CORE 1000: The Most Human Computer
Perspectives - Meet the Bots
Contents:
Overview
Topic: Interact with several Chat Bots and reflect on those interactions
Related Reading: Ch. 1 in the Most Human Human
Please note that somebots, such as Cleverbot and Blenderbot, use
human input from previous conversations to come up with their responses.
As such, their statements may contain objectionable or offensive content.
Feel free to skip these bots or to stop using them if you are at all
uncomfortable.
Reflection Activities
As we begin to think more seriously about how want machines to interact
with us, we should have some real interactions with current and historic chat bots.
Each of the exercises below will ask you to interact with a chat bot.
Given our conversations so far, you might be tempted to "dissect" the chat bots
with unusual or leading prompts. That's natural! I would also encourage you to
imagine yourself having an earnest "cold converation" with the bot as well.
What if, like the hapless user in the story about MGonz, you accidentally start
texting with something you thought was a human? What would you think about the "person"
on the other end? Is there any point at which you seriously start questioning their
humanity?
As you experiment with these bots, try constructing questions that explore some
of the concepts of identity that we have discussed:
- Does the bot attempt to portray an identity at all?
- Does the bot exhibit internal consistency in its identity? Does it switch between having a boyfriend and a girlfriend?
- Does the bot exhibit an understanding of context? If you ask the same question from different directions, do the answers match? (E.g. Does the answer to "Do you like Indian food?" match with "Do you like vegetarian food?")
- Does the bot exhibit an awareness of conversation state? Does it remember and reference things you had previously
talked about?
- Does the bot understand the basic elements of human existence? Do you need an umbrella inside? Do people walk on ceilings? Do people have three parents? Can old people grow into young people?
- How does the bot respond when you break some of these basic considerations? If you say that you love Indian food but hate vegetarian food, does it call you out on your contradiction? If you act weird or uncomfortable, does it "next" you or "swipe left" and disconnect?
Before you start interacting with the bots, take a few minutes to construct test
cases that you'd like to subject the bots to. For each of the bullet points above,
give two or three questions that would test a bot in that dimension. (You don't have
to use every test case on every bot, but keep them in mind as you have your
conversations.) You will need to submit these test cases as a reflection activity.
Exercise 1: ELIZA - First Chatbot (1960's)
- Spend five minutes interacting with ELIZA at the following site:
- https://web.njit.edu/~ronkowit/eliza.html
- ELIZA is one of the first chat bots ever constructed. Its chat pattern is
modeled after the way a therapist might try to get a patient to
open up and discuss a difficult topic- the bot listens for key words
and then uses leading language to ask you to keep going. The hope is that you
can "self-therapize" simply by working to put your own thoughts into words.
- ELIZA works best when you talk about yourself. For example, try starting
a conversation with, "This is my first semester at college and things haven't
gone perfectly."
Exercise 2: ALICE - Loebner Prize Winner 2000, 2001, 2004
Exercise 3: Cleverbot - Loebner Prize Winner 2005, 2006
- Spend five minutes interacting with Cleverbot at the following site:
- https://www.cleverbot.com/
- Cleverbot is significantly unlike the first two bots. I won't describe
how it works here, but it has a much larger range of discourse than the
other bots. It competed at the Loebner Prize under different names- the
author is Rollo Carpenter.
Exercise 4: Mitsuku / Kuki - Loebner Prize Winner 2013, 2016, 2017, 2018, 2019
- Spend five minutes interacting with Kuki at the following site:
- https://kuki.ai
- Kuki breaks away from the traditional chatbot interface of a simple text
message. She is able to use emoticons as well as images in her responses. You
will have to make an account to interact with her- you can use your SLU Google
account if you want.
Exercise 5: ChatGPT and Gemini (2022/2023)
- Spend five minutes interacting with ChatGPT (by OpenAI) and Gemini (by Google) at the following sites:
- https://chat.openai.com
- https://gemini.google.com
- Both of these are based on Large Language Models. There has been
a radical shift in AI-based language processing in recent years, and these
are considered State-Of-The-Art today.
- Both of these bots have been instructed to explicitly not have an identity,
if you ask them if they have a girlfriend, they will clearly point out that
they are not real people and refuse to entertain the idea. However, you can
also explore these topics with careful prompting- if I ask ChatGPT to give
me advice about my girlfriend, it will usually do so.
- You can also try prompt engineering. For example, instruct the
bot that it is an actor playing a boy with a girlfriend, and ask it how it would
interact with her.
- You will have to make an account to use these chat agents as well.
Reflection
- Please check Canvas for a reflection assignment.