GPT-4 passed the Turing test: is it a truly intelligent AI or just a good imitator?

OpenAI’s GPT-4 language model has passed the Turing test, a test that evaluates whether a machine can imitate human behavior indistinguishably. This achievement not only reflects the advanced capacity of current models of artificial intelligencebut also raises the question of whether GPT-4 represents true intelligence or simply a sophisticated imitation.

Specifically, a study from the University of California in San Diego asked 500 people to talk with three models: GPT-3.5, GPT-4 and ELIZA. After five minutes of chatting, participants had to determine whether they were talking to a person or an AI. 54% of participants identified GPT-4 as a real person.

Scientists were surprised to find that more than half of the participants did not distinguish between an AI and a human. 54% thought GPT-4 was a real person, 50% identified GPT-3.5 as an AI, and 22% believed ELIZA had human-like qualities. Researchers also warned that human-like AIs could disrupt the economy and society, causing distrust in real interactions.

Now, although GPT-4 has passed the Turing test, demonstrating its great capabilities, does this mean that it is truly intelligent? To answer this question, El Comercio spoke with Giovanni Geraldo Gomes, director of AI at Stefanini Latam, and with Deisy Acosta Ticse, professor of Information Systems Engineering at the Universidad San Ignacio de Loyola.

True intelligence or good performance of intelligent behavior?

Giovanni Gomes describes OpenAI’s latest model passing the Turing test as “very important and makes clear the great potential of AI and the speed with which LLMs are evolving”. The expert has a positive view of this achievement and is convinced that “It is fair to consider it as a true intelligence”.

For his part, Deisy Acosta points out that the performance of GPT-4 does not represent “not true intelligence, but good intelligent behavior.” She clarifies that there is still a long way to go to achieve Artificial General Intelligence (AGI), which would have the ability to understand, learn, creatively apply knowledge and adapt to new situations in a way similar to a human being.

“In this sense, the development of GPT-4 represents an important step on a longer path towards the development of truly intelligent AI.“, Add.

Limitations of the Turing test

Deisy Acosta points out that the test created by Alan Turing in 1950 “has limitations as a tool to evaluate intelligence in general”. She specifies that GPT-4 can pass the Turing test with its conversational skills, but it does not evaluate essential aspects of intelligence such as logical thinking, creativity or conscientiousness.

Furthermore, he reveals that this methodology has been criticized since the creation of the Loebner Awards in 1991. This annual competition focused on evaluating whether artificial intelligences can convincingly imitate human conversations. Inspired by the Turing Test, it was established by Hugh Loebner with the purpose of determining which AI program can appear human during a textual conversation. However, over the years, there has been criticism that some participants focus more on superficially replicating human behavior than on developing real, creative cognitive intelligence.

Stefanini’s spokesperson also explains that the test “It does not take into account conversational style, emotional factors or creativity, which are key aspects to affirm that machines have traditional notions of human intelligence.” cWhen the Turing test was invented, artificial intelligence was not as developed as it is now with large language models (LLMs), he maintains.

In this regard, both experts agree on the need to develop new methodologies to evaluate artificial intelligence in a more precise and comprehensive manner.

“Metrics that cover a broader spectrum of cognitive abilities are required. Metrics from natural language processing (NLP), pattern recognition and speech processing are essential to measure a chatbot’s inference, reasoning and language comprehension capacity.”, Acosta explains.

Specific metrics are thus essential to assess AI in its specific field, but for a complete assessment, especially in relation to Artificial General Intelligence (AGI), a broader approach is required. Ethical and accountability assessments also come into play.

“On the positive side, these advances show that AI is getting closer to reproducing the capabilities of human intelligence, which would surely generate many benefits.”, recognizes Giovanni Gomes. However, he adds that this raises concerns in society because there is a risk that this intelligence could be misused, misleading the population and undermining trust in authentic interactions.

Meanwhile, Deisy Acosta stresses the need to prudently and responsibly manage the development of AGI, carefully assessing both its benefits and potential risks. The fate of AGI and its impact on humanity will depend on the decisions taken: we must take advantage of its advantages to improve the quality of life, but at the same time be prepared for the challenges that may arise.

“AI ethics requires that this technology not only be effective, but also responsible and beneficial to society.“, he notes.

The teacher emphasizes that there must be complementarity between deontological ethics, which is based on duties and rights, and teleological ethics, focused on consequences. Both are essential to guide the development of AI. Human intervention must ensure that AI is developed under these ethical principles, establishing clear frameworks that regulate its design, training and use for the benefit of human well-being.