Read the explanation for every wrong option. It should say what rules that option out. A vague line teaches nothing.
7. 7.Look for invented detail. An age, a lab value, or a qualifier your lecture never gave can change the answer.
8. 8.Ask when you are stuck. If the source does not settle it, ask your educator. Never use a practice question for a patient-care decision.