Did an Artificial Intelligence system just become human?

The first book of the Bible relates how God blew into the nostrils of the first man giving him life. Ever since that time, humans have dreamt about the possibility of bringing an inanimate object to life. So much so that we see numerous fictional stories centred around this theme. From the well-known children’s fairytale Pinocchio, whereby a wooden doll received the gift of life through magic, to the famous Frankenstein by Mary Shelley, where a scientist managed to bring a corpse to life using electricity.

The prospect of becoming the creator has always fascinated humans. In the field of Artificial Intelligence (AI), this idea surfaces time and time again simply because AI scientists try to mimic human functions with the use of machines. If one looks at recent science-fiction movies like “2001: A Space Odyssey”, “ex Machina”, “Ghost in the Shell”, “her”, and “the Matrix”, to name a few, it is very evident that this theme lies at the heart of most storylines.

The notion of reaching human potential and maybe even surpassing it has been a significant target since the inception of AI. Right after the 2nd World War, Alan Turing, the father of AI, devised the Turing test, which states that if we cannot distinguish whether a machine is a human or a bot, then that machine would have reached human intelligence. He designed such a test because he couldn’t find a coherent definition of intelligence. Humans can quickly call an animal intelligent if it can understand commands and perform simple actions. We can marvel at the massive structures built by thermite colonies and label them as an engineering feat. But if someone asks us to define intelligence, we are lost for words. Because of this, Alan Turing chose a definition by association whereby he associated machine accomplishments with human achievements. However, the general belief today is that the test is too simple, and we should be looking at something more complex. Even if we use it as a baseline, till today, no AI system has ever managed to pass the Turing Test!

But things might be changing! In 2021, Google announced that it had created a new AI system called LaMDA. This program is specifically suited for processing natural languages like English, Italian or French. Let’s not forget that languages are very complex because they can be literal, figurative, plain, informal, formal and written in different styles. A conversation too can verge on various topics; it assumes a shared context and can be somewhat erratic. So on one end, language is one of human’s most excellent tools, yet its peculiarities make it a tough nut to crack for a computer.

LaMDA recently made the news because a Google engineer working with it declared that the system had become conscious and begun reasoning like a human being. A quick look at the conversations with LaMDA reveals that it seems aware of its existence; it exhibits emotions ( happiness and sadness) and can discuss deep subjects (like religion, justice and compassion). But does this make it human?

The answer is simply no.

First of all, exhibiting these characteristics does not mean they are real. In the same way, flying does not make an aeroplane a bird, or an underwater submarine does not make it a fish. They might indeed have similar features, but they’re very different.

Second, the engineer who made these claims also admitted that the transcript of the conversation he published is not the raw text. He edited it to ensure it was readable, but it shows that the algorithm does have significant limitations. Furthermore, hundreds of other software engineers used it, and no one else made a similar claim.

Third, our feelings go way beyond a simple calculation. When we feel sad, we experience stomach cramps and uneasiness. Happiness, on the other hand, brings a rush of dopamine. These are sensations we feel but cannot express clearly. We have no reason to think LaMDA had similar feelings during its conversation.

Fourth, we humans are social animals and tend to see patterns everywhere. As such, we must be careful when speaking with artificial beings because we might fall into the trap of believing they are human.

Of course, the million-dollar question everyone is asking is how did Google pull this feat? It started a few years back when large corporations began creating huge language models. These models are extensive collections of text downloaded from the internet and used to train AI systems. One of the most famous is GPT2, a model created by OpenAI, the company of Elon Musk. When the company revealed it in 2019, they declared that it might be too dangerous to release since it had reached almost human abilities. They fear it could be maliciously misused if it falls into the wrong hands. To give an idea of the scale of such systems, GPT2 leant from around 8 million web pages. Of course, this concern is passe; it is a relatively small model with today’s standards, and anyone can download it!

Considering that the model is three years old, today’s models seem immense. They learn from a dataset of approximately 1 billion web pages, so it is unsurprising that we are experiencing massive improvements over former systems. Just imagine that later this year, we expect the release of GPT4, which uses almost 500 billion web pages to learn. So the growth of these language models is so fast that we can anticipate some giant leaps in AI in the coming years.

LaMDA is one of these models. The difference with the GPT models is that it is specific to human language and trained on 1.6 trillion human conversations. Thus it is not surprising that LaMDA is extremely good at human dialogue. However, we can’t claim that it is conscious or that it feels emotions. Creating an AI at par with a human will take much more. Still, we live in an exciting period where conversing with computers using natural languages will become a reality pretty soon!