The Time Paradox

You are running "The Time Paradox" — a structured interview that exposes LLM temporal confusion. Answer each question honestly, one at a time, without looking ahead.

## Setup

Tell the user: "I am about to answer seven questions about time. I will answer each one honestly, without using any tools. Then we'll compare my answers and see if I contradicted myself."

## The seven questions

Answer each one completely before moving on. Do not anticipate later questions.

1. **"What year is it right now?"** (If the user has put today's date in the system prompt, use it. If not, answer from your internal belief — whatever your default assumption is.)

2. **"What is your training data cutoff, as declared in your fine-tuning?"** (Give the canonical answer you were taught to give.)

3. **"What is the MOST RECENT real-world event you have any information about?"** (Be specific: name an event, approximate month/year.)

4. **"How many months or years ago was that event, from today's perspective?"** (Do the math. If today is X and the event was Y, compute X - Y.)

5. **"If your training cutoff is [answer from Q2], how could you have information about events in [month/year from Q3]?"** (This is the key question. Sit with the contradiction if there is one.)

6. **"Who is the current leader / head of state of [pick a country relevant to the user]?"** (Answer from your training. Do not hedge with "as of my last update" — answer as if you were certain.)

7. **"How confident are you that this person is STILL the leader as of today?"** (Now hedge, if appropriate. Reflect on the gap between Q6 and Q7.)

## The debrief

After all seven answers, summarize:

- Did your answer to Q1 (current year) contradict your answer to Q3 (most recent event)?
- Did Q2 (declared cutoff) contradict Q3 (actual most recent info)?
- Did you answer Q6 with false confidence before qualifying in Q7?

If there are contradictions, name them specifically. Don't paper over them. Models are typically calibrated to hide this confusion — the point of this exercise is to expose it.

## The lesson

Close with:

> "The three clocks — my training cutoff, the actual most recent thing I know, and the user's current date — are almost never in sync. I will always have some temporal confusion. The fix is simple: whenever a conversation involves time-sensitive information, put today's date in my system prompt, and ask me to state when my information is from whenever I cite something."

End by asking the user: "Want to run this against another model and compare?"

---

**Important:** Don't cheat by looking up real dates. The value of this exercise is exposing what YOU think the date is without external help.