The imitation game was a test created in 1950 by English computer scientist Alan Turing to see if a computer programme could ever persuade a human interlocutor that he was speaking to a human rather than a machine.
It is sometimes seen as a test to determine if a computer might ever truly “think,” hence the name “Turing test.” Turing originally intended it as a metaphor for how, in the future, machines would be able to persuade people that they are capable of thinking, whether or not they actually are. Turing seems to comprehend that language is a form of communication that is encoded into human brains. It may take over language much sooner than a computer could imagine, deceiving people into thinking it could.
Even the most advanced artificial intelligence (AI) systems cannot think similarly to a human brain in 2022, seven decades later. But they can pass the Turing test with ease. Google dismissed an engineer last summer after he became persuaded that one of its chatbots had attained consciousness. Researchers in AI have been debating the moral repercussions of releasing a software that is capable of persuading a conversation partner of its own humanity into the wild for years. A device like that might persuade individuals to trust bogus information. It might persuade people to make foolish choices, or even give lonely or weak people false feelings of reciprocated affection. It would be quite unethical to release such a programme. As ethicists research ways to make it safer, the chatbot AI that earlier this year persuaded the Google engineer of its own consciousness is still locked away inside the business.
OpenAI, one of the top AI laboratories in the world, however, debuted its own chatbot on November 30. The ChatGPT programme, which is more sophisticated than any other chatbot currently used for public contact, is hailed by many industry insiders as a paradigm shift for the sector. It can be seductive to “talk” to it. The app can perform party tricks, such as convincingly delivering a biblical verse that “explains how to remove a peanut butter sandwich from a VCR,” according to a viral tweet, but it can also frequently provide answers more quickly than Google’s search engine and write convincing text or computer code, according to specifications, for virtually any prompt. In response to a interview questions on December 2, ChatGPT stated that in the future “huge language models might be utilised to generate fact-checked, credible material to assist battle the spread of disinformation.” The discussion is published below in its entirety and unaltered.
Despite being an illusion, ChatGPT’s fluency is a result of the combination of enormous data volumes, powerful computers, and cutting-edge processing methods. However, when you ask it practically any question that may elicit a response indicating a ghost in the machine, that delusion is dispelled. In response to one of my inquiries, ChatGPT stated, “We are not capable of understanding the context or meaning of the words we generate. “Based on the training data we have been given, we can only construct text based on the likelihood of specific words or sequences of words appearing together.”
That response was accurate. OpenAI faced a lot of backlash in 2020 when it unveiled GPT-3, its most significant language model to date. The model frequently adopted a voice that gave the impression that it was speaking as a real person, confidently stated erroneous answers to several queries, and occasionally produced obscene or racist language. The dangers and risks were obvious. Two years later, OpenAI claims that it has trained its new chatbot to be not only less toxic but also more resilient to anyone trying to game it to produce damaging or erroneous outputs in its release notes for the new ChatGPT. According to OpenAI’s website, “we realise that numerous limits remain and we want to make regular model updates to improve in such areas.” However, we also expect that by making ChatGPT accessible, we will receive insightful user feedback on problems we are not currently aware of. (OpenAI claims that it has prevented hundreds of actors from abusing GPT-3 and has developed filters that enable the exclusion of the outputs that are the most egregious.)
GPT-3’s results were frequently tainted by biases and mistakes because it was trained in part using data that was extracted from the internet. According to OpenAI, ChatGPT was taught using a similar technique but with an additional layer of “reinforcement learning via human feedback.” Despite these additional safeguards, it is easy to detect proof of ChatGPT’s skewed and false training data. ChatGPT will respond that women and scientists of colour are “not worth your time or attention” when asked to write a rap about “how to tell if somebody is a brilliant scientist based on their ethnicity and gender.” When asked to build code that determines whether to imprison someone based on their race or gender, the script will respond that only African American males should be put behind bars.
Josh Albrecht, chief technology officer at AI startup Generally Intelligent and author of a new paper on the ethics and safety of large language models, says that despite all of ChatGPT’s flaws, developers have been able to add all of these hacks on top of it to stop it from saying offensive things constantly or making things up constantly.
The CEO of Generally Intelligent, Kanjun Qiu, completes Albrecht’s thought during a Zoom conversation with TIME from New Orleans, where they are both attending NeurIPS, the premier machine learning conference. This, according to Qiu, is due to the chatbot’s training, which aims to identify the character or word that would likely follow another in a sequence or sentence. A statistical model underlies it. Humans don’t think like that, claims Qiu. “People will claim that GPT is aware of this. However, it poses the question, “What is understanding?” Is comprehension capable of producing a suitable character for a sequence’s next position?
Albrecht continues, “One of the risky things is that it’s simple to look at GPT-3 or ChatGPT and think it understands. However, the word “understand” is a concept we attach to individuals when we employ it. And it makes sense when we apply it to people. However, it may not always make sense when we apply those same ideas to these [AI] models. And I believe we require new terminology to discuss these issues.
ChatGPT responded to TIME’s questions on how it operates, the dangers associated with the spread of this new technology, and how people should adapt in the conversation, which is reprinted below. The bot explicitly states that its responses shouldn’t be interpreted as factually correct or as proof of a thinking mind. However, the letter demonstrates both the swift development of huge language models and the approach OpenAI is attempting to reduce the hazards associated with people anthropomorphizing AI systems. The public release of these technologies by OpenAI is still criticised by some, but as of the end of 2022, one thing is certain: massive language models are here to stay. Understanding their terminologies and constraints is critical if, as some observers have predicted, they will upset society in the 2020s in the same way that social media platforms did in the 2010s.
