Most of us treat hallucinations as an evil, scary thing. With some exceptions.
This negative perception of hallucinations extends to our views of generative artificial intelligence. Although perhaps what generative AI does is more accurately called “confabulations.”
““A hallucination is a conscious sensory perception that is at variance with the stimuli in the environment. A confabulation, on the other hand, is the making of assertions that are at variance with the facts, such as “the president of France is Francois Mitterrand,” which is currently not the case.”
Whatever you call it, the result is not consciously intended. And it can sometimes be bad.
Lying on a job application
Take those AI tools that jobseekers can use to not only apply for a job, but automatically customize their resume for that particular job.
When automatic resume rewrites are not reviewed, the new resume may end up with confabulations, hallucinations, or outright falsehoods.
If my rewritten resume claims two years’ Python experience, that just ain’t true.
And I could lose a job opportunity if I lie on my resume.
Fly (on) an eagle
But those who praise hallucinations as good are not limited to Timothy Leary.
Take the time I intentionally asked Google Gemini’s image creation engine (Imagen 4 at the time) to make something up.
Perhaps I’m wrong, but I don’t see any harm in creating a Tolkienesque illustration of Theodore Roosevelt riding a flying bald eagle. Actually, TR fans may think it’s pretty cool.
By definition, ANY generative AI engine HAS to invent stuff. A prompt can’t specify everything.
Audio inventiveness
Let’s look at another example, the two-plus minute song that formed the audio for my recent reel “The Cooling Blue.”
Now here is the prompt that I used to create that audio track.
“Create a moving song with violin, harp, and guitar about overly long meetings. The opening male spoken words are “meeting hour 1, meeting hour 2, meeting hour 3, meeting hour 4.” The female singer, accompanied by a female choir, sings of her despair in pointless meetings with no purpose. The chorus consists of the choir singing “When will this madness end?””
When you review the prompt you can see many of the elements of the final song.
But I never told Lyria to sing “the coffee turned to ink.” Lyria made that up.
But I like that addition.
And I have another example.
Image inventiveness
This example is from the images that appeared throughout the video. These were also created by Google; is the image generation capability still called Nano Banana this month?
Anyway, here is the prompt for the noon scene.
“Edit the picture so the time is noon and the lead wombat is still droning on and on. The attendees are restless.”
Google executed my image request.
But look more closely.
I did NOT specify that the koala write the note “Make it end…so sleepy.” Or any of the other notes that this particular koala wrote throughout the day.
Nor did I specify the “out of order” note that appeared on the coffee urn at 10:10 am.
(My little secret: that time was NOT supposed to be 10:10. I asked Google to display a time of 10:45. But since so much of the clock training data uses at 10:10 time, Google got confused.)
But I like those additions.
Take two minutes and twenty-four seconds and watch the reel again, taking note of the few elements specified by me, and the many elements that were “made up” by Google.
Hallucinations can be good, evil, or indifferent
Adding a koala note of frustration is a good thing.
Lying on a job application is a bad thing.
And showing a time of 10:10 instead of the requested time of 10:45? It didn’t materially affect my story, so I was indifferent to it.
Bredemarket 