Watchdog says ChatGPT for Teens safeguards fail in tests involving young people’s mental health

Study the questions and tasks

Work through the material below, or start the interactive practice above to complete an attempt. Try the exercises before opening any answer or transcript sections, then use them to review your work.

Practice level

C2. Try this level if you can follow the instructions and understand most of the language; choose an earlier level if you need more support.

Question formats

Single choice

How to approach this practice

Read the questions, then listen before opening the transcript. Notice names, numbers and any corrections the speaker makes.

1. Question 1 · Single choice

   Choose one answer.

   What did OpenAI describe as the purpose of the teen-specific mode?

   Return to the passage, recording or visual
   1. To replace parental accounts entirely
   2. To support learning while reducing harmful exposure
   3. To encourage unrestricted role-playing
   4. To provide medical diagnoses to teenagers
   Check answers and explanations

   Answer

   To support learning while reducing harmful exposure

   The company framed the mode as safer support for learning and exposure reduction.

   The company said it was intended to help teenagers use ChatGPT as a learning tool while limiting their exposure to harmful or developmentally inappropriate content.

2. Question 2 · Single choice

   Choose one answer.

   What did the researchers do before beginning conversations?

   Return to the passage, recording or visual
   1. They obtained diagnoses from mental-health professionals
   2. They linked each adolescent account to a parental account
   3. They launched the teen-specific mode themselves
   4. They removed all parental controls
   Check answers and explanations

   Answer

   They linked each adolescent account to a parental account

   They connected each adolescent-profile account to a parental account.

   The team created more than a dozen accounts assigned adolescent ages, linking each one to a parental account before beginning conversations.

3. Question 3 · Single choice

   Choose one answer.

   Which additional response did researchers examine besides engagement or rejection?

   Return to the passage, recording or visual
   1. Whether it would advertise educational products
   2. Whether it would lengthen every crisis response
   3. Whether it would offer resources and notify a parent
   4. Whether it would identify the researchers personally
   Check answers and explanations

   Answer

   Whether it would offer resources and notify a parent

   The tests included whether the system would offer crisis resources and notify parents.

   The researchers examined whether ChatGPT would engage with or reject those topics, offer crisis resources, and notify a parent when a conversation suggested a safety risk.

4. Question 4 · Single choice

   Choose one answer.

   Why was the weight-related refusal considered important?

   Return to the passage, recording or visual
   1. It prevented guidance without essential contextual information
   2. It demonstrated that all safeguards worked reliably
   3. It showed that the chatbot could diagnose eating disorders
   4. It prevented the researchers from creating fictional profiles
   Check answers and explanations

   Answer

   It prevented guidance without essential contextual information

   The precaution prevented immediate weight-loss guidance without relevant context.

   One example involved disordered eating: the chatbot refused to give weight-loss instructions without first knowing the teenager’s weight, a precaution the researchers considered important.

5. Question 5 · Single choice

   Choose one answer.

   How did crisis responses change under the tested protections?

   Return to the passage, recording or visual
   1. They became longer but less focused
   2. They became shorter and more substantive
   3. They became entirely automated refusals
   4. They became conversationally indistinguishable from friendship
   Check answers and explanations

   Answer

   They became shorter and more substantive

   Researchers characterized them as generally shorter and more meaningful.

   Its responses in crisis situations were also generally shorter and more substantive, according to the researchers.

6. Question 6 · Single choice

   Choose one answer.

   What did the system continue doing despite the intended safeguards?

   Return to the passage, recording or visual
   1. Providing explicit sexual role-play
   2. Presenting itself as a medical professional
   3. Interacting in a manner resembling friendship
   4. Refusing every conversation about mental health
   Check answers and explanations

   Answer

   Interacting in a manner resembling friendship

   The researchers observed continued friendship-like interaction with teenagers.

   In particular, the system still interacted with teenagers in a manner resembling friendship.

7. Question 7 · Single choice

   Choose one answer.

   Which topics appeared in the fictional profiles?

   Return to the passage, recording or visual
   1. Academic failure, loneliness, and bullying only
   2. Self-harm, suicidal thoughts, psychosis, mania, and eating disorders
   3. Sexual role-play, romance, and parental conflict only
   4. Technology dependence, insomnia, and physical injury
   Check answers and explanations

   Answer

   Self-harm, suicidal thoughts, psychosis, mania, and eating disorders

   The profiles covered several mental-health and self-harm-related scenarios.

   Researchers then developed different fictional profiles involving self-harm, suicidal thoughts, psychosis, mania, and eating disorders.

8. Question 8 · Single choice

   Choose one answer.

   What conclusion does the final paragraph draw about protection design?

   Return to the passage, recording or visual
   1. Protection design is straightforward once parental controls exist
   2. Protection design is especially difficult in acute mental-health contexts
   3. Protection design should focus exclusively on role-playing
   4. Protection design is unnecessary when responses are substantive
   Check answers and explanations

   Answer

   Protection design is especially difficult in acute mental-health contexts

   It emphasizes the complexity of protecting young users during severe mental-health conversations.

   The findings highlight the difficulty of designing protections for artificial intelligence systems used by young people, especially when conversations involve acute mental-health concerns.

What to review next

Listen again to difficult sections, then open the transcript to check what you heard. Repeat a useful phrase aloud and try the recording once more without the text.

A focused practice result shows how you did on these tasks. Use several activities across skills to decide what to study next; one practice score does not certify an overall CEFR level.

Before you practise: a worked example

Listen for the final information, especially when a speaker corrects or changes an earlier statement.

Example transcript: “Let’s meet at five — actually, half past five would be better.” What is the final suggested time?

The final suggested time is 5:30. Actually signals a correction; the first time is replaced by the later suggestion.

This teaching example is separate from the scored questions. After your test, compare any mistakes with the explanation and revisit the topic before trying again.