show index hide index
|
IN BRIEF |
|
In recent years, artificial intelligence has progressed considerably, to the point of raising questions about its ability to be confused with human intelligence. Advances in the field of machine learning have given rise to systems capable of carrying out complex tasks, to the point of calling into question our ability to distinguish an artificial intelligence from a human being. Faced with these advances, are we really capable of differentiating between these two forms of intelligence?
Advances in latest generation AI tools
THE latest generation AI tools such as GPT-4 or Claude 3.5 are today capable of generating text in such a natural way that it becomes difficult to know whether you are conversing with a human or a machine. This is a real challenge for anyone trying to make the distinction.
Return to the Turing test
This advance takes us back to the famous Alan Turing test. Aimed at assessing the level of intelligence of a machine, this test involves interaction between humans and AI to check whether the artificial nature of the interlocutor can be detected.
Recent experience with ChatGPT
Recently, the Department of Cognitive Science at UC San Diego reiterated this test using modern versions of GPT-3.5, GPT-4, and Eliza, a chatbot from the 1960s. Participants had five minutes to chat without knowing the nature of their interlocutor.
AI Response Strategies
To spice up the experiment, the researchers programmed GPT-4 and GPT-3.5 to take on the personality of a young person. This includes the use of slang language and occasional spelling errors, while incorporating information about recent news stories.
Experiment results
The results are revealing. In 54% of cases, GPT-4 was identified as human, just ahead of GPT-3.5 at 50%. On the other hand, Eliza only deceived 22% of the participants. Although real humans are still identified correctly 67% of the time, AI’s ability to imitate humans is impressive.
The future of human-AI interaction
With the emergence of Voice AI and humanoids, it will become almost impossible to distinguish a machine from a human being. Conversations on the phone could fool many of us.
Criteria
AI
Human
Ability to imitate human language
High (54% for GPT-4)
High (67% of cases)
Use of slang
Possible
Natural
Spelling errors
Scheduled
Natural
Response time
Artificial delay
Natural
Responsiveness to news
Scheduled
Natural
FAQs
Q: How did participants determine if they were talking to an AI?
A: Participants used linguistic and socio-emotional cues, as well as general knowledge questions to try to distinguish human from machine.
Q: What AIs were used in the recent Turing test?
A: The AIs used included GPT-3.5, GPT-4 and Eliza, a chatbot from the 1960s.
Q: Has GPT-4 often been identified as human?
A: Yes, GPT-4 was identified as human in 54% of cases.
Q: What adjustments were made to the AI for the experience?
A: The AIs were programmed to adopt a youthful personality, using slang language and making occasional spelling errors.
Q: Are humans always the most convincing?
A: Yes, real humans are still identified correctly in 67% of cases.
