人工智能神话测验

当您参加测验时，就会出现答案和解释。

1. A model answers fluently and names a source. What should a reader still check?
   • Only whether the answer is polite
   • Whether the response is long enough
   • Whether the model used a large context window
   • Whether the source exists and supports the claim
2. Does more training data guarantee a better system?
   • No; data can be irrelevant or systematically flawed
   • Yes, if it comes from the internet
   • Yes, regardless of the task
   • No model can learn from more data
3. Why compare a larger model with smaller or simpler alternatives?
   • Larger models always have worse accuracy
   • It may cost more without meeting the task’s needs
   • A benchmark score cannot be measured
   • Model size determines deployment law
4. A demo solves one impressive case. What conclusion is justified?
   • That the system can handle every difficult case
   • That it will be reliable in production
   • Only that the demonstrated setup handled that case; more tests are needed
   • That no tool or context contributed
5. Does automation remove human responsibility for the system?
   • Yes, when outputs are generated automatically
   • Yes, if a vendor hosts the model
   • Only if a model has a high score
   • No; people choose goals, data, interfaces, permissions, and deployment conditions
6. What does one failure tell you about a system?
   • It reveals a failure worth testing, but not that every use is ineffective
   • That the system is useless for all tasks
   • That the model is conscious
   • That all future versions will fail identically
7. What is a benchmark result distinct from?
   • A measured score
   • A future prediction or a statement about consciousness
   • An evaluation setup
   • A task-specific test
8. How should a broad capability claim be phrased?
   • As strongly as possible to attract attention
   • As a guarantee based on a single score
   • At the scope the supporting evidence actually covers
   • Without naming the test conditions