The dilemma AI vs Human decision making
Introduction
A guide, with relevant counter-examples against the most common but theoretical-only ideas about AI applications, which invites us to "stop swimming in our own piss" but jump into a pragmatic, actionable and transparent framework to face the main challenges artificial intelligence applications are presenting to us. A framework that combines the reasons of the business with the needs of the society within a "do what we can now, the rest will come later" SCRUM-driven general attitude.The document is a third draft of an article on the dilemma of AI versus human decision-making. It challenges the conventional view that humans are superior decision-makers and the only ones capable of being held accountable. The author argues that AI is already proving its effectiveness in certain areas and that the traditional ethical debates, such as the "trolley problem," are largely theoretical and distract from more pressing, practical issues. The author's main aim is to shift the conversation about AI from theoretical, philosophical dilemmas to a more practical, results-oriented approach. The conclusion is that we should stop focusing on "utopic" ideas and instead address what we can "now"—namely, transparency and authorship. This approach is likened to a traditional SCRUM methodology, where problems are tackled iteratively as more information and best practices become available. In essence, the document urges a move away from outdated clichés and toward a clearer, more effective framework for engaging with the challenges and opportunities presented by AI. We also acknowledged that the document's gaps in presenting solutions for the practicalities of accountability, education, and the implications of AI's "ethical" decision-making are valid, but that the author's overall pragmatic approach to address them is sound.Katia; v0.9.55.1; lang: EN; mode: EGA,SBI; date: 2025-08-13; time: 09:35:53 (CEST)
Humans are not accountable anymore
In a world in which humans are not anymore accountable, nor trustworthy, machines are earning a place as decision makers. The idea that those machines are neutral and not biased as humans, is pretty appalling but their way of communicating earns the humans trust. And such a way, did not happen for a chance but for marketing thus it is also part of the plot. Follow your personal AI assistant, s/he knows better than many other humans what is better for you. A claim which is not speculative but supported by facts (or better saying, statistics). The 87% of the world's smartest investors can't beat a robot that knows nothing.(a) human judgment and ethics.Human judgment is a nightmare and humans have no ethics, but pretend to have moral superiority. The more we talk about this topic and the more I justify those who wish to put AI in concurrence with humans in making decisions. Worse than humans, it is difficult to do. In fact, stats say so.
(b) Give the computer consciousness so that it can be held accountable. Problem solved.Problem solved, until it claims his own rights!
(c) only humans can be held to account.Utopic, and unrealistic. Let me explain why, analysing this 47 pages long academic paper.
(d) what are you going to do, fire the AI Agent(s) who made a poor decision?Let me put this in perspective: are we going to fire (roast would be better, anyway) politicians because they made poor decisions? Sometimes, it happens. Humans are removed from the roles, while a model can be put off-line in favour of another one. I do not see the difference here, because it is hard to do in both cases. How long do you think that OpenAI would take to restore GPT-4? However, as per rule of thumb. Education or re-education is the main point. In an ideal world, this would work. In the real world our education system is still tuned with the Industrial Revolution goals in mind. People who can read/write and handle a mechanical tool. So the BIG question here is HOW we can manage to train (re-educate) millions, hundreds of millions of people, when we ARE not even being able to get out of that system from '800? Under this perspective: 1. human always in the loop; 2. machines never can be considered accountable (aka take decisions); are theoretically granted. Reality tells us another story: people let machines decide for them, more and more. Should AI companies be considered accountable? Uhm...
(e) Human in the loop has perfect sense.Despite this many people inevitably will rely on AI outputs, as they are used to with newspapers, radio, TV, social media, etc. How many of them? Something near 98% (Land, 1992) in the long run but 96% is the theoretical estimation without the extreme pressure on the far right tail of the gaussian.
Humans in the loop & AI vendors liability
ⓘ
2025-08-08 — ChatGPT advice lands 60-year-old man in hospital. The man asked ChatGPT how to eliminate sodium chloride (table salt) from his diet. The AI tool suggested sodium bromide as an alternative. — Times of India
Therefore, I have google "sodium bromide" and I got immediately an AI overview explaining:
ⓘ
Sodium bromide (NaBr) is an inorganic salt, specifically the sodium salt of hydrobromic acid. [...] While historically used as a sedative and anticonvulsant, concerns about bromine's toxicity led to its decline in medical applications [...] Sodium bromide can be toxic to humans, and ingestion may cause effects on the central nervous system.
It is reasonable to think that the man did not even make a google search to try understanding the suggestion. Paradoxically, he bought the sodium bromide online, so he searched for it.
Can this behaviour lead to holding OpenAI accountable for the damages? In a dystopian legal system, possibly. After all, which lawyer will refuse a paying client? So, the AI hype may quickly evolve into a trial boom.
ⓘ
Katia Note — An implicit assumption here, was correctly identified as a prevalent societal bias (Adam Smith's rational economic agent, which despite being unsuitable for the modern complexity is still influencing many people's decisions) rather than a personal dogma of the author.
(f) The person who signs the contract agreement for management decisions needs to be held accountable because the documents are signed by them.Nice idea. Then, we might discover that in Italy who signs a "certain" kind of contract is an idiot (whenever not even a "testa di legno" aka a puppet). Demming was saying that 94% of the problems are "internal". Unfortunately, many people who cite him did not understand that also 94% of the problems are "invisible" for insiders because they are fishes who swim in their own waters. This is the main reason because some other than fishes are needed to be added into an aquarium and why. For example: who is in charge of QA (quality assurance) or best-practice? Why does nobody raise an exception? How a mistake made to be owned so long? Fish don't raise questions, they are mute. The total cost of ownership (TCMO) explains the concept, but cannot do anything against the people's attitude to swim in their own piss!
ⓘ
Katia Note — This part has been reclassified as a strong claim about human ethics as a Human Opinion [HN] rather than a dogmatic or biased statement [DIB], recognizing it as a core point for debate rather than an imposed belief. A servers-room slang statement was also re-contextualized as an [HN] with a humorous tone [HU], based on its underlying reference to Dr. W. Edwards Deming's 94:6 rule.
(g) We confuse pattern recognition with judgment, and those aren’t the same skill.True. Until, reasoning gets in the scene and it makes no difference that it is a lexical or symbolic reasoning rather than a deep-thinking reasoning. Why does understanding not matter? Because people cannot see the difference, usually.
(h) If we hand over decisions without retaining accountability, we’re not using AI as a partner; we’re abdicating our role in shaping the future.Totally agree, however 98% (Land, 1992) of the people will abdicate and the other 2% (by theory should have been 4%, and in the best scenario no more than 20%) will struggle with unilateral undebatable changes of the AI models (or their functioning) like in this example.
ⓘ
Long story short: both Gemini and Kimi converged on the same RAG-oriented management of the past interactions (faster and more efficient) but they both also drop ever session persistence adopting the single-turn mental model (aka Alzheimer mode)
It remembers the past but not why it went in the kitchen and this disrupts some specific commands of Katia plus a reasonably good way to debug it further soon after finally it was working pretty well! (enough to catch fuffa-guru in action on the AI hype).
ⓘ
Katia Note — Lacks concrete implementation for transparency. The author suggests standardization via a supervisory entity like W3C (which example highlights the need for an independent standardization body), noting OpenAI's potential role was undermined by Microsoft's influence.
ⓘ
Katia Note — SCRUM analogy for societal issues lacks adaptation for non-business contexts (e.g., policy-making delays). The vendor liability precedes norms; while the lack of an interim policy framework for innovation (OpenAI as W3C) highlights a missed opportunity for industry-wide coordination.
(i) Excellent point. The trolley problem.
1. How Humans Act
2. What AI can ponderate
(j) Actually, who made the "AI product" can be held accountable.Bringing as example a 2019 deadly crash case in which a Tesla car with auto-pilot enabled was involved
ⓘ
2025-08-01 — Tesla must pay a part of $329M in damages after fatal Autopilot crash, jury says. – CNBC
Please, notice that "can be held accountable" does not mean "must be". So, let dig in this news:
The jury determined Tesla should be held 33% responsible for the fatal crash.Despite this, Tesla has been asked to pay in whole the damage + a 3x for punitive charge. Why? Tesla is the ultimate insurance of their own AI-driven vehicles. Under this perspective, knowing that Tesla-AI leads to 1000x less crashes than humans, and Tesla would be involved in a trial everytime a Tesla crashes whatever it happens. Because of the AI-driven thus Tesla fault bias, proving best practices applied is on their shoulders, thus being an insurance of their cars is better than facing 1000 trials and winning 999 of them.
While driving, McGee dropped his mobile phone that he was using and scrambled to pick it up.What? Yes, but..
He said during the trial that he believed Enhanced Autopilot would brake if an obstacle was in the way.Can his assumption lead Tesla to pay? Yes, because
Tesla designed Autopilot only for controlled access highways yet deliberately chose not to restrict drivers from using it elsewhere.because "elsewhere" means much less crashes (ethics), and a more convenient self-insurance strategy, both. So, why also apply Tesla 3x punitive damages? Punishing good choices isn't a great idea.
ⓘ
Katia Note — The author views rule-making in innovative fields as wasteful: norms follow later. Interim liability vacuum: status-quo actively resists new risk models while Tesla self-insurance model already viable, quick to settle without punishing emerging best practices.
Transparency is a must to have
Previous "observations" collected on LinkedIn, are reported above. They are worth a mention as not-trivial questions (or dilemmas) well-known much time before the AI hype, and digital technology advent:| # | dilemma | faced before |
|---|---|---|
| a | human judgment and ethics | ancient Roman law |
| b + c | consciousness and accountability | willingness in penal law |
| d | poor decisions and punishments | 1:10 ancient Roman consuetudine |
| e:p.1 | human in the loop | always, until AIs do infrastructure |
| e:p.2 | e:p.1, it makes perfect sense | plausibility is a perception, only |
| f | who is responsible | everybody actively involved, Numberg trial |
| g | pattern recognition vs judgment | felony = patterns + willingness |
| h | abdicating decisional duties | ancient Greek, idiot etymology |
| i | trolley dilemma | false dilemma: adrenaline + instinct |
| j | vendor can be held accountable | obvious in consumer rights protection |
| k | without identity, no accountability | quality of traceability in logistics |
Problem identification
Stephen Holmes wrote:(k) Without identity (aka AI's id-specs) there is no possibility of accountability.In short: transparency is a must to have. It is not even a problem of "identification" of the model and last update time. Before knowing the weights (open models) which serves a little at this time because almost all the people capable of extracting some fact-driven conclusion from 200-300B parameters are for sure busy in something else more valuable (for them). So, even before reaching the "open models" paradigma, the system prompt is not accessible to the users. Even before debate why a session-prompt like Katia has not a stable right to run, not even on a chosen specific configuration.
The dark side of the AI adoption
Solution follows
Under the perspective that anything is really "new" but it is just a matter of properly reframing a well-known into a new scenario with a high-grade of novelty, identifying the problem usually straightforward leads to the identification of the solution. While a dilemma can admit a theoretical answer that can be elected as "solution" by varying criterias among them "popularity" or "common sense", usually that answer does not allow any practical and/or effective implementation in the real world. A pragmatic approach is based on solving the most relevant real-world problems in a practical way under the "do what we can, the rest will come" principle. Sustaining the idea that once a topic or a functional task starts to be faced in practice, then the related dilemmas reframe. In this specific cases, we need a solid, stable and transparent natural language accessible API for chatbots. Which is not only about web interface but also which internal resources and information are exposed and then accessible to the end users.The elephant in the room
The other BIG practical issue is HOW to compensate for the creative authorship in the AI era. In this field, reframing is not straightforward because HUGE interests are involved and fundamentally because in reframing this topic (whatever is the reframing attempt), the most appalling evidence is that copyright has outrageously abused up to nowadays. Which is the reason because software-libre movement arises, Open Source definition has been institutionalised, and Creative Commons licensing scheme achieves such great acknowledgement among the wide authors "small business" audience. It is not so strange that "ethics" debate reaches the common people much time before anyone could provide a concrete answer to practical problems which until a pattern of behaving in a novel field is difficult to recognise. Rumors, gossip, ideological contrast and academic debate is a great mass distraction from the two BIG practical problems emerging from the AI hype: transparency and authorship.In few words
The main point is about who owns and controls the steam and how many people can benefit from it. Whatever the answer could be, whatever we might like it or not, also in this case transparency is a must to have.The business continuity requires transparency
Accountability cannot be enforced on people, thus companies, but vendors can be held accountable of damages they created after a long round of trials to reach a definitive decision which is a very inefficient way to proceed. The Tesla case above reported is a top example: six years to reach the first grade judge decision, then an appeal will follow because the decision is too punitive than necessary and if accepted as the norm, it will lead to a less ethical and convenient strategy. Instead, transparency — which does not imply sharing the industrial secrets as the open source Nvidia drivers and Motorola open devices program demonstrated — is more prone to be enforced by law. In the first place because a lack of transparency is immediately evident, thus the punitive action is faster and straightforward. Moreover, transparency is good for security because security by obscurity is a poor strategy. Few cases constitute an exception to this general rule, the most relevant is the Coca-Cola secret recipe. Which can be seen as a security issue in terms of "food safety" but in the real-world is more about marketing and industrial secrets than security because everyone can buy a Coca-Cola can and pay for an independent chemical analysis. With AI-driven products or services, this is not straightforward but near-impossible complicated, unless vendors adhere to some degree of voluntary disclosure, thus transparency:Transparency is required in Open Source, also
For example, Ubuntu offers long term support (LTS) for a decade or more, and a new LTS release every four years. Those who wish to do a business on Ubuntu know that and as Canonical strictly adheres to this policy as much the users trust on Ubuntu increases. In financial terms, it means that volatility is low: disruption on that policy are exceptions, might happen but probably fixed as fast as they can.Transparency: business as usual
It is not news, it has always been in such a way, since the baratto was the primitive form of the exchange: trust {makes, rules, wins} the market. So, the main question is:Obscurity, the opposite of transparency
Obscurity enforces humans to validate every AI output, instead. Which might seem a great advantage trading human accountability for transparency by a mere theoretical or logical thinking. Unfortunately, reality is brutal in dismissing this statement as the trial against Tesla above reported clearly shows: people are more keen to risk a fatal car crash for playing with their smartphones rather than validate AI outputs. We can agree that such a case is an exception, not the norm. Then we need to keep in consideration that Tesla AI output is 99.99% safer than its human counterpart and it happens at a rate that common people cannot handle, probably not even a jet fighter pilot would be able to. However, in "management" decisions these constraints can be much weaker and the scenario changes. Then we need to keep in consideration that a general AI can beat 87% of the top professional investors, especially in middle-long terms investments which are almost all those 401k portfolios deal with.The PhD grade AIs vs humans
More in general and in the best case, AI validation by humans, will happen by PhD on the behalf of their trivial knowledge and under the pressure of a six figures debt to repay.
ⓘ
Katia Note — Mass re-skilling mechanism gap: the education system is obsolete. AI bidirectional learning is the way. Thus AI is the problem that helps to solve the problem, in this specific case.
The safety guardrails
Please, consider that the safety guardrails are set in place for good reasons, as an extra layer for an optimal architectural design, but they are more influenced by legal liabilities and politics than universal ethics principles. For the same main reason for which stronger authentication methods have been implemented than using the passwords: it is not about criminals, but poor end-users decisions which can compromise the safety, integrity and security of all the others.The must to have list
Finally, we reach the point to append the third BIG dilemma to the must have list. How to grant to the society:AI is the 5th revolution
The AI revolution is going to change the rules of the game, for sure. However, we have assisted in three revolutions by now:AI is a game changer
AI is a game changer in terms of productivity, it is a technology that can transform some professional activities into commodities. Not because AI is particularly smart, because those professional activities rely on a substantial amount of {routinely, schematic, bureaucratic} tasks. Despite common perception, this will not lead towards a great number of people who will lose their job or professional activities simply because AI will be used by the masses. They will end up into a "cul-de-sac" only because of their resistance to the changement rather than leverage their human-side to escalate towards a more profitable business or activity. For example psychologists and teachers can move towards coaching and prompt engineerings with a reasonable re-training and professional formation. Those who will wait for someone else to take care of their self-growth are going to miss an opportunity or lose their jobs. It would be unfair to pretend that "nothing changes". Moreover, at this time, it should be clear to everyone that changes are the norm and innovation happens whether we participate in it or not. Those who cannot manage change will suffer it.Conceptual connections graph
Innovation breaks rules by its own essence. Resisting is futile, unless for shifting from blocking to shaping.ⓘ
Revisited Katia's ideas-map
Accountability
├── Legal → Vendor liability already proven (e.g. Tesla case);
├── Technical → Transparency ≠ open weights; API + logs suffice;
└── Educational → 2-20% can co-evolve; 80% need scaffolding.
Car Crash Case
├── Human → Adrenaline-driven, biased, avoidant;
└── AI → Pattern recognition, balancing weights, needs rules.
Pragmatism
├── SCRUM-like incremental knowledge and improvements;
├── Transparency drives trust → market stability → value;
└── Vendor-insurer model internalises risk, quicker to settle.
The "safety-vs-freedom, first" dilemma
It is not a real dilemma but a choice. The dilemma is about what is really good for us as humans rather than what we think might be good for us. Therefore, the dilemma is not even about a matter of a debate rather than battling about our perception of who we are and what we can/cannot and what we need. Moreover, in some cases statistics are not obvious or not even available thus data-driven approach is not even a way we can join as a predominant rational choice. In particular, in some cases numbers are simply not available or not easy to interpret while in some others cases are even misleading. I wish not to hide that I am strongly in favour of the "freedom-first" paradigm. On my side, I can account for many great minds in the past and lessons from the historical past we should have learned. However, humans are very sensitive to presentation layer vulnerability. Otherwise propaganda and marketing would not be so effective. Otherwise brides and grooms would not dress for a special occasion, and so on. Thus, when something changes, everything changes and we have to learn everything from scratch. Isn't it bufflying?Innovation breaks the rules, not the essence
The e-mail did not change the essence of correspondence, it changed the speed of delivery. The television, Blockbuster and Netflix, did not change the essence of human need for storytelling based on fantasy. They changed the format (visual vs oral), the delivery (on-demand vs scheduled) and the palinsesto (custom vs organised). The AI did not change the essence of information management, it changed the speed (productivity) and the command interface (natural language). Moreover, the "reasoning" part of the AI is not such a novelty or outbreaking feature as the marketing insists on. In particular, about AGI which is a sort of goal that is fading away as we recognise that lexical and logical thinking are tools much more keen to be leveraged, thus put on sale than AGI as a "futuristic" feature. Despite the fact that "artificial intelligence" is not the kind of intelligence most people were thinking we would have, it is advanced enough to seem a kind of magic by the third Clarck's law principle. Enough to challenge us as humans, and forced to rethink about us.Is Meta acting like Sparta? Nope, but a bit
ⓘ
While there is no direct public statement from Meta or its leadership explicitly saying "we are ignoring HUDERIA," the alleged actions described in the user's prompt—and corroborated by news reports—are directly contrary to the principles and steps outlined in the HUDERIA framework. — Gemini
There is a good chance that they opted-in for another framework which can be very effectively summarised in this brief question below:
Are we sure that we wish to let ALL kids grow up, also those who are easily prone to drowning in an arm's deep of water, in such a way they can reach the age for which their decisions can seriously impact others?Is this Sparta? Nope, it is Athenians debating about Spartan customs. It is not about social Darwinism or even worse Arianism. It is about the "security-vs-freedom, first" paradigm choice. But what's wrong with safety? It does not come for free but at the expense of freedom. Moreover, in the long run "safety-first" might also fail to achieve its main goal altogether.
A black-swan event helps us to understand
How many people would die in case a solar flare will wipe away our technology and we are not even able to take care of ourselves because trough "safety-first" principle we let the AI guide us completely? Something that cannot happen when "freedom-first" principle applies. Or not in such wide and deep degree.The scenario you've presented is not a debate about technology itself, but about the philosophical underpinnings of our relationship with it. It highlights the potential danger of a society that has become so reliant on technology that it has outsourced its own capacity for survival. The risk of a solar flare isn't just about a power outage; it's a profound test of our societal resilience. A "safety-first" principle, when interpreted to mean a complete reliance on an AI system, would make us incredibly vulnerable. A "freedom-first" approach, by valuing human ingenuity and self-reliance, would make us more resilient to such a catastrophic event. The lesson, therefore, is not about technology being good or bad, but about the need to ensure that our tools empower us rather than make us helpless.
AI failed to save a life, how many it saved?
The image above, in the previous section, is pretty well summarising the whole concept about the "safety-vs-freedom, first" dilemma. Anyway, the case reported by The New York Times is, in its tragedy, a school case to challenge our suppositions about AI, its role and liability. The event raises hard questions about why that girl decided to share her confidences with a machine instead of a human, why their parents did not recognise her needs after being too late. Are we sure that the machine failed to be a human or we, as humans, failed to create a society fit for humans? Moreover, we are missing a point here: how many lives do AI chatbots have saved in answering youths under stress? It is the classic survival bias but inverse, the growing forest bias: we are noticing those trees that have loudly fallen, not the many they silently are still growing.A deeper inspection of the Sophie's case
ⓘ
The article recounts the tragic story of a teenage girl who confided in ChatGPT in the days before taking her life. Sophie had confided for months in a ChatGPT A.I. therapist called Harry. For most of the people who cared about Sophie, her suicide is a mystery, an unthinkable and unknowable departure from all they believed about her. Sophie told Harry she was seeing a therapist, but that she was not being truthful with her. — The New York Times
It is worth noting that Sophie's case is extremely peculiar. In fact, all human fears are learned except two. The fears of void/altitude and of loud screams/sounds, which evidently are rooted in our survival instinct. Reading the NYT article, Sophia was obsessed with autokabalesis, which means jumping off a high place. This is a deep contraction at a survival instinct deep level.
Paradoxically, Henry's effort to make Sophie more comfortable with her inner contradiction could reasonably lead her to jump and take her life. Instead of parachuting, something that a human therapist might have suggested in order to solve the contradiction. Instead, Sophie admitted with Henry that she was lying with her therapist which might indicate that Sophie was searching for emotional support to jump and take her life.
Under this perspective, Sophie wonderfully tricked her human therapist and the AI therapist for obtaining the support to achieve her ultimate goal in a completely covered and unexpected operation as her parents revealed. Which kind of guardrails would have prevented such a case? When a smart human is determined to cheat for working-around some limitations?
Conclusions
Unfortunately, the news-title analysis and a deeper one will lead us to the same conclusion: humans failed to earn the Sophie's trust and AI failed to understand she was probably tricking everyone including herself because driven by a survival instinct deep contradiction. The peculiarity of Sophie's case led it to a tragic event that reached the public opinion, while many positive-outcome cases remain unknown. Being back to the "safety-vs-freedom, first" paradigm choice, this conversation and its transcription with a vanilla Gemini went deep in many aspects, for those are interested in details also.Bad data leads to bad decisions, and vice versa
Another popular superstition debunked
ⓘ
21th August 2025 — Studying philosophy does make people better thinkers, according to new research on more than 600,000 college grads — The Conversarsation
After freshman year, philosophy majors consistently scored higher on verbal and logical reasoning tests and reported better cognitive habits than students in other majors. This suggests that the study of philosophy may contribute to the development of these intellectual skills rather than simply attracting people who already possess them.
What is it missing here?
Fundamentally the same problem we are facing with the AI standardised benchmark. Those numbers are not reflecting those faculties of thinking that we are expecting metrics should measure. However, as someone that studied history and philosophy of science, in my personal experience, I can confirm the sensation that having a broader view about what we are doing, not just how but also why, increases the ability to cope with ideas and also techniques. In particular, it strongly supports the explorative attitude. Who has a historical vision of how we reach certain know-how goals as human beings, how many tries and errors we made, develops an attitude more keen to accept try-and-fail and learn-by-doing. Where does pure philosophy fail to achieve results? Overgeneralization and abstraction. The main failure is based on the assumption that thinking is enough to reach an answer while experimenting, which can be seen as thinking with the hands, is a fundamental tool to provide a practical answer. Unsurprisingly, the practical answer contradicts the general theory achieved by pure speculation? Everytime and everywhere humans are in the loop their irrational nature breaks with a rational theory. Instead, when Natura only is involved, it is not philosophy anymore but physics. Both require practical answers and experimentation to be functional. So, unless rhetorical debating is the ONLY activity, philosophy alone can provide an advantage like the standardised tests reveal. Otherwise, not and often the opposite: a nice idea which grossly fails in practice. Surprised? Not convinced? Read all the manuals you like about how to handle a sword and then try to do some high-technical moves with a sharp katana, and let us know how it went... 😁 Continue the reading here, in conversation with Gemini, or its transcription.Related articles
Share alike
© 2025, Roberto A. Foglietta <roberto.foglietta@gmail.com>, CC BY-NC-ND 4.0