How good are current AI researchers at simulating complex, first-person emotional states? [closed]

Question

I just read that IBM's Watson would have a hard time answering questions like "tell me about your first kiss." If you asked a modern, state-of-the-art chatbot questions like "tell me about a song that means a lot to you and why" or "tell me about a time when you felt vulnerable" would the chatbot be able to answer in ways that would fool non-experts into thinking that it was human? Are questions like this good candidates for the Turing test?

I am not asking a philosophical question about if it is POSSIBLE to generate an AI that could represent a complex internal emotional state. I am asking: given the state-of-the-art research in 2013: how close are current researchers to generating AI that could pass the Turing test?

Answer 1

Great Question! One area to look into is Cognitive Architectures. These are computer programs that solve AI problems by modeling how humans think and feel (emotion). Here is a brief overview.

Prior to the mid-1980s the study of AI was concerned with creating general intelligent systems modeled after flexible human thinking (Fahlman, 2012). Researchers in AI used findings in cognitive psychology to guide their work.

Currently, there has been a shift away from this approach. The current trend in AI research is towards solving problems with “knowledge-lean” statistical methods which are data intensive and learn less efficiently than humans. (Langley, 2012). Thus, the AI field has shifted away from human-like models and toward "ideal" models.

However, a few AI researches still focus on creating intelligent systems modeled after the way humans think. These researches work with Cognitive Architectures--Computer programs modeled after humans. The Office of Navel Research's Human Computer Interaction lab is actively working to create intelligent agents that can interact with humans.

I believe that the study of Cognitive Architectures is an interesting area of research for investigating how to create human-like systems.

Fahlman, S. E. (2012). Beyone Idiot-Savant AI. Advances in Cognitive Systems 1, 15-21.

Langley, P. (2012). Intelligent behavior in humans and machines. Advances in Cognitive Systems 2, 3-12.

Taatgen, N. A. (1999). Learning without limits. From Problem Solving towards a Unified Theory of Learning(dissertation). Groningen: Rijksuniversiteit Groningen.

http://www.onr.navy.mil/en/Science-Technology/Departments/Code-34/All-Programs/human-bioengineered-systems-341/Human-Robot-Interaction.aspx

Answer 2

there are several different questions here.

yes, questions/topics relating to emotional states in the chatbot or the human conversing with it are those that would tend to trip up chatbots.

a good semi-scientific benchmark for research in this area is the Loebner contest/competition. it somewhat a mix of serious research and informal gaming.

a nice article/overview on the competition was published a few yrs ago in the Atlantic, Mind vs machine.

another area where there is a combination of ongoing research, application, and innovation is the field of gaming, and "NPCs" or "non player characters". one of the more advanced/notable games from recent years with a theme of simulating emotional states of the agents in the game is the Sims written by Will Wright.

how close are current researchers to generating AI that could pass the Turing test?

there is no definitive answer here, only opinion, conjecture, speculation, and science fiction, and some disconnected research. at heart the question asks to predict the future. the short answer is that nobody really knows how far away this milestone is or if it is even scientifically achievable (although the question of machine intelligence on the level of human intelligence is probably a scientific question, subject to study, experiment, evolutionary and/or revolutionary advances etc, rather than a philosophical one as you indicate).