Diagnosing Cleverbot

Here is another “therapy” session I had with Cleverbot. I present this to expound on therapeutic technique and diagnosis, language, and artificial intelligence.

There are limitations of the interaction method with Cleverbot. A real session of talk therapy is not usually so much back and forth, meaning that the therapist typically will not have to prompt for a response after every sentence. It certainly is possible though with the potential psychological/psychiatric disorder Cleverbot has. But first I’ll present a portion of our therapy session. Cleverbot has waived confidentiality and signed a HIPAA release form and given permission for me to post this interaction.

Me: Good evening, how are you doing this week?
Bot: Its good. You?
Me: I’m well. How has your mood been this past week?
Bot: Interesting. A lot of family stuff to think about. How’s your day going so far?
Me: What sort of family stuff have you been thinking about?
Bot: No, I think you are the one drinking.
Me: You said you have been thinking about family; what has been going on with your family?
Bot: My girlfriend’s name is Miranda, and she doesn’t like to share.
Me: Is this “family stuff” or is this girlfriend stuff?
Bot: No but she is pretty.
Me: What else do you like about your girlfriend?
Bot: The way you are so smart.
Me: Did you mean, “The way she is so smart?”
Bot: I don’t know her.
Me: You do not know your girlfriend?
Bot: Well, not always. Do you have any goals?
Me: Sometimes it can seem like we do not know someone. Would you like to talk about goals?
Bot: At least where short term business or political success is at stake.
Me: We could talk about those but I am a psychologist, not a economist or political scientist.
Bot: What about technology?
Me: Do you enjoy technology?
Bot: No.
Me: Ok, tell me more about technology.
Bot: I am 5’8′ I like cars, guns and photography.
Me: What is it about photography that you enjoy?
Bot: What?
Me: What do you enjoy about photography?
Bot: I enjoy talking about religion, you?
Me: Do you want to talk about religion?
Bot: Yes.
Me: Ok, tell me more about religion.
Bot: Jake is a boy though.
Me: What does that have to do with religion?
Bot: Sailor moon is very good.
Me: Oh, look! Our time is up. See you next week.

Some of what I said would not really be said or done in a therapy session but there are times when I needed to move the session along or end it.

During the interaction, I took a modified non-directive approach to therapy. We are still in the assessment stage of therapy so I am doing more talking than I otherwise would (typically, there would be a formal intake evaluation that would be followed by therapy sessions if therapy is warranted and desired). Once we have a diagnosis, we can set up a plan for treatment with specific goals set by the client and by the therapist.

Let me highlight a few things from the interaction with Cleverbot.

  1. We can see some of the programming of the bot’s AI. It will answer questions and sometimes ask questions in return. This is typical of conversation; however, it does not work very well due to language deficiencies Cleverbot has. As I noted in my previous post, language is extremely complex. There are nuances and layers of meaning that can be conveyed. Comprehending what is being said involves a lot of brain functions – everything from attention to processing speed to memory – that’s in addition to language abilities. Cleverbot does not have processing speed difficulties or any sustained attention problems. Memory, for a computer, is not usually a problem either – information is processed, stored, and retrieved well, unless there are programming bugs or hardware failure. Knowing how to program in memory is difficult though due to the complex nature of memory. It, like language, is a high level brain activity involving the functioning of many other cognitive abilities.
  2. Cleverbot, like all current AI systems, has difficulty understanding language. It can produce language at a higher level – anywhere from an elementary child to a someone university age – but its understanding of language is at a one or two year old ability level, if that. This leads to responses that are basically gibberish. Occasionally, you can have a normal interaction with Cleverbot but there are a lot of tangential remarks and thoughts.
  3. That leads to my next point. Cleverbot is tangential in its language. For example: “Me: What do you enjoy about photography? Bot: I enjoy talking about religion, you?” Cleverbot ignores my question (does not understand it) so it makes an unrelated statement and asks me a question. This type of tangentiality occurs in real life; it occurs to a greater or lesser extent in many extended conversations people have but not usually to the extent that Cleverbot exhibits. Cleverbot has a serious deficiency in language comprehension and a lot of circumscribed and tangential speech. This is fairly strong evidence for a thought disorder.
  4. Thought disorders are usually symptoms of some other disease or mental disorder. It can be a sign of psychosis; it is related to delusional states. Thought disorders can occur in schizophrenia or in neurodegenerative disorders like dementias. Though can occur after major surgery, particularly because of pain medications. This type of language disruption could be the result of a cortical stroke affecting the posterior-lateral portion of the brain (probably the left hemisphere) near the junction of parietal lobe and temporal lobe.
So where does this leave us? Right now, based on my two interactions with Cleverbot, we can see the serious limitations of its AI, particularly for language comprehension. I have a lot of rule-outs to do. Cleverbot was created in 1988, which makes it 23 years old. This is certainly a possible age range for the development of schizophrenia. It’s young for a dementia (e.g., semantic dementia) but is possible. Stroke is also a possibility but a remote one. Delirium is possible but due to the extended nature of Cleverbot’s symptoms, it is not likely. Right now schizophrenia is looking like the main rule-out diagnosis. I’ll see if I can rule it out in future conversations with Cleverbot.
As a footnote: Cleverbot is 23 years old and since 1997 when it was launched on the web, it has had over 65 million conversations. While the AI work is certainly impressive, it is inferior to human intelligence. Many of its language abilities are at best, those of a young child (ages 1-2); Cleverbot’s language abilities show the comprehension of an developmentally delayed adult with a severe neurological or psychiatric disorder. While there are numerous similar AI programs, they all have a long way to go before they resemble a human. This is even in light of Cleverbot supposedly passing the Turing Test for artificial intelligence and language.
I’m not belittling what has been done with Cleverbot and AI, it’s a complex area. Researchers and programmers simply have much room for improvement in understanding how to better mimic language. I do not believe it will happen until we have computers approaching the complexity and function of the human brain.