How to Administer an Effective Turing Test

Today is the first time in a few years that I failed to administer a Turing test when I should have. It didn't cost me anything, I just felt silly for talking to a robot for longer than I should have. Essentially it cost me time and pride. I'm really good at that stuff. The fact that I can make that mistake makes me concerned for the rest of you who aren't as good at it as I am. It's a very important skill. As AI becomes more pervasive in our lives the ability to administer a good Turing Test is very important. It's also a term that I use a lot, so knowing what it means will help you read my posts. So, here is your guide!

Administering Effective Turing Tests

First, let's define what a Turing Test is. Alan Turing was the individual who helped utilize the first computers to unravel the Enigma Machine, which led the Allies to saving many casualties during WWII. He theorized that computers would eventually become intelligent enough to outwit humans. So he came up with a test. The test is quite simple:

Can a human, in a text based chat, identify that the entity they are talking to is a robot.

If a human fails to identify a robot, said robot is said to have passed a Turing Test. In Turings time he envisioned limiting these chats to certain topics. So people would have competitions that were specific to a topic, like history or something else.

In modern days we don't have this luxury. It's not history professor vs 5 history bots and 5 other history professors. It's you vs something that sounds like your best friend. This is a much more challenging Teste to implement. Becoming good at identifying malicious robot actors is, or will become, an essential skill in our world today. I will outline two tests, 1 for text based chats and 1 for voice (a problem even Turing did not anticipate).

For Text Chats

1. Entertain initial conversation even if preliminarily suspicious. This is actually very important as you can fairly easily catch robots in lies just by engaging in conversation.
2. If you become suspicious push the conversation farther than the AI gets comfortable. This will unveil poorly trained AI that, like a shitty salesman, are going for quick wins against the naive.
3. Throw in s0me numbers. A human w1ll immediately rec0gnize this sentence, a robot will trip and share complete nonsense unless trained exceptionally well. (cough quantum paradoxes cough)
4. Finally, if you aren't sure, a human should graciously accept an ask for a phone call or, if "family/friends" would be happy to exchange a tidbit of how you met or a fond memory... an AI will get defensive.

For Voice Chats

The process for VoiceChats is a little different, mostly because we don't have the handy placing numbers in text trick. Though it is a very similar process.

1. Again, entertain initial conversation and see if anything seems obviously off. This also lays the ground work for simple queries that can trip AI up. "You mentioned ... blah ... why did you mention that?" is a very powerful question that will catch a LOT of AI that do not have memory of your conversation, which would be very common for AI texting you or reaching out over chat apps.
2. Attempt to move the conversation to a different mechanism. "Hey, can you text me later and I'll give you a call when I'm free?" or "Will you reach out to me from Telegram or WhatsApp" (both of these are excellent encrypted forms of communication that are free to use)
3. Utilize a tone that is not yourself. A robot that is trained to impersonate a friend or a relative will trip up if you are not behaving like yourself! Whereas a friend will just laugh and call you an idiot.
4. Same step from text chat. Don't overruse this tactic or you're providing training material and inviting future attacks. Save step 4 for times when you're really concerned and the above steps fail to convince you one way or another.

That's it. It's that simple! Remembering all of the steps and using the appropriate tactic in the right situation is that part that is difficult. It's a skill we all should practice... even when being reached out by a profile you recognize. If you're not sure... do the test!