The search for an artificial intelligence capable of simply maintaining a conversation in a friendly way turned out to be seriously long. When Alan Turing, the father of modern computing, presented his famous test to show the true intelligence of a computer program, he set a very ambitious goal for the hardware. If a computer could convince a group of human judges that they are talking to a person - if it could hold a conversation - that would be evidence of an artificial intelligence that has evolved to the point of being indistinguishable from a human.
This glove was dropped in 1950. So far, no computer program has succeeded in passing the Turing test. Individual pseudo-feats do not count.
Why false feats? Because they are considered failures rather than achievements. For example, Joseph Weisenbaum back in 1966, when computers were programmed with large punched cards, developed software for natural speech processing - ELIZA. “Eliza” was a machine that was supposed to carry on a conversation, pretending to be a psychotherapist; you can talk to her today.
Talking to "Eliza" is a bit weird. She often paraphrases what you said, so for example, if you say "I feel depressed," she might respond, "Did you come to me because you feel depressed?" When she doesn’t understand what you’re saying, she replies “yeah” or “tell me more.”
During the opening lines of the dialogue, especially if you treat her like your doctor, “Eliza” can be quite convincing. When Weisenbaum noticed this, he was slightly alarmed: people were ready to see more human in the algorithm than there was human in it. Soon, although some of the subjects nevertheless realized that they were dealing with a machine, they revealed their deep experiences and secrets to it. They poured their souls out to the machine. When Weisenbaum's secretary spoke to “Eliza,” although she knew she was speaking to the program, she still insisted that Weisenbaum leave the room.
Some of the unexpected reaction that ELIZA has elicited may be that people are more inclined to open up to the machine, realizing that no one will judge them, even if the machine cannot say anything related or help at all. The Eliza Effect was named after this computer program: people tend to endow machines with human traits or think of them as human.
Weisenbaum himself, who later became deeply suspicious of the impact of computers and artificial intelligence on human life, was amazed that people were willing to believe that his script was human. "I never thought that a very short acquaintance with a simple computer program would lead to such delusional thinking in quite normal people."
The Eliza effect may have alarmed Weisenbaum, but it has intrigued and fascinated others for decades. You may have noticed it on yourself, talking to AI like Siri, Alexa or Google Assistant - these short answers seem too real. In your sane mind, you know that you are talking to a large chunk of code stored somewhere on the air. But subconsciously, it seems to you that you are talking to a person.
Promotional video:
Further more. Facebook opened its Messenger program and made software available to people and companies to create their own chatbots. Natural speech processing has progressed by leaps and bounds since the 1960s. Today you can find welcoming chatbots like Mitsuku, who recently won the Loebner Prize, which is awarded to machines about to solve the Turing test. She will answer whatever you write to her. The idea is simple: if there is already an app for ordering pizza, why not order a pizza by asking, for example, a chatbot like an old friend?
Startups like Semantic Machines hope their virtual assistants can interact with you like secretaries, while still being able to extract information from the Internet. Soon they will be everywhere.
But people who make chatbots - both social and commercial - face a common problem: People, perhaps subconsciously, accept chatbots for people and get frustrated when they can't keep up a normal conversation. Frustration with misunderstandings is often linked to high initial expectations.
Until now, no machine has been able to learn to understand the context - take into account what was said earlier, refer to it and respond based on the current position of the dialogue. Even Mitsuku will often try to recall the topic of conversation after a couple of lines of dialogue.
This is clear. The conversation can be multifaceted and complex. There may be hundreds of answers to whatever you say that make sense. When you create additional layers of conversation, these factors multiply, multiplying all kinds of dialogue variations. It's harder than playing chess or go.
But that doesn't stop people from trying to create new chatbots. Amazon recently launched the Alexa Prize, in which the AI winner will be awarded a $ 500,000 prize plus an additional $ 1 million on top if the development team can create a "social bot" that can talk to human users for 20 minutes on a variety of topics. …
Topics identified include science and technology, politics, sports and celebrity gossip. The finalists were recently announced: chatbots from universities in Prague, Edinburgh and Seattle. The finalists were selected based on Alexa user ratings.
Having narrowed the area of conversation to a specific range of topics, the chatbot begins to skillfully bypass the problem of context. It is much easier to simulate a conversation that deals with domain-limited topics.
Developing a machine that can support almost any human conversation can be challenging. Perhaps it will require general artificial intelligence for a complete solution, rather than previously used approaches with recorded responses or neural networks that associate input data with responses.
But there will certainly be a machine that will conduct meaningful dialogue and which people can enjoy. The Alexa Prize winner will be announced in November. The Eliza effect means we will trust machines sooner than we thought.
Ilya Khel