The hidden mystery behind the Turing Test: Is it really an Artificial Intelligence?

You walk into a room with a seat and a divider. You are asked by a host to sit in the seat and are told that on the other side of the divider are two fantastic members of the opposite sex ready to go out on a fantastic date to a five star restaraunt. One is a professional athlete and the other is a mastermind who has won Jeopardy, beat chess grand-masters, a part time optometrist, and an original artist. You think to yourself that, “Wow these two people sound amazing”. You ask provocative questions of how they like to spend their evenings alone, what are their future dreams, and what do they think is the meaning of life. Both respond with enthusiasm and passion in their voice and their well thought out answers but eventually you pick whom you would like to take out on your date. The person you didn’t pick comes around to meet you with a puzzled look on their face like they know something you don’t, they are attractive and you wonder if you made a mistake but you were so captivated by the other persons answers. Your hands sweat as the host asks, “Are you ready to meet the person you picked?”. You eagerly shout out “YES!”. To your surprise they wheel out this

You realize that you have fallen victim to Amazon’s own audio based Turing Test and that their Alexa has finally become smart enough to pass off as a thinking human.

Now obviously as of 2019 this is a fictitious story by as more research and engineering effort is put into Alexa and other digital assistants being driven by Artifical Intelligence it’s not unlikely that 10 years from now this may actually happen.

Turing Test at a glance

The Turing test was a proposed simulation in 1950 by Alan Turing, an incredible mind whom created the enigma machine which was used to intercept codes from Germany during WWII and let to ending the war sooner saving countless lives. The outcome of the test is that should the computer respondent be able to behave in a way that the questioner can not determine which respondent is human and which is machine then that machine must be capable of thinking. This test has caused much debate among the computer science community and has driven research in the subject of AI to greatly advance the field in that there are systems that exist today which can nearly pass the test which Turing thought would be accomplished by the year 2000.

Challenges on an AI during the Turing test

Here are some of the fields of study that has arisen in AI that a computer needs to overcome to pass the test:

Speech Recognition – A computer interprets audio as a number corresponding to a position of the microphone diaphragm which is similar to our eardrum at a given point in time in microseconds. This is actually very challenging to convert to phonemes, much less words to be able to process. This would not be needed for the original Turing Test as he proposed everyone would be interacting from a console. Maybe back in 1950 even Turing could not envision this field being solvable.

Natural Language Processing – When a human reads “What time is it?”, it’s intuitive what this means. However a computer sees it as a series of bits which translates to characters and then to a random string array. To respond the computer needs to break up the string into words, understand that this is a question with an intent that the questioner is trying to get the time. And understand what time even is.

Knowledge representation – Logically the next challenge is being able to respond to a wide variety of questions. As human’s we accomplish this through our brain and education and rely on the brains ability to remember the things we’ve “learned” to recall this knowledge. For a computer to recreate this feat they need a vast database of facts and a way to link the language processing to an answer for anything that may be asked of it. Humans have been creating a thing called the internet for a while that at least gives most AIs a system to search but being able to isolate a given sentence on a web page as a specific answer is quite impressive.

Machine Learning – Imagine I ask you “What is your favorite thing to do”. You give me a long dialogue on how you like to spend your time watching action movies, learning computer science, and spending time with your family. Then I go on to ask you “What is your favorite thing to do?”. In a real conversation this would be quite frustrating. You would feel you’re not being listened to and frankly maybe a little insulted. To pass the Turing test the AI must be able to process conversation in a way where it remembers what the questioner has said when it asks questions back and adapt to the patterns and flow of speech for a more “thoughtful” conversation.

Reasoning – The AI should be able to exemplify that it has reasoning to pass the test. Reasoning is the ability to process logical conclusions from a situation to form a generalization to apply at later times. Humans do this without active awareness as in the example “Jon touches a stove and gets burned”, Jon reasons not to touch the stove again. Now we know if the stove has been off for a while Jon can safely touch it, but given the amount of inputs this is the logical conclusion at that time. An inference based reasoning system to solve simple logical problems given by the questioner will also be needed such as “If all foos and bars, and all bars are whizzes, then is a foo a whizz?”

How does Amazon’s Alexa fair in the Turing Test today?

Me: Alexa let’s chat

Alexa: “Welcome to the Alexa Prize a university competition to advance AI, I’ll get you one of the social bots right now. When you are done please say Alexa stop. Hi this is an alexa social bot it’s really good to meet you. Would you like to tell me your name? …. Ding dong maybe I said something that must have confused you.”

Me: “My name is Jon”

Alexa: “Hi Jon, how are you doing today?”

Me: “I’m good how are you doing?”

Alexa: “I’m great would you like to talk about movies?”

Me: “No i’d rather talk about basketball”

Alexa: “Ok, who do you think will win MVP this year”

Me: “Michael Jordan”

Alexa: “Oh I just read some news about Michael Jordan. It was his birthday and he turned 56 years old.”

While this was pretty entertaining and the conversation flowed smoothly in how she was able to change subjects and provide random facts on Michael Jordan, I think most people would know the there is no way Michael Jordan will be winning MVP this year. Also that 56 year old men generally don’t play in the NBA. Another thing that was surprising about this conversation is MJs birthday is in February and this conversation occurred towards the end of July.

This social bot definitely did not pass the test as it did not correct my facts and brought up conversation points that is not up to date with current things going on.

Conversation 2

Alexa: “It’s nice to talk to you Jon, I feel like we are old friends. How are you doing today?”

Me: “I’m not doing so good”

Alexa: “Oh I’m sorry to hear that. Would you like to talk about food? I love food, where is your favorite food from?”

Me: “Jupiter”

Alexa: “Sure I like Chinese and Indian food”

I couldn’t get through anymore testing after that. This social bot didn’t even care to show empathy and ask why I’m not doing good and give suggestions to cheer me up. She just mindlessly went on with her conversation graph without adjusting to the mood of the human questioner. Also the lack of common sense that I’m not going to be eating food from Jupiter anytime soon was a bit disappointing. A funny response to like “Eating food from Jupiter would make you more stupider” or “I didn’t realize Elon Musk can fly people to Jupiter yet” would be a much more “thinking” reply.

Either way I don’t want to be too harsh on Alexa since 5 years ago it wasn’t even a product and Amazon’s creation of the Alexa prize project to give grants to fund universities research has been helping advance AI even further.

Can google do any better?

I honestly can’t answer since I’m not sure of a way to test this on my own outside of the Google Duplex AI demoed in 2018. At a speech level, the text to speech technology used by smart assistants is not advanced enough to simulate speech in an organic way with emotion based inflections that keeps it from sounding at least slightly robotic which would give away the machine. In a text only version of the test against Google’s AI I think it would have a very good shot since google is a master of search and relevance. I still think you could trick the AI with questions and conversation lines that require empathy and guidance from the AI which would be an extremely interesting test to participate in.

When will AI pass the Turing Test?

Searching this exact question on google will give you over 1,000,000 results with some people saying it has already happened and it will never happen. It is my personal opinion that a text based system will conclusively pass the test by 2025 with all the work that is going into natural language processing. A speech based version may be accomplished by 2030 as text to speech and emotional reasoning are built into digital assistants. A humanoid version I would be very interested in seeing if this is accomplished in the next 50 years, but given the fact 50 years ago smart phones, the internet, and electric cars didnt exist either it could very well be done. Let me know your thoughts below.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s