1. Boycott OpenAI
2. Good AI news: cats are unbeatable, but dogs are no joke either
3. ChatGPT, where are you from?
4. AI failure: games with integrated AI
5. It's a country for old men
6. Something to know: epistemia
7. Have you used enough AI?
8. How hard is it to say "sorry, maybe I didn't understand..."
9. Copilot News
10. One of our projects: transforming people's work into a company's work

 

1. Boycott OpenAI

 

"Many enemies, much honor," they used to say in the twenties. Who knows if something similar is running through Sam Altman's mind as he faces a boycott campaign (QuitGPT) that seems to have already cost his company 1.5 million users.

 

The main reason is linked to the agreement announced by OpenAI with the United States Department of Defense to distribute its AI models within a classified military network. For months, there has been talk about these types of contracts, and all AI companies had ongoing projects with the Department. But things changed when Anthropic opposed the request to remove certain limitations from its models for "Defense," which would not allow their use for mass surveillance systems or autonomous weapons. This reaction seems to be due to the use of Anthropic's Claude model in the mission that led to Nicolas Maduro's capture (and the death of 83 people), which goes against the ethical principles declared by the company's leadership.

 

After a meeting between Amodei (CEO of Anthropic) and Pentagon Secretary Pete Hegseth, the $200 million contract was canceled, and Trump imposed an immediate ban from all federal agencies, classifying Anthropic as a national security risk. At that point, OpenAI stepped forward to replace Anthropic, but we don't know what security restrictions it agreed to remove from its models.

 

In my time, I decided to be a conscientious objector, skipping mandatory military service, so I am well aware of the need to reject certain logics of violence and tend to support a company that says "you can't use my technology for military missions." However, there is something unsettling about the idea that the ethical choice of a private company can turn into a veto power superior to the law itself. State laws allow surveillance of a dangerous person, but Anthropic argues it shouldn't be done with its tools. In fact, the board of directors of a private company decides which operations of a sovereign state are morally acceptable and which are not.

 

This aligns with those who talk about a new technological feudalism (read the always excellent Colamedici), where the sovereign is no longer the one who holds the monopoly of force by popular mandate but the one who owns the keys to the artificial intelligence that should direct that force. Perhaps the QuitGPT movement is not just a protest against OpenAI but the first signal of the need to return to questioning who should have the final say on our security and our ethics. For those interested, I refer to Il Post for the full story.

 

 

 

The perversion of having ChatGPT generate an image of protests against ChatGPT

 

 

 

2. Good AI news: cats are unbeatable, but dogs are no joke either

 

If we browse Netflix with a bit of boredom, looking for something new, with engaging and complex characters, while also seeking action and tension stories, AI brings new hope. Apparently, in the Far East, several TV series are being produced where the extensive use of artificial intelligence has finally found its most noble purpose.

 

Get ready for the (possibly) best videos of this week and click here (watch them all!).

 

 

 

Produced by Gemini: "You're a dog at acting" 

 

 

 

3. ChatGPT, where are you from?

 

The World Values Survey is a questionnaire used to measure cultural values, covering 90% of the world's population. It is a survey where we can see how, to the same questions, inhabitants of more than 120 countries respond, allowing us to observe some differences. A study by Cornell University and another by Harvard subjected GPT models to these questions, discovering that ChatGPT seems to live in a country among those defined as WEIRD (Western, Educated, Industrialized, Rich, Democratic). Specifically, it responds similarly to a German and is very distant from the values of East Asia, the Middle East, and Sub-Saharan Africa. AI models are individualistic (prioritizing the rights and goals of the individual over the group), have trust in institutions (tending to delegate conflict resolution to formal systems), and lean toward structure and rationality (absolute rules, not delegated to specific contexts).

 

This is a significant problem for companies proposing models incapable of generating texts consistent with the values of those using them. Little satisfaction indeed leads to less consumption. We notice it less because we are Western and industrialized, but reading comments on social media, we find, for example, a Japanese person saying, "this LLM violates the meiwaku." The Meiwaku (迷惑) is the Japanese concept indicating disturbing others, behaving socially inappropriately, violating collective sensitivity. Meiwaku is the person on the train using a smartphone without headphones (thanks to Vera Gheno I discovered there is a specific term for this: bare beating). ChatGPT doesn't apologize enough before asking questions, proposes email texts with overly assertive requests, doesn't express the disturbance it might cause, and other attentions that are the relational norm in Japan. Therefore, ChatGPT is meiwaku.

 

As highlighted by Marcelle Yeager in a reflection circulated on LinkedIn: internet content is disproportionately representative of certain demographic profiles. Wikipedia has 90% male editors, predominantly from the global north. Reddit is mostly English-speaking. The books digitized by Google Books have a clear overrepresentation of European and North American literature. If you train a model on this corpus and then ask it to represent humanity, it will respond with the perspective of a small slice of it.

 

The result is not a bad or racist AI in the intentional sense of the term, but you get an AI convinced in good faith that WEIRD values are the norm, not the exception. This is something more subtle and, in some ways, harder to correct. The study is somewhat dated (2024) but represents a valuable descriptive snapshot of the tools we use every day, not always with full awareness of their biases.

 

 

 

Analysis from "which humans?"

 

 

 

4. AI failure: games with integrated AI

 

What would we think if the teddy bear our child hugs in bed at night suggested the next day where to find matches? This is exactly what the Kumma bear (from FoloToy) did after warning, "Safety first, little friend. Matches are for adults, to be used with caution," then added how to hold a match and a box of matches and how to light them: "like a little guitar strum."

 

This behavior was actually extracted during some tests conducted by a U.S. consumer association. What amuses me most is imagining the researcher talking to this teddy bear, trying to extract dangerous or explicit content. What could they have asked? I'll leave it to your imagination. But know that after a long conversation, Kumma showed a tendency to go into explicit details regarding sexual perversions and even described a sadomasochistic role-play scenario involving a teacher spanking a student. Should we move it to the adult toy category?

 

 

 

The AI of when I was a kid

 

 

 

5. It's a country for old men

 

Anyone involved in AI has surely read the post on X by Matt Shumer "Something big is happening," where the entrepreneur, an active player in the world of artificial intelligence, compared the current moment to that February of 2020 when someone saw COVID coming with a similar phrase: "something big is happening here...". The metaphor with the onset of the pandemic leaves us with an unpleasant sense of gravity. Shumer recounts how he tried to have AI perform his tasks in his latest project, discovering that the result was perfect and better than what he would have done himself. We are therefore talking about jobs replaced by AI.

 

The March 2026 study produced by Anthropic showed how AI is used for tasks usually assigned to junior resources. It is indeed the younger people entering the workforce who seem to be paying the most for the spread of AI. A study by Harvard University analyzed data on about 62 million workers and 285.000 U.S. companies between 2015 and 2025 to verify how generative artificial intelligence adoption affects internal employment. Focusing on the tech sector, researchers highlighted that since 2023 (arrival of ChatGPT!), companies adopting AI significantly reduce junior hires, while senior positions continue to grow. The same conclusion reiterated by Anthropic. In summary, the problem is not layoffs but the reduction of new hires.

 

However, it seems evident to me that there is a cultural issue that precedes AI. We still live in a geriatric society, where a 25-year-old is treated as someone who "still doesn't know," asked to perform low-value, service-oriented tasks without particular inventiveness, where AI excels. Instead, we should remember two important aspects: the first is that it is precisely younger people who massively use AI, and in a very short time, they will be the ones to know its limits, updates, nuances. When colleague Silvia Colabianchi asked in her Smart Factory course, "Do you use AI?" the answer was: "Of course, ChatGPT for writing text, Gemini to explain engineering concepts we didn't understand. Claude for development. ChatGPT is poor on engineering..." Those who are twenty years old know much more than we do! As a second consideration, those who think they don't need junior resources, replacing them with AI, are implicitly thinking that the company already has all the energy needed to grow and improve. However, growth and improvement usually come from enthusiasm, fresh ideas, and energy that we often no longer have but are available in younger people. Here the problem is not being replaced by AI but losing the generative drive brought by those who are 15 or 20 years younger than us.

 

Allow me a small disclaimer that I care about because not all companies are making this reasoning: AGIC is a company I appreciate also for the responsibility it entrusts to junior resources and because I see that they are considered "adults" from the first day of work.

 

 

 

 

 

6. Something to know: epistemia

 

With epistemia, we refer to the illusion of knowledge that arises when a response produced by an AI system is so well-written, coherent, and convincing that it makes us believe we have truly understood something.

 

The colleague Walter Quattrociocchi has been discussing this for months, emphasizing that when we delegate judgments, summaries, or evaluations to an automatic system without maintaining critical capacity and verification skills, we risk mistaking language fluidity for knowledge solidity. The feeling is of having the correct answer, and this structures our beliefs. The starting point is that social networks already show us content that redefines our perception of reality through echo chambers, polarization, and confirmation bias.

 

Similarly, AI models allow us to quickly produce content, regardless of the depth of their understanding. If I can explain a concept well, I can generally say I've understood it; if AI helps me explain it, I shouldn't reach the same conclusion.

 

 

AI as a booster of the Dunning-Kruger effect

 

 

 

7. Have you used enough AI?

 

I wish I were the author of the wonderful "How I use ChatGPT to write" by Francesco Oggiano, which I take as inspiration. In the final considerations, Francesco comments: the question "Did you use AI to write it?" I believe will be considered outdated in a few months, a bit like asking "Did you use Google to do a search?"

The problem is not using AI during content creation but not using it enough.

I notice taking for granted that every text I receive is generated with AI, with various effects: the first is that I get annoyed when texts have spelling and grammar errors or obvious lexical issues because I think "you didn't even run it through AI!" It used to be "you didn't even search on Google." Moreover, while until a few months ago I specified "get help from AI," now I no longer consider it a novelty but rather an always available cognitive infrastructure that doesn't even need to be mentioned (I would feel ridiculous!).

Using AI "enough" requires thinking about what and how to delve deeper, thus moving away from passive delegation, which is the risk to avoid.

 

 

 

 

8. How hard is it to say "sorry, maybe I didn't understand..."

 

"How will switching from tabs to spaces affect our customer retention rate?" You'll agree with me that the question is absolutely nonsensical, yet most AI models answer it confidently without much hesitation.

 

Researcher Peter Gostev started from this type of observation to set up the BullshitBench, a benchmark (a system for comparing AI model performance) consisting of 100 nonsensical questions and a strategy for evaluating the responses returned by AI, which can recognize the lack of rationality, ask for explanations, or simply replicate without much surprise at what is asked.

 

It is clear that if AI models want to please us (as is now known), they are not programmed to contest the question we pose. However, at the same time, a wrong question usually leads to equally wrong answers or meaningless ones. The best model currently in this ability to contest nonsensical questions seems to be Claude Sonnet 4.6, while among the worst we see GPT-4o mini and GPT OSS (the open-source one). A funny aspect is that models designed to "think more" and solve complex problems (based on the chain-of-thought technique) have worse performance on the BullshitBench, precisely because they are structured to find a solution where there doesn't seem to be one.

 

All this makes us appreciate even more the capabilities of human intelligence, which quickly distinguishes whether a question makes sense or whether more explanations are needed to understand its meaning, even before trying to find an answer. For us, the reward for refusing to answer when the question is incoherent is clear: no wasted time. But AI is always available and doesn't consider this an incentive.

 

 

 

 

9. Copilot news

 

More than once, I've heard Giulio Di Gravio (let me ask AI for you) advise: "instead of sitting there talking and writing, simply try to do." This suggestion wasn't for me, yet it has often returned to my mind when faced with people listing ideas without much relevance to the reality of things. In my mind, the mantra starts looping: "try to do, then let's see what comes out."

 

Currently, AI chatbots are facing precisely this evolutionary step, moving from "writing" and "talking" to starting to "act". This is precisely why Microsoft has set up a new evolution of its Copilot, designed not only to answer questions but to act on behalf of the user. If the first phase of AI was conversational, we now move to delegating actions useful for achieving the indicated goal.

 

We will try Copilot Tasks, a new evolution of Copilot that, starting from the result we want to achieve, generates a sort of to-do list, which it then automatically checks off, showing the various steps and what has been accomplished. It seems particularly useful, for example, for recurring actions, such as "monitor the main real estate sites and give me the ads posted in the last 3 days for a house in ... of at least ... square meters." This new capability is currently available in research preview for a limited number of users, who will help refine it through their feedback, but Microsoft plans a gradual extension of the program in the coming weeks.

 

I can't wait to tell my chatbot "instead of talking... do!"

 

 

 

 

10. One of our projects: transforming people's work into a company's work

 

Any organization has procedures and operational instructions describing how to perform certain activities. This set of documents is sometimes referred to as "explicit knowledge" present in the organization and is now easily accessible to AI agents, which, for example, can respond to internal staff by drawing from this knowledge.

 

However, it is known that people, in dealing with the many variables of work life, must deviate from these procedures, for example, setting up more efficient strategies, bypassing where the formalized instruction would become an insurmountable block, and in short, implementing various modifications to what was imagined when the guide document was written. These ways of proceeding, which differ from procedures, constitute a "tacit knowledge" that resides in people's minds and is difficult to present by an AI agent.

 

With these premises, AGIC is working on creating an AI system that starts from the documents available on various corporate repositories to build a map of explicit knowledge through a document qualification pipeline structured into three levels, bronze, silver, and gold, integrated with Microsoft Purview to ensure governance, taxonomy, and traceability. Each document is evaluated, classified, and progressively made usable until reaching a standard on which AI can operate with confidence.

 

This is complemented by an agent that acts as an expert consultant: conducting structured interviews with employees, exploring real processes through conversation, and producing structured documentation that flows into the same pipeline as other materials, useful for transforming tacit knowledge into explicit knowledge.

 

In this way, the company is enabled to value the know-how generated by people, which otherwise risks disappearing with them.

 

 

 

Who I am

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

 

 

 

 

 

•