June Newsletter

1. The Wonderful Humanity

As someone with a technical background, atheist, I never thought I would read an encyclical. And I never thought I would find it perfect: useful, stimulating, courageous, capable of guiding enlightened minds. Reading the document “Magnifica Humanitas” by Pope Leo XIV took me far from the cliché that sees deep faith as opposed to science. The encyclical is extremely interesting and is the first political document I feel represents me. Some excerpts: “Each generation inherits the task of shaping its own time”, “If technological development proceeds without adequate ethical and social maturation, it may happen that means increase without humanity growing proportionally: there is ‘more to have’ but not ‘more to be,’ and the person risks being valued primarily based on the performances they guarantee.”

The encyclical does not provide an evaluation of AI behavior but explicitly criticizes a situation of power concentrated in the hands of a small group of people. In fact, the Pope writes: “We do not need a more moral AI if this morality is decided by a few”. The position is curious because historically the Church allowed about 250 cardinals (all male, of course!) to decide what was moral or not. But we appreciate the step forward. Leo XIV explicitly speaks of exploiters and invisible labor, used unethically, but also of the need to overcome enthusiastic or fearful positions, taking responsibility for choosing between progress that serves the person and peoples, or progress that bends them to power dynamics.

An interesting analysis of the document was published by Andrea Colamedici, who highlights the Pope's choice to consider AI as “cultivated” and not “created” or “built.” The chosen term is interesting because one cultivates a plant. Instead, we know that models are “trained” and that they “learn.” The choice of terms is important. I train someone who must achieve superior performance. I learn through a very sophisticated cognitive process that until today has been the prerogative of humans and animals. I have never heard of a plant learning something. Now the Pope chooses the term “cultivate” to downgrade AI models to an element useful for our survival, like a tomato or grain. However – again Colamedici – the same term “cultivate” suggests that “the machine today has changed its nature and has moved to the side of what grows”, in a non-mechanical, only partially controllable manner.

The encyclical shows a Pope perfectly in tune with his time, who rejects the hypocrisies of economic convenience but also embraces the complexity of the current technological situation.

2. AI is not a Parrot

Let me preface this: I am about to self-promote my work. But when you achieve an important result, it’s easy to lose the compass of humility. Last month, an article written by Andrea Falegnami, Andrea Tomassi, Michele Levorato, and myself was published in the prestigious international journal “AI & Society”, where we argue a simple yet original concept: many people who insist on defining AI models as simple “parrots” that probabilistically produce text are wrong. These models are not solitary elements but enablers of an intelligent system made up of AI, individuals, data, interfaces, documents, processes, incentives, and much more. It is necessary to accept that in a work context (typically knowledge-based), it is more productive to collaborate with AI than with a colleague, thus modifying, improving, and making more useful the moments of interaction with that person.

The invitation is to not romanticize human skills in a world where many organizations struggle to reward those with more value (think, for example, of public administration), to realize that AI is effective in performing human activities only until we elevate people to be more human and less machine-like. If so far, I’ve asked someone to check data in a document by opening software and verifying its accuracy, I’ve used their potential poorly. The same goes for responding to support request emails with information already available online, and so on. Let’s abandon the discussion “AI is not like a human,” “it’s better than a human,” and similar, to embrace one focused on analyzing the system that contains AI and reasoning about what virtuous contribution a real flesh-and-blood person can make.

I admit with pride that we stood firm against the request to change the title, which the reviewers found too ironic: “Stochastic Parrot or Not, It’s Still Less Chicken Than the Colleague Two Desks Away.” The paper is available here, and contains many other useful considerations for placing AI in contexts made of people, systems, organizations, in a more conscious way.

Gemini: create an image with title stochastic parrot or not […]. Evidently, it hasn’t read the paper.

3. Trump Requests to Block Claude for Non-Americans

A few days ago, the U.S. government sent a directive to Anthropic ordering it to suspend the use of its most performant AI models for those who do not hold American citizenship, citing cybersecurity reasons. Anthropic blocked its Mythos 5 and Fable 5 models for its entire clientele, regardless of citizenship.

For several months now, Claude has been considered the most satisfactory AI model by many, and it cannot be excluded that Trump is trying to use its widespread adoption as a strategic lever, capable of putting entire countries in difficulty. It is unclear why an American citizen should be able to use these models if they are capable of generating cyberattacks. Honestly, I think this might be a move orchestrated by Anthropic itself, which could either be trying to reduce the costs generated by these huge models or shut them down quickly due to some security flaw it has discovered.

We already knew this, but AI is becoming a strategic geopolitical asset, and those who do not have their own AI models risk being excluded from an increasingly widespread productive resource. Do you remember when WhatsApp stopped working for a few hours and entire companies came to a halt? In that case, had the malfunction persisted, we would have quickly had many alternatives to move our work conversations. But in this case, if Trump managed to shut down ChatGPT, Claude, and Gemini, we would only have Chinese alternatives and some European models, unfortunately less performant.

Irony (?) produced by my Copilot

4. Something to Know: Very Sycophant and Little Sicophant

The title here could have been “Something I Should Have Known,” because in the last issue of the newsletter, I made a mistake that I now try to correct. I defined AI models as "sicophants", convinced I was labeling them as flatterers, servile, always ready to agree with me to please me. Too bad it doesn’t mean that at all, as an attentive friend kindly pointed out to me, to whom I extend my thanks. The term "sicophant," in Italian, means informant, spy, professional slanderer (we can call upon Treccani). At most, it is also the name of a beetle with an elegant blue and green metallic sheen. Certainly, "sicophant" is not the translation of sycophant, an English term that instead means precisely the servile and opportunistic flatterer I had in mind.

It is therefore one of those terms that English speakers call "false friends", false friends: words so similar to an Italian term that they immediately suggest an incorrect translation. Thus, something "terrific" is not terrifying but fantastic and exceptional, a "rumor" is not a noise but gossip, and so on with "preservative" (a preservative), "morbid" (unhealthy), "ape" (monkey), or the insidious "pizza with pepperoni," which is devoid of peppers, instead enriched with unexpected slices of spicy salami.

A true sicophant

5. This Monet is Ridiculous

The story is as follows: on X, a user provocatively wrote “I just generated an image in Monet's style using AI. Describe, in as much detail as possible, what makes it inferior to a real Monet”.

Hundreds of people responded, explaining with great competence the distance between this AI-generated work and one by the real artist. To understand, the average comment is “What a sadness to have to point it out. There is no cohesion in the choices of depth and color. The reflection of the tree smudges on the water lilies without any regard for spatial depth or contrast. The amalgamation between water lilies and algae in the background is scandalously vague, like most AI-generated art.”

You’ve already guessed how this story ends: the painting is actually a real Monet, from 1915, currently on display at the Neue Pinakothek in Munich. A study published in Nature conducted an experiment on several hundred people, showing that we are unable to distinguish images created by humans from those created by AI, that we generally prefer AI-generated artworks, but that if we think they are AI-generated, we like them less. The image on X was accompanied by the caption “Made with AI,” and it was certainly the presence of this label that activated the harsh judgment of those observing it. I believe it is a mechanism we know well and that pushes us to remove from the texts AI generates for us all those traces that seem to reveal it was written by the machine, because we fear the judgment of those who read thinking “Ah! It was written with AI!” For example – just to be clear – I always remove the horizontal dashes that enclose parenthetical statements, even though sometimes they would fit quite well 😉.

Curiosity: the author of the post generated an NFT from it (turned it into digital art) titled “inferior image,” then sold it for $42,000. In my opinion: brilliant.

6. Managers are the New Bottleneck

In an article from the Harvard Business Review, a manager complained: “every 30 minutes someone creates something I have to look at,” revealing a new problem many of us are now experiencing. AI has enormously accelerated the pace of work: people execute ideas, produce deliverables, and launch projects much faster than our organizations can manage. Each person uses AI chats and agents to produce at great speed, putting pressure on those coordinating them. The manager becomes the new bottleneck, overwhelmed by the volume of decisions, reviews, and feedback to provide. The real limit has shifted from production to judgment, and the temptation is to neglect this responsibility, only giving directions, priorities, decisions.

I also feel this way, and I believe we need to recalibrate our work methods to the new timelines. Since I am in a phase of life where I try to guide younger people, I often find myself struggling to review their ideas, project proposals, articles, which they write aided by AI. But for a few weeks now, I’ve realized I also need to dedicate time to reading what AI has written, to which I had assigned some tasks to perform for me. The feeling is that AI would do much more if only I had more time to tell it what and to review its work. I have become the bottleneck.

This new condition is evident to those using the project management technique (widely used) called agile development, whose SCRUM framework is the most well-known: instead of planning everything in advance and delivering the final result after months, work is done iteratively and incrementally, with short cycles of work, continuous verification, and progressive adjustments. In this logic, a complex project is divided into sprints, defined time windows at the end of which there must be something concrete to show. The sprints of an IT development project usually last between ten and fifteen days (the classic two working weeks). After those days, used to develop the planned features, the team meets to test and validate them, decide what to keep and what to correct, and plan the next sprint.

The problem is that today, with AI, a development "sprint" can last a few hours. The machine produces in an afternoon what previously took two weeks. But the human verification moment, where you look, judge, and provide direction, has remained stuck in its old rhythm. To keep up, we should meet to check results not every two weeks, but two or three times a day. And it is precisely there that we realize the constraint is no longer the speed of construction but the speed of decision-making. AI has taken away the excuse of production fatigue and is leaving us with that of judgment.

7. Who Graduates and Who Enrolls in University Today

Those about to graduate started their university studies shortly before generative AI became pervasive. At the time of enrollment, years ago, it was thought that studying computer science guaranteed finding a job and earning a good salary.

With the boom of generative AI, we discovered that machines find it very difficult to produce ironic text but are excellent at writing in mechanical languages like those used in programming. All this completely changes the scenario for those about to graduate in computer science or computer engineering: on https://layoffs.fyi/, we can see about 120,000 layoffs in the tech sector, open positions for juniors have decreased, and job interviews increasingly focus on problem-solving and critical analysis of what is achieved with AI. These are not the classic technical skills of those who develop code, who tend to present themselves showing what they have created, with what tools, possibly with a demo of such digital products. This is a problem if they then struggle to explain the problem, represent user requests, the architecture, the data source, risks, error cases, the evaluation approach, and subsequent evolutions.

Even those about to enroll in university take into account the current scenario where AI agents write software, and for the first time in many years, enrollments in computer science courses are decreasing in number. At Duke, introductory computer science courses have dropped by about 20% in one year, and at Princeton, a cohort of computer science graduates is expected to be 25% smaller within two years, in line with the national figure of an 8.1% decrease in enrollments in the 2025-2026 academic year, the sharpest decline in any field of study according to the National Student Clearinghouse.

In China, where there are about 3000 universities, the Government has decided to close about 12,000 degree courses, considered obsolete in the AI era. Obviously, the computer science course remains open and is even considered very important, but the attempt to build skills that today seem unavailable is evident. I believe it is obvious that the university not only has the task of preparing for the world of work but also of educating people based on their aptitudes and talents, also contributing to those sectors that, although they may not generate profit, remain fundamental for the well-being of a community (just think of healthcare or culture).

My impression is that no one knows how the situation is evolving, and those entering university today must learn a profession that we professors do not yet know how to teach.

8. Good AI News: A Special Vaccine

A group of researchers from Cambridge University has completed the first human trial of a vaccine whose active ingredient was entirely designed by an AI system, marking the first time the active component of a vaccine, designed through computer simulations, has been administered to people. Unlike traditional vaccines, built around a single strain already in circulation and therefore destined to lose effectiveness as the virus mutates, this approach aims for universal protection. The AI analyzed all available genetic sequences of a coronavirus family and identified traits common to the entire family, including variants not yet emerged, combining them into a so-called “super-antigen” capable of training the immune system to recognize the entire viral group rather than a single target.

The trial involved 39 healthy volunteers aged 18 to 50 and demonstrated that the technology is safe and free of significant side effects, and activated immune defenses not only against SARS-CoV-2 and SARS but also against animal coronaviruses that have never infected humans. The observed immune response was described as “modest” by the researchers themselves, and the duration of protection and efficacy on a larger population remain to be clarified, which is why a subsequent study with about 200 participants is already planned. But regardless of the results, the possibility has opened up to test the same method on influenza, avian flu, and hemorrhagic fever viruses like Ebola.

9. One of Our Projects: Data Coach

It is becoming increasingly simple to independently develop agents that perform tasks on our behalf. For example, I have an agent that takes the text of each issue of this newsletter I write and checks its spelling, looks for phrases I can shorten (because I tend to be verbose), verifies the links I’ve inserted, and suggests paragraphs to review that might be less fluid. However, when an agent needs to be distributed within a company, making it available to many people, connecting it to documents, data, information systems, it is necessary to carefully evaluate the data base being used. AI models are now a commodity without particular value, and if projects succeed, it is because these models read the right data. Consider that AI amplifies any errors present in the data, and in fact, the primary cause of AI project failure is the lack of information for the machines to use.

AGIC's approach to AI projects often involves a figure we might call “data coach”, dedicated to analyzing both the available data and the context that must be associated with such information. A customer record is data, but knowing why that customer receives a particularly discounted offer is context. A contractual clause is data, but knowing which clause applies to the specific case is context. The data coach is also able to guide the choice of the best model considering the data to be analyzed, because the most expensive model is not always the best: generally, it is the most powerful, but for many problems, a model requiring a smaller budget is sufficient. Finally, our data coach helps to understand where an agent is needed and where it is not needed at all.

Who I Am

Hello, I’m 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 new AI tools available, as well as observing and reflecting on digital evolution.