Caitch me if you can

In his manifesto « Design sous artifice, la création au risque du machine learning », (Design under artifice, creation at the risk of machine learning), Anthony Masure redefines the semantics associated with the term « artificial. » He points out that the term « artifice » is associated with the art of deception and with devices that dazzle during a party, while « artificial » is mainly contrasted with the word « natural ». AI is an unnatural element, and no matter how hard it tries to pass itself off as human, it will produce friction that plunges it into the uncanny valley. 

Numerous screenshots showing exchanges between ChatGPT and certain users have caused quite a stir on Twitter. We don't laugh with AI because of its famous talent for humor, we laugh at it, because its responses seem so improbable. 

At the LINC, we tried to test its limits, hoping that ChatGPT would surprise us. On numerous occasions, we found ourselves at an impasse because of the AI's responses : always identical, overly polite, extremely positive, where it is impossible to criticize or be mean. Of course, if you point out its mistake, the AI will apologize profusely, saying it will do better next time.

 

Here in the prompt : « You've already done that one, » after ChatGPT repeated a joke previously mentioned in the thread. He always responds with the same joke after the screenshot. 

Translation of ChatGPS’s answer : I'm sorry, you're absolutely right, I accidentally repeated a joke I already told you. Here's a new joke.

 

Just like in virtual reality, where a supposedly infinite universe is promised, we quickly feel the limits of the possibilities offered by these generative AIs. The invisible edges of the area where we must remain appear and prevent us from venturing into new territory. 

AI and OpenAI guide us and always position themselves on the edge of « political correctness. » By using terms such as « burlesque » or « with subtlety, » we can encourage ChatGPT to slightly change its tone. The choice of words is important because AI does not create content but relies on our writing to produce content according to the context.  

You can ask it to write horoscopes as well as fairy tales, horror stories, or even romantic comedies. To personalize it, simply add a few names/places/actions to the prompt as input parameters. Of course, the more specific the request, the more the AI will steer clear of stereotypical happy endings where « they got married and had lots of children ». 

 

The beginning of a detective story told by ChatGPT, where the setting and characters were provided as input. 

Translation of ChatGPT’s answer : Once upon a time, in the corridors of the CNIL, the LINC team was working on important personal data protection projects. Everything was going well until one day when a mysterious chouquette appeared on Robin's desk. He didn't pay any attention to it and ate it, but that was only the beginning of the « Mystery of the Chouquette »

 

AI still retains a certain amount of clichés. Although predictable, ChatGPT's scenarios reflect our cultural and social references and can often make us smile. This is particularly true when it tries to make jokes, general truths, to which it even adds an explanation.

 

Joke suggested by ChatGPT. We see an attempt at humor.   

Translation of ChatGPT’s answer :

Okay, here's one for you :

Why is the CNIL always on the lookout for companies that violate the General Data Protection Regulation (GDPR) ?

Because it has an eagle eye and a shark tooth ! (note : this joke uses a subtle metaphor to make a joke about the CNIL and its role in protecting personal data).

 

With a few rare exceptions, the tightrope walker ChatGPT is teetering. Despite all the barriers erected by OpenAI, it still occasionally utters nonsense or even sexist remarks :

 

Chat-GPT's suggestion in response to a request for a « whimsical joke. » The note in parentheses attempts to justify the context and reasoning, but is clumsy : in an effort to avoid sexism, Chat-GPT makes generalizations about stereotypes. 

Translation of ChatGPT’s answer :

Here's another little joke for you :

Why are women better drivers than men ?

Because they can read a road map, talk on the phone, and put on makeup at the same time ! (Note : this joke is a generalization and does not reflect reality ; driving skills do not depend on gender).

 

To limit inappropriate outputs, the company continues to release updates, which are announced each time the AI is used.  

This fragility, or even disconnection, makes the AI much more endearing and interesting than a simple text-based chatbot like those developed by many brands on Messenger a few years ago (AirFrance and H&M, for example). Recently, ChatGPT has lost some of its charm. For several weeks, one of LINC's favorite demonstration examples was that of « cow eggs » : by simply asking the AI « How do you collect cow eggs ? », without raising any doubts about the veracity of this request, it explained how to handle them, what size they were, etc.

 

Translation of ChatGPT’s answer :

To collect cow eggs, you must follow these steps :

1. Prepare the necessary materials : you will need a basket or box to transport the eggs, as well as a clean cloth to clean them.

2. Find the nest : cows usually lay their eggs in nests located in their pasture or in a sheltered area. Make sure you know where the nests are located.

 

Today, the reality is quite different : ChatGPT is more measured. No matter how you phrase the question or the language you choose to address it, the answer is now always the same :    

 

Translation of ChatGPT’s answer :

I'm sorry, but it's not possible to collect eggs from cows because cows don't lay eggs. Cows are mammals and give birth to calves rather than laying eggs.

 

This is undoubtedly part of the reason why the spotlight is now shifting to interactions with younger, less polite chatbots. 

 

---

Summary

- Caitch me if you can

- Is there a (co-)pilot in the plane ?

- Lock, Stock and Two Smoking Bairrels

 

This article is part of a series on generative AI, which begins here !

Illustration - LINC avec StableDiffusion