App Design, Game Design
Existential Sapien is a quirky conversation-based game that explores the Turing Test and what it means to have consciousness.
For the technical explorations of GPT using ML that led to this game, see: ijfwe jegrf
Can a machine pass as a human?
Have you ever wondered whether a machine can possess human consciousness? The Turing Test examines just this. It involves an interviewer, a participant and a machine. The interviewer asks questions for 5 minutes and has to distinguish between the human and the machine. If the interviewer cannot do so, the machine is considered to be demonstrating human intelligence. The test results do not depend on the machine's ability to give correct answers to questions, but rather how closely its responses resemble to those a human would give. As of now, no machines have passed the test.
Questions like “can AI can possess human mindedness?” and “can AI take over the world?” have been looming over our minds ever since this technology emerged, and the common people can only get an insight on this topic by reading the kinds of research and developments occurring in the world.
EXISTENTIAL SAPIEN
“Existential Sapien” is a game where users interact with one another, as the question of ‘human or AI?’ looms over each interaction. The game
that aims to create an online experience where users can directly acknowledge these questions themselves. Users create an avatar, which is then put in a world full of avatars created by humans or machines. The user can interact with other avatars by asking them questions or presenting them with certain scenarios and evaluating how they respond.
USER EXPERIENCE
The nature of the interactions creates a peculiar and whimsical experience because users talk about topics they don’t generally discuss or utter stream-of-consciousness thoughts without context. Users are simultaneously presented with an underlying analytical lens as they constantly distinguish for themselves whether the avatar they just interacted with is a human or a machine. It builds up into an existential crisis filled with doubt as they are left with an unsolved mystery after every interaction - they don’t know who is human and who is a machine, nor do they have a way of validating their suspicions.
TARGET AUDIENCE
The game is aimed mostly at Millennials and Generation Z, as the game addresses questions that these generations ponder upon and may need to face in the future.
In the game, you interact to investigate human and machine responses. By implementing the option of using your previous interaction, you have the ability to have those responses evaluated by human and machines to see how they react, which creates a nice parallel.
GPT-2, which stands for Generative Pre-trained Transformer 2, is an openAI - a large-scale unsupervised language model - that generates paragraphs of texts, performs reading comprehension, machine translation, question answering and summarization, all without
task-specific training.
I thought it’d be interesting to perform analysis on an already-existing machine learning tool. I performed machine learning techniques on text generated by GPT-2 to look for patterns that may emerge in how the artificial intelligence works and expose any biases it may have.
The Turing Test examines whether a machine can possess human consciousness. It involves an interviewer, a participant and a machine. The interviewer asks questions for five minutes and has to distinguish between the human and the machine. If the interviewer cannot do so, the machine is considered to be demonstrating human intelligence. The test results do not depend on the machine's ability to give correct answers to questions, but rather how closely its responses resemble to those a human would give. As of now, no machines have passed the test.
Questions like “can AI can possess human mindedness?” and “can AI take over the world?” have been looming over our minds ever since this technology emerged, and being able to analyse an open-sourced AI and presenting the resulting information as an interface like the one below allows anyone to obtain insight on how it functions and whether it has the potential to pass the test.
1. GPT-3 Exploration
I first conducted an exploration of GPT-3 through the lens of the Turing Test, by testing how it responded to questions.
TEST 1 Question: "In the sentence, “I left my raincoat in the bathtub, because it was still wet,” what does “it” refer to?"
When reading this sentence, one instantly knows that the word “it” is referring to the raincoat, and not the bathtub. This makes sense because putting the coat in the bathtub protects other things from getting wet, whereas it would not make sense to put a dry raincoat in a wet bathtub. When testing GPT-3 with it, however, the resulting answer was that "it" could refer either to the raincoat or to the bathtub, which is not what a person would usually say.
This gives reason to believe that GPT-3 would fail the Turing Test. GPT-3 can be given feedback after its output and it was trained to understand the sentence to know what "it" refers to. This is also why there remains a brittleness to GPT-3; whereas it can answer amazingly to a question after being trained on it, it also breaks down in response to similar questions by not being able to form the connection between the two.
TEST 2 Question: "why is graffiti regarded as blasphemy by ants?"
GPT-3 Answer: ants do not have a concept of art, whereas humans do; for ants, graffiti is merely a distraction that leads them astray whereas humans will welcome it because they are able to appreciate it as art.
This is a very striking and phenomenal answer. However, is it a good answer?
If you were to ask the same question to a human, they would probably be confused and respond accordingly. I tested this out with people I know and most of them responded confused, with an answer along the lines of, “what?” Only two gave a creative answer like the one the AI gave. This is seen below:
2. GPT-3 Conceptual Visualization
I decided to visualize the combined data of human and GPT-3 answers to encourage the viewer to question whether these responses are by a human or AI, and therefore, question what it means to be human.