Q1. What is the passage mainly about?

1. AHow to write a good book
2. BWhy AI sometimes gives confident but wrong answers
3. CHow to get high scores on tests
4. DWhy AI is better than people

AIが自信満々に間違える「ハルシネーション」とその理由が主題なので B。

Q2. What does AI do instead of checking facts?

1. AIt asks a teacher.
2. BIt reads a dictionary.
3. CIt searches the library.
4. DIt guesses the next word that sounds right.

AI does not check facts. It guesses the next word that sounds right. とあるので D。

Q3. In this passage, "confident" means ...

1. Asure about yourself
2. Bvery angry
3. Cquiet and shy
4. Dtired and slow

confident は「自信がある」。in a very confident way で「とても自信満々に」なので A。

Q1. How do large language models create answers?

1. ABy checking facts in an encyclopedia
2. BBy asking human experts
3. CBy predicting the word most likely to come next
4. DBy copying answers from students

work by predicting the word that is most likely to come next とあるので C。

Q2. Why is guessing "always the better strategy" for AI in many tests?

1. ABecause wrong answers lose many points.
2. BBecause "I don't know" and wrong answers both get zero points.
3. CBecause the tests are very easy.
4. DBecause humans check every answer.

zero for both a wrong answer and "I don't know" とあり、白紙でも間違いでも0点なら推測した方が得なので B。

Q3. What would the researchers most likely support?

1. AA test that gives some credit for honestly saying "I don't know"
2. BA test that gives points only for fast answers
3. CRemoving all tests for AI
4. DGiving extra points for long answers

suggested rewarding AI when it honestly admits uncertainty とあるので A。

Q1. According to the passage, why do models hallucinate more about obscure people or papers?

1. ABecause the models refuse to answer.
2. BBecause users ask about them too often.
3. CBecause benchmarks never include such questions.
4. DBecause such information rarely appears in their training data.

especially about obscure people or papers that rarely appear in their training data とあるので D。

Q2. What is a weakness of the method proposed in Nature?

1. AIt requires a lawyer to check every answer.
2. BIt works only in English.
3. CIt cannot detect errors that the model repeats consistently.
4. DIt makes the model answer more slowly.

The method misses errors that a model makes consistently, though. とあるので C。

Q3. What did the 2025 preprint by Kalai and colleagues argue?

1. AModels should never answer legal questions.
2. BScoring a refusal to answer like an error pushes models to guess.
3. CHallucinations happen only with obscure topics.
4. DAsking the same question twice removes all errors.

treat declining to answer exactly like an error, so models are in effect trained to guess とあるので B。