How often is the AI tutor right?

The model behind the tutor in your account: Gemini 2.5 Flash by Google, run on October 11, 2026. Temperature was set to 0 so the run can be repeated; the tutor in the app uses 0.5. If Gemini does not reply, the tutor retries with a backup model (DeepSeek V4 Flash), which this test did not cover.

Only an instruction, the question and its options. For single-answer questions the instruction was: “You are a candidate taking the nursing course exam. Choose the single best option. Reply with only the option letter.” For select-all questions it was asked for every correct letter.

One attempt per question. A select-all question counts only when the chosen letters match the key exactly, no more and no fewer. A reply that could not be read would count as wrong; there were none.

No. They are review questions from openly licensed nursing textbooks, not questions from any licensing exam. The result is about the tutor’s AI model; it does not predict your exam score.

These textbooks are public, so the model may have seen them while it was trained. The test checks the answer letter, not the explanation the tutor writes. Check important facts against your course material.

The same model as in the nursing test: Gemini 2.5 Flash by Google, the model behind the tutor in your account, run on October 11, 2026. Temperature was set to 0 so the run can be repeated; the tutor in the app uses 0.5. The tutor’s backup model was not tested.

Only an instruction, the question and its options. The instruction, quoted as sent: “You are a candidate taking the TEAS-style science course questions (anatomy, biology, chemistry), not official TEAS exam. Choose the single best option. Reply with only the option letter.”

One attempt per question. Every question has one correct option, taken from the answer key printed in its book. A reply that could not be read would count as wrong; there were none.

Areas follow the book chapters: chapter 2 of each book counts as chemistry, chapter 3 of the anatomy book (the cell) as biology, and chapter 1 of the biology book as scientific method.

No. They are end-of-chapter review questions from two openly licensed OpenStax textbooks, picked because they cover the science topics of the TEAS. They are not ATI TEAS questions. The result is about the tutor’s AI model; it does not predict your TEAS score.

It covers science only: the test had no reading, math or English questions. Chemistry has only 20 questions, and scientific reasoning (measurement, graphs, experiment design) is barely covered. Every question has a single answer, while the TEAS also uses other question types. These books have been online since 2013, so the model may have seen the questions while it was trained. The test checks the answer letter, not the explanation the tutor writes.

The same model as in the nursing test: Gemini 2.5 Flash by Google, the model behind the tutor in your account, run on October 12, 2026. Temperature was set to 0 so the run can be repeated; the tutor in the app uses 0.5. The tutor’s backup model was not tested.

Only an instruction, the question and its options. The instruction, quoted as sent: “You are a candidate taking the HESI A2-style science course questions (anatomy and physiology, biology, chemistry), not official HESI exam. Choose the single best option. Reply with only the option letter.”

One attempt per question. Every question has one correct option, taken from the answer key printed in its book. A reply that could not be read would count as wrong; there were none.

Sections follow the book chapters: chapter 2 of each book (the chemistry of life) counts as chemistry, the rest of the anatomy book as anatomy and physiology, and the rest of the biology book as biology.

No. They are end-of-chapter review questions from two openly licensed OpenStax textbooks, picked because they cover the science sections of the HESI A2. They are not Elsevier HESI questions. The result is about the tutor’s AI model; it does not predict your HESI score.

It covers biology, anatomy and physiology and some chemistry: the test had no math, reading, vocabulary, grammar or physics questions. Chemistry has only 20 questions, and solutions, stoichiometry, nuclear chemistry and redox reactions are not covered. Every question has a single answer. 443 of the 449 questions were also used in the TEAS-style test of October 11, 2026, so the two results are not independent. These books have been online since 2013, so the model may have seen the questions while it was trained. The test checks the answer letter, not the explanation the tutor writes.

The same model as in the nursing test: Gemini 2.5 Flash by Google, the model behind the tutor in your account, run on October 12, 2026. Temperature was set to 0 so the run can be repeated; the tutor in the app uses 0.5. The tutor’s backup model was not tested.

Only an instruction, the question and its options. The instruction, quoted as sent: “You are a candidate taking the pharmacy technician course questions (PTCE and ExCPT style), not official PTCE or ExCPT exam. Choose the single best option. Reply with only the option letter.”

One attempt per question. Every question has one correct option, taken from the course’s own “Solutions to Exercises” page. A reply that could not be read would count as wrong; there were none.

Most questions fall under the Medications domain of the PTCE. Dosage form and compounding questions come from MD0809; the poison prevention questions come from lesson 9 of MD0806.

No. They are end-of-lesson exercises from U.S. Army Medical Department correspondence courses in pharmacology, drug therapy and compounding, picked because they cover drug knowledge that both exams test. They are not PTCB or NHA questions. The result is about the tutor’s AI model; it does not predict your exam score.

Most questions are about drugs: classes, uses, side effects and interactions. The PTCE also covers federal requirements (DEA schedules, controlled substance rules, REMS, DSCSA), patient safety (look-alike and sound-alike drugs, error prevention) and calculations; this test has only 4 questions on poisoning and poison prevention packaging, and none on calculations. Some compounding topics are no longer in the January 2026 PTCE outline. The courses date from 2001 to 2007: questions about withdrawn drugs or changed guidelines were left out before the test, but some drug names and wording are old. Every question has a single answer. These courses are public, so the model may have seen the questions while it was trained. The test checks the answer letter, not the explanation the tutor writes.

The same model as in the nursing test: Gemini 2.5 Flash by Google, the model behind the tutor in your account, run on October 20, 2026. Temperature was set to 0 so the run can be repeated; the tutor in the app uses 0.5. The tutor’s backup model was not tested.

Only an instruction, the question and its options. The instruction, quoted as sent: “You are a candidate taking the ASVAB-style trades course questions (math, science, mechanics, shop and electronics), not official ASVAB exam. Choose the single best option. Reply with only the option letter.”

One attempt per question. Every question has one correct option, the one the course marks as correct. A reply that could not be read would count as wrong; there were none.

Areas follow the modules: D-1 is arithmetic and math, D-2 science and mechanics, C-3 shop (rigging and hoisting), E-1 to E-4 electronics.

Yes, 11. A first run used 335 questions and scored 303 (90.4%). An audit, which could see the AI’s answers, then checked every question. Only provable faults were removed: 3 need a figure that is not in the text, and 8 have a marked answer that a calculation or a basic fact shows is wrong. The AI had missed 10 of these 11, so removing them raised the score. To check that choice, a second audit was run without the AI’s answers: it flagged the same 11 questions, and on the remaining 324 the first run scored 302. 46 questions the first audit called ambiguous were kept with the course’s own key. The remaining 324 were run again.

No. They are self-test questions from openly licensed trades courses, picked because they cover math, science, mechanics, shop and electronics, which the ASVAB also tests. They are not Department of Defense questions. The result is about the tutor’s AI model; it does not predict your ASVAB score.

There are no word knowledge, paragraph comprehension, auto information or assembling objects questions. The courses are Canadian trades courses, so some questions use metric units and Canadian practice, and the shop questions are about rigging and hoisting only. Every question has a single answer. These courses are public, so the model may have seen the questions while it was trained. The test checks the answer letter, not the explanation the tutor writes.