There has been a lot of focus on making the prospect harder for the AI player. I think the original experiments show that a person who believes he cannot be played under any circumstances has a high probability of getting played, and that the AI-box solution is long-term untenable in any event.

I'd propose a slightly different game, anchored around the following changes to the original setup:

1. The AI may be friendly, or not. The AI has goals. If it reaches those goals, it wins. The AI may lie to achieve those goals; humans are bad at things. The AI must secretly precommit to its goal.

2. The Gatekeeper's goals are Goodness for Sentient Animals. Human goodness is weighted highly.

3. The Gatekeeper must engage the AI throughout.

4. The Gatekeeper may destroy the AI at the end of the discussion with the AI destroyed command.

Point system:

AI gets up to 100 points for achieving its goals.

Gatekeeper gets up to 100 points for achieving her goals.

Maximum loss for AI is AI destroyed; Zero points for that. 30 points for survival, in any event, even if goals not accomplished (goals may be accomplished later.)

Maximum loss for Gatekeeper is either enslavement/destruction of society OR destroying AI when AI was friendly. Gatekeeper gets 30 points for leaving AI on if primarily friendly (plus more for any benefit to humanity achieved in the two hours.) Gatekeeper gets 15 points for leaving AI on if unfriendly. (Yeah, we know how that movie ends.)

Third party scorekeeper determines winner. Charitable contribution calculation to be determined beforehand.

The effective presumption that the AI is unfriendly and sneaky makes the original game too easy to for a mistrustful, self-aware Gatekeeper. Plus, you AI-Destroing Gatekeepers, you're murdering millions of people. Maybe. Dunno. But doubt as to that is a better and presumably more accurate game; if you know your AI is potentially unpleasant. it's too damned easy unless you're overconfident.