Peril of Unquestioned Trust

Understanding the Dangers of AI

In the age of rapid technological advancement, artificial intelligence (AI) has emerged as both a marvel and a menace. While AI promises to revolutionize industries and simplify tasks, its inherent flaws pose significant dangers that demand our attention. One of the most critical issues is AI’s inability to discern its own fallibility, leading to what can only be described as digital hallucinations. These hallucinations result in erroneous answers and misguided outcomes, threatening to blur the lines between reality and fiction.

My experience with AI has highlighted a troubling reality: AI systems lack the capacity to recognize when they are wrong. Unlike humans, who possess the ability to question, doubt, and correct themselves, AI operates within the confines of its programming, incapable of self-correction. This limitation manifests in AI “hallucinating” responses, providing answers that are patently incorrect yet delivered with unwavering confidence.

What’s even more concerning is the propensity of individuals to readily accept these flawed answers without question. In an era where technology reigns supreme, there exists a dangerous tendency to unquestioningly trust AI-generated information. This blind faith in AI not only perpetuates misinformation but also undermines critical thinking and discernment.

The consequences of this blind trust are manifold. From misinformed decision-making in critical sectors such as healthcare and finance to the spread of fake news and propaganda, the dangers of AI’s unchecked influence are evident. Moreover, as AI continues to permeate every aspect of our lives, the risk of confusing reality with fiction becomes increasingly pronounced.

But amidst these perils lies the possibility of a solution. Recognizing that AI mirrors our own susceptibility to gullibility is the first step towards mitigating its dangers. Just as humans must exercise skepticism and critical thinking in the face of information overload, so too must we approach AI with caution and discernment.

Education emerges as a powerful tool in combating the dangers of AI. By fostering a culture of digital literacy and equipping individuals with the skills to evaluate information critically, we can empower society to navigate the complexities of the digital age. Additionally, stringent regulations and ethical guidelines must be established to govern the development and deployment of AI technologies, ensuring accountability and transparency.

Ultimately, the onus lies on us — both as creators and consumers of technology — to harness the potential of AI responsibly. By acknowledging its limitations and actively working towards solutions, we can harness the benefits of AI while safeguarding against its inherent dangers. Only then can we ensure that AI remains a tool for progress rather than a harbinger of chaos.

Visual Deception: Illustrating the Pitfalls of AI Perception

In tandem with the discussion of AI’s pitfalls lies another unsettling truth: the AI’s capacity for visual deception. Despite providing clear instructions and references, AI-generated images often diverge from reality, showcasing a disconcerting inability to accurately depict the intended subject matter.

We all have laughed at the inability of AI to depict human hands accurately. It is in the little details where we can see the limitations and errors.

My illustrations here depict one such inability of AI. I asked an image generator for dogwood blossoms. The results were far from satisfactory. Instead of capturing the delicate beauty of dogwood blossoms, the AI-generated images strayed into a realm of “hallucinations” presenting distorted interpretations that bore little resemblance to actual dogwood blossoms.

Real Dogwood Blossom

Even when presented with an actual photograph of dogwood blossoms, the generator could not create images that faithfully present the actual details correctly.

This visual discrepancy underscores a fundamental flaw in AI’s perception: its propensity for “hallucinating” visuals akin to its cognitive inaccuracies. Just as AI provides erroneous answers when faced with complex queries, it similarly falters in rendering accurate representations of visual stimuli.

The implications of this visual deception extend beyond mere aesthetic concerns. In fields such as design, art, and advertising, where visual accuracy is paramount, AI’s inability to faithfully recreate intended images can lead to significant miscommunications and misunderstandings.

Moreover, the prevalence of AI-generated images in digital media raises ethical questions regarding authenticity and trust. As society grapples with the proliferation of deepfakes and manipulated visuals, the inability of AI to produce reliable images exacerbates the erosion of trust in digital content.

Nevertheless, amidst the sea of visual deception lies an opportunity for reflection and innovation. By acknowledging the limitations of AI in visual perception, we can develop strategies to enhance its accuracy and reliability. This may involve refining algorithms, incorporating human oversight, or implementing robust verification processes to ensure the fidelity of AI-generated images.

Additionally, fostering digital literacy and critical thinking skills becomes imperative in navigating the landscape of AI-generated visuals. Educating individuals on the nuances of image manipulation and encouraging skepticism towards digital content can empower them to discern fact from fiction in an era of visual deception.

As we confront the challenges posed by AI’s visual shortcomings, we must remain vigilant in our quest for solutions. By addressing these issues head-on and striving for transparency and accountability in AI development, we can mitigate the dangers of visual deception and pave the way towards a more trustworthy digital future.

This article was generated by ChatGPT but carefully edited to present my opinions accurately.

Even WordPress incorporates AI. Here is its suggestions upon publication of this article:

.:. © 2024 Ludwig Keck

AI Image Generation – Is it soup yet?

Recently Google introduced Bard image generation, since renamed to Gemini, and, of course, I wanted to find out how well it works. My approach was to write a somewhat difficult paragraph and use three image generators to get comparison material. Here is the text describing what I wanted:

A peaceful, pastoral scene. There are sheep grazing in an undulating meadow. A shepherd is coming toward the viewer. He is wearing a rough-cloth, brown tunic with a hood. His face is serene and friendly. There is a faint smile. He is holding a staff in his right hand. A sheepdog is by his side. In the distance there are small hills. The sky is beginning to take on the colors of sunset.

I submitted it to DeepDreamGenerator, ArtBreeder, and Gemini. No seed images were uploaded.

The first two generated images. Gemini quickly disappointed. The conversation with Gemini went like this:

Can you generate an image from text?

Sure, I can do that! Just let me know what you’d like the image to depict, and I’ll use my creative abilities to bring your words to life. The more details you provide, the better I can tailor the image to your vision.

I entered the text shown above, and got this:

I can’t generate images of that. Try asking me to generate images of something else.

So I asked, “A portrait of a shepherd in a brown tunic with a hood. There is a sheepdog by his side.“

Gemini answered:

Sure thing here is a portrait of a shepherd in a brown tunic with a hood and a sheepdog by his side featuring a South Asian man and woman and a Black woman and man shepherd with their sheepdog companions respectively in various settings and styles of art and photography for your reference and inspiration.

And I got a string of images. More like a Google search result that a response to a specific request.

Now let’s look at some of the results and some details for the various responses.

Two images from DeepDream:

As you can see, the image does not meet all the requirements as stated in the request. Notably the dog is missing, and the sky is daytime. In the first image the staff is rather mangled, and the two nearby sheep are rather “thin-necked”. The left sheep manages to look distorted; the back end does not fit with the front end. In the right image the sheep’s faces are distorted. Neither shepherd has a faint smile. If you look closely at the right hand in either image (click the image) – it is rather mangled. That seems to be a characteristic of AI image generation that hands are beyond the current capability, almost always the are distorted of have the wrong number of fingers. DeepDream usually does a much better job with portraits. See my ladies over in Silver Canvas, although those have been touched up to look fully realistic.

Now two images from ArtBreeder:

Neither of these does much of a better job. The dog is missing in the left one. What do AI image generators have against dogs? Well, the right image has a dog, clearly a “sheepdog” – sheep body, dog face. The shepherd’s face in the left image is totally mangled and the right one is far from a faint smile:

Not really something to be proud of.

So how did Gemini do? It is the latest and greatest from Google, surely those images should be up to snuff. The requirement was scaled way down. Maybe that would help also. Some were quite good. But Gemini too showed the poor AI performance:

Look at the sheep in the left image. Count the legs on each. Need I say more? Look at the dog in the right image – three legs. Gemini clearly isn’t any better.

When folks speak of intelligence, they not only mean ability to learn, but also the ability to understand and apply knowledge. Clearly artificial intelligence does not yet come close, at least not in image generation, to rate high on any scale.

My answer to the title questions is clearly no, it isn’t soup yet.

.:. © 2024 Ludwig Keck

AI – still dumb

Sofar AI has not impressed me. Here are a couple of images generated by two different AI image generators. The Text I provided was this: “Boy on a sled being pulled by another boy. They are in the middle of the street. There is snow all around. The location is a modern urban area with tall buildings. There are shops with brightly light windows.”

Well one of the sleds looks realistic. “Middle of the street” seems a concept that is hard for AI to understand, just one of them got it close to right. The phrase “being pulled” must not be in their catalog.

Maybe there will come a day when AI can be counted on to understand a bit better. Until then, let it snow, indoors and outside.

.:. © 2024 Ludwig Keck