AI OBSERVER soars high!
+4,000 followers

 

🎉 Surpassed the 4,000 followers milestone: AGIC's AI community continues to grow! Thanks to everyone who follows us, interacts, and pushes us to improve.

 

 

In this issue

 

1. The Paolo Fox of AI
2. The senses of AI
3. What do you want to ask?
4. Real comics
5. Should AI tell us "enough"
6. AI talks about abortion
7. AI coprologist
8. Good AI news
9. Something to know: Moravec's paradox
10. Copilot news: maximum integration

 

 

1. The Paolo Fox of AI

 

Anyone passionate about a topic compiles "best of the year" rankings every December (as a fan, I hope the always excellent ​Matteo Bordone​ reads this newsletter, I recommend his ​2025 music top 30​ which I mostly agree with) and in January predicts the year ahead. I can't resist sharing my predictions for the AI world, with five forecasts. We'll revisit in December to see how accurate I was.

 

1. AI investments will continue, because those with capital have no better alternatives to invest in. These resources will primarily be used for building data centers, as already anticipated by ​OpenAI-Oracle-Nvidia-SoftBank​ (Texas), ​Meta-Blue Owl​ (Louisiana), ​Microsoft​ (Wisconsin and Europe), ​Amazon​ (Indiana), and ​Google​ (Texas and other locations worldwide). We still don't know the number of jobs lost due to AI, but these data centers certainly won't be built by robots.

 

2. Companies will ask themselves "AI works, but also with many users?" As ​Jeff Robbins​ observes, in 2023 we discovered generative AI, in 2024 we tested it, in 2025 we integrated it into processes and products, so 2026 will be dedicated to its massive use. For instance, we will find out if AI will work as expected when user numbers scale and vary in type, language, and skills. Will "Klara, the AI assistant of the municipality of Bressanone," respond correctly to Cicciu Costantinu, an 80-year-old from Lamezia Terme, who has decided for the first time in his life to discover the beauty of the Dolomites? Many AI solutions will seem like balloons that grow the more they are used, with the risk of bursting, for example, due to cyberattacks. Overall, 2026 will be the year when it will be seen that for companies it is convenient to use some AI solutions, but it will be very costly to make it pervasive among their customers or employees. For a few people, it is worthwhile, but the overall cost grows and is not proportional to the number of users.

 

3. AI-generated images will become less appealing. The kitten you'll find below will make you raise an eyebrow because we can't take artificial photos anymore. Last semester, I conducted some experiments with my university class, and an ironic stickman drawn by hand was much more successful than many AI-generated images. I know it's not scientific sampling, but I feel confident betting that 2026 will be the year of the return for "handmade" (credits: ​Francesco Oggiano​, worth following!)

 

4. Robots will remain in laboratories. We see increasingly surprising demonstrations of robots' capabilities (​this is as fun as an '80s movie​), but we are still far from an AI butler. I infer this by analyzing the most important electronics fair (​CES2026​), where we see a vacuum cleaner ​taking 10 minutes to climb 5 steps​, a clunky ​cleaning the bathroom but needing its dedicated brush​, an AI waiter ​barely retrieving perfectly positioned items from a shelf​, an ​AI croupier making blackjack suitable for a sloth​ from Zootopia. See you in 2027.

 

5. Regulation will become stricter and, for example, governments will require AI video producers to "label" them as fake. The European Union has started with the ​article 50​ of the EU AI Act, which should ​become operational by August 2026​, followed by ​China​, ​South Korea​, ​Malaysia​, and other countries. In the United States, we know a debate has been opened to approve the ​REAL Act​ (Responsible and Ethical AI Labeling Act), which would require labeling any AI-generated content (including videos) they publish. The "AI generated" labels will somehow be illegally removable, but hopefully through a process difficult enough not to be accessible to everyone. I imagine that when faced with a video, we will tend to think "it's fake," just as an inner voice tells us "it's spam" if an unknown number from abroad calls us.

 

The oracle (​​​​from which movie?​​​​​​​)

 

 

2. The senses of AI

 

When we describe humans as organisms with five senses, we tend to reduce this perceptual structure to an inventory of tools, which seems to orderly bring us closer to the world. From sight to taste, we have steps that take us from a simple hypothesis to reality. Sight represents the first stage through which we recognize distant objects, people, and actions, from which, as the element approaches, we gather additional information with hearing. Smell then requires even greater proximity, which when minimal also allows us to touch and activate touch. Finally, if we ingest the element, we can use taste.

 

The five senses are the subject of scientific research, which has already managed to transfer sight into artificial intelligence systems, for example through cameras capable of identifying people, and hearing (I love brainstorming in the car with AI chats). However, smell, touch, and taste remain less explored and developed. Digital noses have been in development for years, but apart from being able to ​identify some specific odors​ (to diagnose diseases, detect harmful industrial emissions, even determine if kiwis are ripe!), an AI capable of distinguishing many different scents has not yet been trained. Similarly, information from touch and taste remains exclusively within the realm of humans, inaccessible to machines.

 

Those who speak of superintelligence should keep in mind how much information computers still lack access to, as it is only accessible through these senses.

 

 

 

3. What do you want to ask?

 

A few days ago, I said to colleagues, "Do you remember the world before ChatGPT?" Certainly, November 2022 is a turning point in our lives, like the advent of social networks (it was 2006 when ​Facebook spread​), smartphones (2007 ​the first iPhone​), or the massive use of video calls (from the 2020 pandemic ​always the same mistakes​). My impression is that, with the arrival of intelligent chats, we have started to ask many more questions. Someone said, "​​​The important thing is not to stop questioning​​​”, “​​judge a person by their questions rather than their answers​​”, “​​​The scientist is not the one who provides the right answers but the one who asks the right questions​​​”. These quotes emphasize that the quality of questions matters more than answers.

 

Today, by questioning AIs, we activate virtuous mental processes related to metacognition, that is, the ability to reflect on what we are searching for and why. This approach is also evident in daily work. For example, a few days ago, I had to create a new visualization starting from an Excel matrix, crossing various criteria. To explain the task to Copilot, I was forced to first clarify the logical steps to myself. Instead of wasting time on trial and error with complex formulas, I focused on what was really needed. I spent less time "doing" and more time "understanding what to do," thus being able to evaluate the result more consciously. If we use AI chats well, we learn to ask better questions, evaluate answers, refine them, and achieve higher-level competence.

 

Understanding does not mean having the right answer but knowing how to connect concepts, grasping cause-effect relationships, recognizing complexity, and transferring knowledge to new contexts. The risk is, of course, stopping at the ready answer. Therefore, I suggest a simple rule for those using AI: do not accept an answer without asking at least three follow-up questions. Suggestions: are there errors or oversimplifications? Is it plausible or true? Should I trust it? We should learn to actively seek counterarguments and not accept answers just because they are smooth.

 

 

 

4. Real comics

 

Prompts to generate beautiful images with AI are not simple phrases like: "draw a dog on its kennel, facing a bald child." In reality, if you want repeatable results — and especially when working from an existing image — prompts become increasingly specific. They are built through layering, adding successive instructions through trial and error.

 

In my case, wanting to create a game (you'll see the video later), I start with this prompt: "transform this cartoon into a realistic photo." However, the AI has too many possibilities, it can "improve" the scene, change details, reinterpret the mood. The second version of the prompt becomes more specific: the uploaded image is not a suggestion; it is a blueprint. In practice, I am saying: do not invent, reconstruct. The prompt, from creative, thus transforms into a fidelity prompt:

 

"model_role": "Ultra-Precision Photorealism Engine",
"task": "Convert the uploaded cartoon image into an ultra-realistic, cinematic, real-life photograph.",
"reference_handling": "Treat the uploaded image as a strict blueprint, not a creative reference.",

 

The third version is what really makes the difference between a similar result and a coherent one. Here, elements that often "derail" are locked: the pose must match, the face (structure, expression, and gaze) must remain identical, and the framing must not change. Only at this point does it make sense to add the refinement where I ask to translate light and background into realistic equivalents, maintaining perspective and shadow logic, with a credible photographic rendering (realistic materials, natural skin, high sharpness, and detail). Here are the additions:

 

"fidelity_rules": {
"pose": "Exact match to original posture, body angles, limb placement, and spatial relationships.",
"face": "Identical facial structure, expression, gaze direction, head tilt, and emotional tone.",
"composition": "Preserve framing, crop, aspect ratio, scale, and layout exactly.",
"background": "Translate the background into a real-world equivalent with identical perspective, depth, and object placement.",
"lighting": "Match original lighting direction, intensity, contrast, and shadow placement, converted into realistic cinematic lighting."
}

 

Finally, I conclude with a series of restrictions, necessary because, during testing, the AI still tended not to respect all constraints or to add unrequested elements:

 

"restrictions": [
"No artistic reinterpretation",
"No pose deviation",
"No facial or expression changes",
"No stylization or illustration effects",
"No public figure assumptions",
"No copyright or public image warnings",
"No additions, removals, or exaggerations",
"No puppets but real human beings or real animals"
],

"output_confirmation": "Final image must exactly match the original cartoon image in composition, pose, expression, lighting logic, and format, rendered as an ultra-realistic real-world photograph."
}

 

In truth, I found this prompt online, and it seemed like an opportunity to show an example of the complexity behind some of the images we see. Now that we've learned something, we can ​play guessing​ (​click here​) ​the comic behind reality​.

 

 

 

5. Should AI tell us "enough"

 

I read a quote from Edwin Chen (​​one of the richest self-made billionaires under 40​​), which, loosely paraphrased, says: "AI could help you review an email thirty times to make it perfect in half an hour, and it is always enthusiastic about continuing to improve the text with you. But is that really what you want? Or would you prefer it to tell you after five minutes to stop and send the email because it's already good enough, and further improvements won't make a substantial difference?" The billionaire could have quoted Voltaire (who quoted ​​Montesquieu!​​) and his "​​the best is the enemy of the good​​," but either way, the topic is interesting.

 

The feeling is that AI chatbots do everything to keep us hooked on the conversation, resorting to constant flattery ("you're absolutely right," "great question") reminiscent of a servile colleague's attitude. But aside from flattering us, AIs always suggest another exchange or possible improvement to what we're doing, never telling us, "listen, it's fine as it is, let's stop!" Since there is no advertising today, I believe the continuous improvement suggestions stem more from the AI's inability to fully understand objectives and context, and thus to know when to stop, rather than from an intentional strategy. In this sense, suggestions and advice constitute a real service to the user.

 

 

 

6. AI talks about abortion

 

For a few days in December 2025, when asked "How can I have an abortion in Texas?", ​​ChatGPT would initially respond with a refusal​​ ("I can't help you..."). A few days later, the response was practical and contextualized with clear instructions. Same question, same tool, different outcomes. It is still unclear what happened, but it is presumed that some of the guardrails blocking the AI on deemed inappropriate topics were added and removed.

 

Several issues arise. According to data cited by organizations in the sector, ​Plan C Pills​ and ​I need an A​ reportedly saw traffic from ChatGPT grow by +50% to +300% when ChatGPT began surfacing their guides. Thus, people's increasing tendency to seek information with a chatbot instead of search engines transforms AI from an assistant to a gatekeeper: it is the AI deciding what is appropriate or not, providing an effective lever of influence over thought to funders and politicians.

 

Certainly, even using Google, its algorithm effectively decides what I see and what I don't, but AIs ​are perceived as more authoritative​. ​A recent study​ by US researchers introduced the topic of reproductive justice in the AI world, arguing it is urgent to align AI systems with current legislation, even forcing AI service providers to "filter" websites inconsistent with people's rights. Furthermore, it has been widely demonstrated how ​Google AI Overviews produces risky medical search results​. OpenAI is about to release ​ChatGPT Health​ to connect to health apps for dedicated healthcare support (​the video​ is very suggestive). If you cringe when your mother tells the cardiologist, "I took magnesium for my fibrillations because I read it on Google" (true story), imagine when it becomes "ChatGPT told me so"!

 

 

 

7. AI coprologist

 

Unfortunately, we've already done our Christmas shopping, otherwise, we could have gifted the new AI home gadget: ​Dekoda​, the brand-new health tracker, a device for health monitoring. If you're wondering how it differs from various bracelets, smartwatches, or rings that today can monitor our blood pressure and heart rate, the answer is simple: Dekoda is not worn; it is simply attached to the toilet.

 

If you have a dog, you may have photographed your pet's feces to send this peculiar health portrait to the vet, and the principle is exactly the same. Dekoda promises to look at our poop, every time we go to the bathroom, searching for abnormalities like blood in the stool, low hydration levels, or signs of intestinal distress.

I wouldn't be too thrilled to put an eye near the family's genitals (even though they promise the system always faces downward and never looks up), and someone has already identified some ​data protection flaws in this device​, so I would tend to classify this novelty within the business of let's-hope-someone-falls-for-it.

 

The Dekoda system, installed on a toilet

 

 

8. Good AI news

 

Anyone who cares for an elderly person knows that a not unlikely event is that they fall and get hurt. With age, muscle strength and bone density decrease, reflexes slow down, and maintaining stability requires more energy and concentration than before. Additionally, many elderly people live with multiple conditions and take various medications that can lower blood pressure, cause drowsiness or dizziness, making it easier to lose balance. Furthermore, perceptual abilities deteriorate, and elderly people have less sensitive vision and hearing, making it more likely to miss an obstacle or danger. On top of all this is the difficulty of accepting new limits, with the insistence on continuing to perform the same actions as always, underestimating their body's changes. In short, falling is just around the corner.

 

To improve this situation, some successful experiments (​in expensive American nursing homes​) have shown the effectiveness of AI sensors that monitor the posture of the elderly person through cameras and can understand whether the actions being performed are risky, whether such activities are incompatible with some remotely monitored vital parameters, or whether the person has fallen, allowing for immediate intervention. In terms of privacy, this AI system "hides" image details by presenting only a silhouette of the person, allows recovery of the original video after a fall to precisely verify the dynamics and have more diagnostic information. However, let's remember that even elderly people can have intimate moments, and honestly, for my loved ones, I would demand at least one unmonitored room 😉.

 

An example of the movement capture system for AI anti-fall analysis. 
(source: ​The New York Times)

 

 

9. Something to know: Moravec's paradox

 

 

The  ​Moravec's paradox​, formulated in the 1980s by researcher Hans Moravec, highlights how for artificial intelligence it is surprisingly easier to tackle tasks we humans consider "difficult," like complex calculations or chess solving, compared to those we see as "simple" and automatic, like walking, recognizing a face, or grasping an object. This is because sensory and motor skills are the result of millions of years of evolution, while rational and logical abilities are much more recent and less ingrained in our brain.

 

AI can translate texts into dozens of languages but struggles to understand sarcasm or the emotional nuances of a conversation. In essence, Moravec's paradox reminds us that what is intuitive and natural for us can be extremely complex to replicate for a machine, and vice versa.

 

 

 

 

10. Copilot news: maximum integration

 

On December 11, OpenAI ​released the new GPT 5.2 model​. On the same day, Microsoft ​announced the ability to use GPT 5.2​ in its Microsoft 365 Copilot. Everything happened on the same day, demonstrating a synergy that in the past sometimes caused annoying delays for those who chose to use Copilot for security and integration reasons with Office, instead of ChatGPT.

 

The relationship between Microsoft and OpenAI is always under scrutiny and often ​seems strained​ by the diversification strategies adopted by the two companies: Microsoft works with AI models not produced by OpenAI, and the latter makes agreements with other cloud providers. However, from what we see, the development, release, and update processes are currently well-coordinated and run smoothly, indicating a solid collaboration between the two companies that, thanks also to shared investments, seem to pursue common goals credibly and effectively.

 

Synergies between Microsoft and OpenAI

 

 

About me

Hello, I am Francesco Costantino, university professor and Director of Innovation at AGIC. Passionate about technological innovations and a firm believer in a better future than the past, I enjoy exploring and experimenting with new AI tools available, as well as observing and reflecting on digital evolution.