AI generated content

While advanced analytics techniques enable personalization on a scale, machine learning and decision support systems aid professionals every day. While all over the net, same tech is used to hunt for clicks, views and engagement.

As I read more and more content on the internet, started to develop a “sixth sense” if an article is AI or not. If an image is AI or not. However this is unreliable, thinking, can we distill to a few signs that work well for current state?

Maybe. Time will tell how fast even these signs will disappear.

• Reply or the early paragraphs start with the repetition of the main statement: due to inner workings as a statistical engines, genAI puts word by word (tokens, actually), it is reliant on this start to increase precision of answer.
• Adjectives, verbs often used in marketing materials, rarely in live conversations: this one is just a yellow flag, though makes sense that published materials overcome not-published speech. E.g. elevate, unleash, unlock, inspiring, etc. if an author would tell you his protagonist’s traits instead of showing.
• Oversimplified advice, bullet points, high level concepts: while it is not unique to AI, as it does not have general intelligence, putting a concept to daily usage, examples, is still human quality, I always look for it and my biggest turndown reading article if it only speaks in general terms.
• Color schemes of images, weird body parts and text messed up: the palette seems to have a distinct feel, it is a more abstract, vibrant yet artificial than ever before.
• Very short clips, background versus foreground quality differences, details lost: additional to images, and even with the advancements of CGI, however this is a fast developing field too.

I am lucky to be old enough, having read dozens if not hundreds of books written by humans, verifyably. Seen hundreds of paintings, drawings in person or internet. Seen movies in cinema for decades. So I have experience distinguishing the new wave from human styles. What about younger generations?

We will face a new challenge raising them up. Teaching the difference, especially when we ourselves struggle to make a difference, will be hard.

How do you distinguish AI content if it is not tagged?

P.S.: I thought about voice imprints… and have no clue how to distinguish generated versus real without the source telling me, anyone in my network an idea on that one?

Voice and song generation was available even earlier, like Vocaloid from 2007, so makes sense that we have the most experience in this.