Design ‘Seemingly’ Intelligent Machine

It’s not hard to hear the news that AIs defeated grand chess masters. Could these AI machines actually think?*

Definition of Intelligence

We often confuse the meaning of ‘smart’ and ‘intelligent’. Let’s say, if someone solves a difficult math problem without being stuck, we say to him/her as “Wow, you are really smart!” or “You are very intelligent.” Probably the definition of being intelligent in our mind is ‘the ability to understand and solve difficult problems.’

…but what exactly is intelligence? I searched the word on The Oxford dictionary website. It defines intelligence as ‘the ability to acquire and apply knowledge and skills’. Aha, there is a big difference between my understanding and the actual definition. The intelligent one should be able to think on its own and solve a problem with it, no matter how simple the problem is.

Can machine think without human interventions? How do we evaluate that a machine has the thinking ability? This is quite tricky to answer because the definition of thinking might be very different depending on the criteria.

How to Evaluate Machine’s Intelligence

Explaining Turing Test

The term machine thinking reminded me the most well known artificial intelligence(AI) evaluation method, Turing test. Introduced by the British computer scientist Alan Turing in 1950, the Turing test is designed to evaluate natural language conversations between a human and a machine that is programmed to mimic human-like responses. If human evaluators can’t tell the machine from the human during the chatting, the machine is said to have confirmed that it has intelligence.** In other words, machine thinking means the ability to imitate human languages.

Despite of its limitation that it can only test simple text conversations, Turing test has become the most successful AI evaluation tool. Through his paper Computing Machinery and Intelligence, Turing expected that artificial intelligences will be able to generate a part of the words and educated opinion that one will be able to prove the existence of artificial intelligence without expecting to be contradicted.***

Here is an interesting question arises; Why Turing used natural conversations to evaluate artificial intelligence? I couldn’t find a specific reason on his paper but it is quite easy to guess.

Can Machine Fool Human?

What is conversation? My definition of conversation is an elaborated interaction happened between more than two people. To continue a conversation, both people should be able to think and respond proper answers of counterpart’s questions. That is, if machine don’t have enough ability to think, it would be impossible to have a long conversation with human.

Supposed that machine is intelligent enough to analyze conversation and provide freaking proper answers to human. Can machine perfectly fool human? Well, I would say no. In order to have a human-like conversation, machine should be able to mimic and respond non-verbal communications such as gestures and onomatopoeic words. Also, real humans make mistakes or sometimes lie to others, which is hard for machine to find triggers of errors. The Turing test has a huge limit that it can only be conducted to simple, text-based conversations. When it comes to spoken conversation, it is so easy to catch the difference between human intelligence and AI.

Chatbots: Designed to Be Imperfect

Duolingo recently started Chatbot service for language learners

Although it is impossible to fully surrogate human intelligence, that doesn’t mean that imperfect AIs can’t improve our lives. In fact, I think AIs should be designed more like machine so all people can recognize them as machine. It’s my belief that we would experience chaos and depression if we can’t reliably tell the difference between human and computer. So far, some companies have introduced several interesting AI-based services and stimulated our curiosity. However, none of them have perfectly replaced human intelligence.

One good example is Duolingo Chatbot that I’m using now. It is designed almost like an ordinary text messenger, helping me to learn new languages though chatting with quasi-humans. It’s fun and very effective. The best part is, I haven’t experienced the uncanny valley because I know that it is the machine (It only responds limited answers regardless of what I typed). If it was so natural that I confused my artificial language teacher as a human, I would be freaked out!

Conclusion

In summary, a machine’s intelligence is determined by the ability to use natural language, but so far we haven’t experienced impeccable AIs that can replace human intelligence because of the limitation of machine languages. Nevertheless, I believe there are a lot of possibilities for designers that create many interesting, useful AI-based services.

Reference

*https://fossbytes.com/this-ai-based-program-teaches-itself-chess-in-72-hours-becomes-chess-grandmaster/

**https://en.wikipedia.org/wiki/Turing_test

***Turing, A.m. “Computing Machinery And Intelligence.” Readings in Cognitive Science, 1988, 433–460. doi:10.1016/b978–1–4832–1446–7.50006–6.