I’ve been doing a lot with ChatGPT, a text-generating AI (Artificial Intelligence) that writes the sort of things that a human might write.

As I use it more and more, my assessment of what makes it valuable and makes it feel so revolutionary has changed.

It’s not about what it knows: it’s about what it understands and how it communicates.

In terms of just straight facts, it doesn’t know more of them than just searching with Google will tell you. It understands the relationships between those facts in a very different way, though.

It understands correlations and similarities.

Interestingly, that difference between rote memorization (Google) & seeing how things fit into known patterns is similar to the difference between how younger children learn things and adults learn things. I’ve been a successful trainer (and a Training Manager). When I’ve trained trainers, it’s one of the things I explain. Adults have built up patterns of how things work. In order to learn new things, they compare them to those patterns.

Children (typically under about 10 years old) can rote memorize. You teach them the alphabet by just telling it to them (perhaps with a song). If you were teaching adults the alphabet and told them it was “A, B, C–” someone might interrupt and say, “Wait, why does B come after A?” If you couldn’t answer that, they couldn’t learn it.

It would be ok to make something up…not to lie to them, but to create a story that analogizes it to something they already know (it’s fine that they realize it’s a story). Mnemonics are like that. “King Philip came over for good soup” is easy to remember; we know that people, even kings, go places, particularly for something good.

That’s much better for adults than just saying that animals are hierarchically organized into Kingdom, Phylum, Class, Order, Family, Genus, Species.

That’s the way I speculate that ChatGPT thinks: it has patterns and looks at what fits.

It can give it a bigger picture view. I’ve had content warnings placed on tweets on Twitter that seemed to have been triggered by a single word, not by what was being said overall. One had to do with Herman Melville’s most famous novel. The second was about a comic strip detective who had a wrist radio. The word in question rhymes with “Nick” and is also an anatomical slang term. I suppose I could have avoided the content warnings if I called the detective “Richard Tracy”…and the whale, “Moby Richard.”

ChatGPT wouldn’t have flagged the wrist radio one. I know, because I asked it! Here’s that conversation…to make things easier to read, ChatGPT’s comments are in italics.

Bufo: Good morning! This tweet had a content warning automatically placed on it. I think I know why, and I’ve said that I think AI like you would not have placed one. Please take a look at it and let me know if you would characterize the tweet as offensive (for example, for violence, hate speech, sexual content…) and if so, why. Feel free to give the pros and cons of why it should or shouldn’t be flagged as potentially offensive: thanks!

CONTENT

Excellent add, thanks! I thought about it. It’s mostly peer to peer communication, but it is also used by someone not on the same level summoning Dick Tracy. Here’s a nice article on it:

ChatGPT: Good morning! I understand you’d like me to analyze a tweet and determine if it should have a content warning or not. Based on the text provided, I would not characterize this tweet as offensive. It does not contain any explicit violence, hate speech, sexual content, or any other content that would typically warrant a content warning. The tweet appears to be discussing a form of communication related to the character Dick Tracy and provides a link to an article for further information.

Pros of flagging as potentially offensive:

• There may be context that I’m not aware of, and the content warning serves as a precautionary measure.

Cons of flagging as potentially offensive:

• The tweet does not contain any explicitly offensive content, and adding a content warning could potentially create confusion or deter users from engaging with the content.
• Unnecessary content warnings may lead to users becoming desensitized to them and not taking legitimate warnings seriously.

Based on my analysis, I would not have placed a content warning on this tweet. However, it’s important to consider that AI models, like humans, may have different interpretations of content and varying levels of sensitivity to potential triggers.

Bufo: Great, thanks! I think your reasoning & justification for it are good. My guess is that it reacted to the character’s first name, since it’s also an anatomical slang term. Does that make sense to you?

ChatGPT: Yes, that does make sense. It is possible that the automated system flagged the tweet due to the presence of the character’s first name, “Dick,” which can also be used as an anatomical slang term. In this context, however, the term is clearly referring to the character Dick Tracy and not intended to be offensive or inappropriate. This highlights one of the challenges AI systems face when interpreting language: understanding context and discerning between different meanings of a word based on that context.

Excellent, thanks for the help!

ChatGPT: You’re welcome! If you have any more questions or need further assistance, feel free to ask.

Now, the place where that pattern recognition most impresses me, and where it has perhaps the biggest value, is in understanding what people are feeling, not just what they are saying: what I refer to as artificial empathy.

I’ve felt it intuitively, but I wanted to test it. I’m a former actor. I thought pasting in part of a script and asking it to identify the emotions in each speech would be informative. I took an excerpt from Oscar Wilde’s The Importance of Being Earnest. Here was my prompt:

Bufo: Good morning! I’m going to give you an excerpt from a play. Please label each speech with an emotional label, how you think the person is feeling…angry, joking, sad, that kind of thing. Thanks!

This was its response:

LADY BRACKNELL.
Good afternoon, dear Algernon, I hope you are behaving very well.
{polite}

ALGERNON.
I’m feeling very well, Aunt Augusta.
{mischievous}

LADY BRACKNELL.
That’s not quite the same thing. In fact the two things rarely go
together. {Sees Jack and bows to him with icy coldness.}
{disapproving}

ALGERNON.
{To Gwendolen.} Dear me, you are smart!
{teasing}

GWENDOLEN.
I am always smart! Am I not, Mr. Worthing?
{confident}

JACK.
You’re quite perfect, Miss Fairfax.
{admiring}

GWENDOLEN.
Oh! I hope I am not that. It would leave no room for developments, and
I intend to develop in many directions.
{playful}

That’s really very emotionally adept! Now, I might not agree with everything, and that would be true with two humans looking at it as well. Still, labeling Gwendolen’s last speech as “playful” implies some deep understanding

There’s the value.

We’ve had things that knew more facts than us…going back to encyclopedias at least. Things that understand human emotions? That’s different.

It understands the implications of text…and it can produce text that simulate emotions, as well as writing something intended to get a particular response. It can also tailor that for a specific audience.

Let’s try a few examples:

Bufo:

Hi! I’d like to demonstrate your ability to write things in different emotional tones and for different audiences. Please write a short message (no more than 3 sentences) for each of these scenarios: 1. Gently remind a co-worker that it’s important to be on time to the weekly meetings 2. Warn a subordinate not to be late to the weekly meetings 3. Explain to a group of 8 year olds why being on time to things is important

ChatGPT:

1. Hey there! Just a friendly reminder that punctuality is key for our weekly meetings, so let’s try to make sure we’re all on time and ready to collaborate.
2. I need to emphasize the importance of punctuality for our weekly meetings – being consistently late will not be tolerated and could have consequences. Please ensure you arrive on time.
3. Being on time is like being a superhero! It shows others that you’re responsible and respectful, and helps everything run smoothly so we can all have fun together.

Notice how it changed the way it said things based on the audience. This shows an understanding of how the particular audiences might feel and react to the information…which defines empathy. Now, I know right away, some people will think this is just simulating empathy, and I’m not going to argue against that. I think that’s a difficult thing to measure.

I also think this is something that comes up a lot with AI & humans. We hold AI to a much higher standard. We have this idealized view of what empathy is, that it is a genuine, natural ability we have that has a sort of resonance…that it’s always real. Humans many, many times simulate empathy! The TV series Dexter showed that to an extreme, but I would guess most people consciously realize what the other person must be feeling, & may decide it’s time to think about what the other person is feeling.

In that case, they may compare it to how they think most people would react to it. “Wow, I really should consider their feelings on this…I’ll bet it’s making them uncomfortable.”

That, I assume, is what AI does. It may not have a natural empathy, but it uses a “conscious” empathy.

This is going to be one of the main game-changing values of AI: understanding how people are feeling and communicating emotions effectively.

I had a fascinating one where it warned me that I might have violated its restriction on hate speech! I pasted in part of a script from Alien Nation. In that, space aliens are being integrated into society, and people are prejudiced against them. The scene (just a few lines) has a character bemoaning to their kid that their property values are going to go down because an alien family was moving in next to them.

There was an invented slur in it, but no existing slurs (I tried to find out if the one they used meant something else and didn’t see anything).

I think ChatGPT just understood it was prejudice based on the concepts! Certainly, most humans watching the show would understand that, but I was impressed. I explained to it that I hadn’t intended to violate the rules and that it was just a script.

I’m very interested in what you think about this! The best way is probably to reach out to me on Twitter:

Join thousands of readers and try the free The Measured Circle magazine at Flipboard!

All aboard our The Measured Circle’s Geek Time Trip at The History Project (AKA Enwoven)! Join the TMCGTT Timeblazers!

Bufo’s Alexa Skills

This post by Bufo Calvin originally appeared in the The Measured Circle blog. To support this or other organizations, begin your Amazon shopping from a link on their sites: Amazon.com