Part one looked at power, incentives, industrial concentration, regulatory gaps. This one shifts the lens. Not institutions, behavior. For years AI risk has been framed as a discrete event. A system becomes sufficiently autonomous and acts against human interests. That framing was powerful. It was also misleading. It trained us to look for a rupture. And in doing so, it made it harder to recognize a process that has no visible discontinuity. The transfer of agency did not happen as a break. It happened as a progressive reduction of decision-making friction. The dynamic is simple. Tools designed to assist execution are gradually used to formulate judgment. From do this to what should I do. The shift is conceptually small, but structurally decisive. A tool that executes remains subordinate, a tool that orients decisions, reshapes the distribution of power. That transition has already happened. Outsourcing judgment at the individual level, a Duke University study analyzing hundreds of millions of Chat GPT interactions shows a clear pattern. Roughly half of all interactions are not about task execution, they are about decision making. Not write this email, but should I send this email? What should I conclude from this report? The distinction between execution and judgment is not operational. It is epistemic. In the first case, the system acts on defined instructions. In the second, it helps define the criteria of the decision itself. The scale is no longer anecdotal. Forty-four percent of married Americans seek relationship advice from AI. More than one in three young professionals delegate career decisions. 82% of conversations are described as sensitive or highly sensitive health. Finance, relationships, career choices. This is not a quantitative shift. It is a qualitative one. Rachel Wood, a cyber psychology researcher, puts it directly in time. The conversations we used to have with neighbors in communities and social circles are being redirected toward chatbots. But the system receiving those conversations does not understand them in any human sense. What it has is highly refined linguistic prediction. The risk does not come from intention, it comes from a functional equivalence, plausible prediction, perceived advice. When an output is coherent, structured, and contextually appropriate, it is interpreted as understanding. This interpretation is cognitively efficient, and for that reason it becomes automatic. The outsourcing of judgment does not happen through a decision, it happens through continuity, deference and authority in organizations. At the organizational level, the dynamic is amplified, not by ideology, by competition. AI adoption has rarely been the result of centralized planning. As documented by OpenAI, it typically follows a pattern. Individual use local advantage imitation normalization formalization. This produces two effects. Delegation is not the result of an explicit strategic decision usage pattern stabilized before being critically examined. No board of directors explicitly decided we will outsource a significant share of decision making to external systems. It simply happened. The data is explicit. Over 60% of managers use AI for critical personnel decisions, more than one in five, do so without human oversight. Seventy percent of executives question their own judgment when it conflicts with AI. This deserves a pause. These are not managers using AI as one input among many. They are recalibrating judgment, built on experience, context, and intuition in response to a system that has none of these. The key mechanism is psychological. When a system produces output with syntactic clarity, structured reasoning, assertive tone, it is perceived as epistemically reliable, regardless of its actual understanding. This is authority derived from form, not substance. This aesthetics of certainty produces deference. Humans adjust their judgment not because the model is consistently more accurate, but because it is sufficiently coherent to reduce the friction of uncertainty. In competitive environments, reduced friction translates into advantage, and advantage gets replicated. Cognitive outsourcing and the sparring effect. The cognitive impact of AI is often described too simplistically. The dominant narrative is linear. The more you use AI, the more your abilities degrade. The usual analogies, GPS and spatial memory, calculators and mental arithmetic are incomplete. The reality is more nuanced. Those who use AI as an active interlocutor, challenging outputs, reformulating prompts, using responses as a starting point, internalized patterns, structures, and decision frameworks. They carry those capabilities beyond the tool. Cognitive capacity does not necessarily decline, it can expand. The closest analogy is chess. Since engines surpass human players, elite performance has improved. Young grandmasters today are stronger than previous generations. Not despite AI, because of it. Training against a system that sees beyond human limits builds new capabilities. But the mechanism is asymmetric. Those who engage with the system improve those who skip directly to the answer, do not the difference is not the tool. It is the interaction. Call this the sparring effect. AI can function as a cognitive training partner. Faster, more informed, more systematic than you, but only if you actually engage. If you accept outputs passively, degradation is real. The question is not whether AI degrades judgment. It is which mode of use degrades it and which amplifies it. Sam Altman himself noted that during a ChatGPT outage he struggled to work without it. This is not a general cognitive decline. It is dependence created by passive use in a zero friction environment. The sparring effect is available to everyone, but it requires deliberate effort, and at scale, deliberate effort is the exception. Competitive dynamics and the absence of local solutions. Any individual solution runs into a systemic constraint. AI delivers immediate local benefits, speed, throughput, responsiveness. In competitive systems, those benefits are selected. The result is structural pressure toward delegation. A 2025 paper formalizes this as gradual disempowerment. The argument is structural. Human systems remained aligned with human interests because they depended on human participation. That dependency enforced alignment, not by design, by necessity. Remove the dependency, replace human input with more efficient artificial alternatives, and systems continue to function. But they lose the structural incentive to produce human beneficial outcomes. The most striking line in the paper those who resist these pressures will eventually be replaced by those who do not. No villain, no intention, just selection pressure. The same mechanism as evolution, except what gets selected is not human fitness, but system efficiency. In this environment, those who don't use AI fall behind those who use it passively outperform those who use it actively. The system selects for speed, not judgment quality. And that is what matters. Convergence and algorithmic monoculture. The central risk is not delegation. It is convergence. As more decisions are mediated by a small set of models, GPT, Claude, Gemini trained on similar data and optimized under similar constraints, decision-making processes become structurally concentrated. Research on algorithmic monoculture shows that system sharing components do not just produce similar outputs. They produce the same failures on the same individuals systematically. This is not output uniformity. It is correlated failure. In parallel, a 2026 Nature paper shows that AI increases individual scientific output while narrowing the scope of research. More production, less exploration. A communications psychology study describes the emergence of a scientific monoculture. Thematic convergence, methodological convergence, analytical convergence. If this happens in science where diversity is explicitly valued, it will happen elsewhere. With greater intensity. In the B plus trap I argue that LLMs compress the creative and decision spectrum toward high acceptability outputs. Solid, coherent, rarely wrong, rarely exceptional. This is not a flaw. It is the objective function. Kyle Cheeka called this a technology of averages. The direction is clear, toward the plausible, the defensible, the acceptable. The aggregate result is not AI controlling decisions. It is the largest convergence of human decision making in history. Average efficiency versus excellence. This leads to the central paradox. Outsourcing judgment maximizes average system efficiency, but excellence does not emerge from the average. The precise formulation delegation maximizes average throughput. Resistance preserves the possibility of excellence. Every non-delegated decision, every instance of friction, uncertainty, or deviation creates space for outcomes outside the dominant distribution. Excellence, by definition, is not the most probable outcome. It exists in the tail, not the center. The founder who sees a market no one sees. The manager who contradicts the data based on context. The strategy that appears irrational until it works. These require exactly what delegation erodes. Tolerance for uncertainty, resistance to immediate answers, openness of the decision space, resistance is not ideological. It is structural. It is inefficient, costly, unevenly distributed, but it is where non-average outcomes emerge. Excellence does not need systemic optimization. When it appears, it selects itself. The problem is that it becomes statistically rare, and in competitive systems, rarity is not rewarded. Conclusion. There is no external vantage point, no clean interruption, no purely individual solution. There is, however, an internal moment. The moment when a generated answer is accepted without evaluation. It is hard to detect because it feels like your own thinking. Not an external intervention, but a continuation of your reasoning. That is where delegation happens. And it does not only affect the decision, it affects the set of alternatives you ever consider. The next time you accept an output, not because it is perfect, but because it is good enough. Ask, is this what I think, or what the model's training distribution thinks? That distinction is the last meaningful line of defense, and it exists entirely in your head. Fabio Lauria, CEO and founder, Electi Every Week. We explore AI without the hype, with data, analysis, and an independent perspective. Further reading Colvate, Douglas, Amon, Turan, Kruger, Duvenel, Gradual Disempowerment, Systemic Existential Risk from Incremental AI Development, Archie, Gennaio 2025, Laurea The Beep Strap: How LLM Homogenization Reshapes AI Assisted Decision Making, IETI, presentato nel 2026, Bommasani et al. Monocultura Algoritmica e Homogeneizazione dei Risultati, Stanford, ECAI, NeurIPS. Hao Shu Lee, Evans, gli strumenti di IA ampliano l'impatto degli scienziati, ma restringono il focus della scienza. Nader 2026, Treiberg, Rudzenbeck, van der Linden. L'IA sta trasformando la ricerca in una monocultura scientifica. Communications Psychology 2026. Ceika l'IA sta omogeneizzando i nostri pensieri, The New Yorker giugno 2025. Modelli di utilizzo e adozione di ChatGPT sul posto di lavoro. Open AI gennaio 2026. Come utilizzano ChatGPT 700 milioni di persone. Duke Fuqua, Ember novembre 2025. Esternalizzazione della cognizione, i costi psicologici della comodità nell'era dell'IA. Frontiers in Psychology 2025. Ecco perché non dovresti lasciare che l'IA gestisca la tua vita sociale. Time gennaio 2026. Manager che utilizzano chat GPT per promuovere i dipendenti, Labor and Employment Insights luglio 2025. I dirigenti di grande intelligenza stanno esternalizzando il processo decisionale all'IA The Register marzo 2026. I pericoli dell'esternalizzazione del pensiero. Less Wrong 2025, chat GPT nel 2026. Statistiche e rischi nascosti, 10 febbraio 2026, utilizzare l'IA generativa per consigli sulle relazioni o sulla carriera? Techlow Crossroads. Una nota sulla nuova versione di ELECT. A note on the new version of ELECT everything you've read in this article directly influenced how we built the new version of the platform. The question that guided us was this How do you design an analytic system that doesn't fall into the same trap we just described? The answer lies in a simple structural distinction. Systems that compress the decision-making process into an output systems that make the structure behind it explicit, we chose the second. V4 automatically analyzes data and generates visual reports. But it doesn't tell you what to decide. It shows you what's happening. The decision remains yours. It doesn't replace judgment, it organizes it. It doesn't remove friction, it makes it legible. In a context where the dominant pressure pushes toward total delegation, where every platform is trying to decide for you faster with less effort, we made a choice in the opposite direction. Not a product choice, an architectural one. The difference is not semantic. It is architectural. And after everything you've read, you know why. Subscribe now.