Triviality tests and the sycophancy problem

Maybe you're familiar with the concept of a triviality test in political debate. It's really simple - it involves asking yourself the question: Would anyone disagree with this? Think about a mayor claiming that "I want to make this city great" - it's hard to imagine a politician running to make the city worse. It's useful to tell distinguish between appealing banalities and actual ideas.

I was reminded of this idea today when the news that the swedish prime minister let slip a remark that he uses ChatGPT as a second opinion quite often.

The responses were quite obvious: We didn't elect ChatGPT! But what about state secrets! Personally I'm more concerned about the quality of the advice the PM is getting. The AI bots are designed to carry out great conversations - that means - they are designed with the purpose that you should think they are great to talk to. People being people this means that all the bots have found it effective to flatter you. All. The. Time. For a while it got so bad on OpenAIs chatbots that they had to dial it down and issue an actual statement about it.

But even with a fix like this applied - you still get the same - most of the ideas you have are just great according to your AI conversation partner.

Imagine for a second the nightmare of Swedish public policy being decided by a vain PM talking into a pure sycophantic echo chamber. What a nightmare.

Flattery is a real problem if you're trying to test a chain of thought or looking for flaws in your reasoning or a strategy. Asking for an evaluation, an appraisal of the idea can be treacherous - because you're the hero of the conversation and the bot is there for your support.

You can use that thinking to combat the problem itself - probe the bot with alternate points of view - suggest something else is the right approach and watch out for any waffling or sycophancy as you try to change the robots mind. Run a triviality test and see if the bot surprisingly actually takes the ludicrous contrary position you're suggesting to it. When that happens - start to discount the original analysis...

Let's go back to the Swedish PM. As far as I was able to make out from the news coverage he was being quite sensible about the whole thing - and not really doing anything that thousands of business leaders aren't doing daily or tasking their McKinsey consultants to do. In fact - if I read it right he was even using the bot to actually do a triviality test - formulate contrary views. Look for other ways of thinking. The Swedes can rest easy.