When Your Team's AI Agents Outpace Trust
When AI becomes your team's primary collaborator, trust beliefs can't stabilise fast enough to keep up.
AI adoption alters the signals that organisational trust judgement depends on, without replacing the underlying mechanisms. Three dynamics follow: competence signals compress toward the mean as the tooling floor rises; the tools themselves evolve faster than mental models can track; and the structure of work shifts toward small human teams directing AI agent stacks, reducing the human-to-human interaction surface where trust evidence is generated. The compound effect is not a trust collapse but a noisier allocation of trust — decisions made at full speed on degraded inputs. This is a design problem rather than a behavioural one. Behavioural prescriptions operate at individual speed in an environment moving at technological speed. Structural interventions — designing for outlier visibility, rewarding adaptation, making calibration legible — can move faster, but require frameworks that can be tested empirically rather than defended by argument.
The previous article [1] argued that organisations run Putnam dynamics at sprint velocity — the same belief-formation mechanisms that build social capital across decades, compressed into quarters. That argument left a question open. If the mechanisms are compressed but otherwise intact, what happens when the environment they operate in is itself changing rapidly? AI is the obvious case to examine. It is reshaping how work is produced, how capability is expressed, and how teams are structured — all on timescales shorter than the trust-formation cycles that depend on those signals.
This article works through that question. It is not an argument that AI breaks trust. The cognitive architecture for trust judgement is robust and continues to function. The argument is narrower and, I think, more interesting: that AI alters the inputs to trust formation in specific ways, and that the alterations compound. What the alterations produce, and what to do about them, is the rest of the article.
I. The competence signal compresses
AI accelerates convergence to the mean. When everyone has access to the same capability amplifier, the floor rises — work that was previously difficult becomes routine, and the baseline quality of output increases across the board. The productivity gain is substantial even without sophisticated use of the tools. This is broadly positive. But it compresses the visible range of competence. The gap between the 30th percentile and the 70th percentile narrows, because AI lifts the lower performers more than it lifts the higher ones. The distribution tightens.
What counts as competence in this environment is itself a moving target. The skills that differentiated an effective AI user six months ago are already commoditised. The prompting patterns that produced exceptional output last quarter are now built into the default tooling. Competence is no longer a state you achieve — it is a rate of adaptation you sustain. And a rate is harder to observe and evaluate than a state.
But the competence belief problem is one of signal-to-noise. In a compressed distribution, the evidence required to distinguish genuine excellence from competent-enough becomes harder to observe. Humans do not stop making trust decisions — our cognitive architecture makes hundreds of trust judgements daily, in and out of work, and it does not wait for perfect evidence. What changes is the reliability of those decisions. When the signals are finer-grained but the decision-making pace stays the same, the result is more frequent misjudgements — trust placed where it should not be, withheld where it should not be, calibrated to signals that no longer differentiate. We have not yet adapted to reading competence in an AI-augmented environment. The old heuristics — speed of delivery, polish of output, confidence of presentation — are precisely the signals that AI compresses to the mean.
Identifying and rewarding competence in organisations was already a problem with known pathologies before AI entered the picture. Pluchino, Rapisarda, and Garofalo demonstrated computationally that promoting the best performer — the intuitive strategy — can systematically degrade organisational efficiency when competence at one level does not predict competence at the next [13]. I had the pleasure of discussing these dynamics with Andrea Rapisarda during a week-long summer school in Lipari in 2019 — the fragility of competence measurement in hierarchies was already a central concern in computational social science well before AI amplified it. Their Ig Nobel Prize-winning finding — that random promotion outperforms merit-based promotion under certain structural conditions — revealed how fragile competence measurement is even in stable environments with clear performance signals. AI makes this fragility acute. The signals that promotion, delegation, and trust decisions depend on are now compressed into a narrower band, measured against a shifting baseline, and produced through a process that is increasingly opaque to the evaluator. The problem is not new. It is suddenly a lot harder and a lot more consequential.
The outliers — the people who use AI to do things nobody else thought to attempt — still produce distinguishable output. But they are harder to see against a background where everyone’s work looks increasingly similar. And in an organisation running trust dynamics at compressed timescales, “harder to see” translates directly to “slower to trust” — which means slower to delegate, slower to empower, and slower to build the high-trust equilibrium that makes exceptional work possible in the first place.
II. The tool outpaces the mental model
The human in the loop remains accountable. This has always been the case and AI does not change it. A surgeon is accountable for the outcome whether they use a scalpel or a robotic system. An engineer is accountable for the architecture whether they wrote every line or directed an AI coding agent. The tool does not diffuse accountability — it raises the bar for what competent use of the tool looks like.
What has changed is the rate at which the tool evolves relative to the human’s ability to understand what they are using.
In a stable tooling environment, a professional builds a mental model of their tools through experience. They learn the capabilities, the edge cases, the failure modes. This mental model is what makes accountability meaningful — you can be held accountable for the outcome because you understood, or should have understood, what the tool would do. Competence beliefs about you incorporate your demonstrated mastery of the tools you use.
AI disrupts this by changing faster than mental models can update. The model you used last month has new capabilities this month. The prompting patterns that worked last quarter produce different results this quarter. The boundary between what the tool can and cannot do shifts continuously. A professional who was genuinely competent in their use of AI three months ago may be operating on an outdated mental model today — not through negligence, but because the tool moved.
This is what I would call the perpetual adjustment problem. The traditional trust-building cycle — demonstrate competence, accumulate reputation, earn delegation — assumes that the capability being demonstrated is relatively stable. When it is not, the cycle cannot complete. You demonstrate competence with version N of the tool. By the time the observation translates into a stable competence belief in your colleagues, you are operating on version N+2, and the demonstration is no longer current.
The disposition and opportunity beliefs described in the previous article [1] face the same instability. Disposition beliefs — does this person act in my interest? — depend partly on predictability. When someone’s capabilities and methods shift continuously, their behaviour becomes harder to predict, which slows disposition belief formation even when their intentions are constant. Opportunity beliefs — do structural conditions allow them to act? — shift as the organisational norms around AI use evolve, often faster than the norms can be articulated.
The result is not a trust breakdown. The belief formation mechanisms work as they always have. What fails is the convergence condition. Trust requires that beliefs stabilise long enough to become actionable. When the environment shifts faster than beliefs can settle, the organisation exists in a state of perpetual belief instability — always updating, never converging, never reaching the stable trust state that enables efficient collaboration.
A colleague — a CTO navigating exactly these dynamics in a large engineering organisation — described what he is observing as “perpetual cognitive load.” That framing is precise. It captures the felt experience of what the theoretical framework describes: the mental cost of continuously re-evaluating trust judgements that never settle. Every delegation decision, every code review, every architectural discussion requires a fresh assessment because the evidence base has shifted since the last one. The cognitive load is not from the work itself — it is from the trust recalculation that the work now requires.
III. The mean absorbs the signal
There is a third dynamic that interacts with the first two, and it operates at the population level rather than the individual level.
In any community, trust formation depends partly on differentiation. You trust specific people for specific things because you have evidence that distinguishes them from others. The doctor you trust with a difficult diagnosis. The engineer you trust with a critical system. The manager you trust with a difficult conversation. These trust relationships are built on observed differentiation — this person is notably better at this thing than the available alternatives.
AI compresses differentiation. When the same tool is available to everyone, the outputs converge. Documents look more similar. Code looks more similar. Analyses look more similar. The surface-level quality rises uniformly, which is good for the organisation’s baseline output but corrosive to the differentiation signals that trust formation depends on.
This is Galton’s regression towards mediocrity [10], applied to AI-augmented organisations. Galton observed in 1886 that extreme traits regress toward the population mean across generations. Secrist documented the same pattern in business performance [11]. Recent work has extended this principle directly to large language models, arguing that because AI systems optimise for statistically probable patterns, their outputs converge on the safe and predictable — a dynamic termed Galton’s Law of Mediocrity [12]. When these tools are embedded in organisational workflows, they import this regression into the output of every team member who uses them. The floor rises. The ceiling does not rise at the same rate. The distribution compresses. The outliers — those who use AI to attempt things nobody else would try, who combine domain expertise with tool mastery to produce genuinely novel work — still exist. And over time, they will separate from the mean decisively, because AI amplifies the gap between “using the tool competently” and “using the tool creatively.” But in the medium term, during the adoption phase that most organisations are currently navigating, the compression effect dominates.
The trust consequence at community scale is significant. In Putnam’s framework, social capital depends on differentiated trust relationships — knowing who to turn to for what. When differentiation compresses, these relationships become harder to form and maintain. The community does not lose trust. It loses the granularity of trust — the specific, targeted confidence in specific people that makes complex coordination possible without excessive oversight.
This connects directly to the compression argument from the previous article. Organisations run Putnam dynamics at sprint velocity. When differentiation compresses at the same time that belief convergence slows (because competence signals are finer-grained and the tool keeps evolving), the high-trust equilibrium becomes harder to reach — not because the basin disappears, but because the path to it becomes longer while the dynamics run faster.
IV. The compound effect: perpetual instability
These three dynamics interact. Competence signals compress to the mean, making differentiation harder to observe. The tool evolves faster than mental models can track, preventing beliefs from stabilising. And the loss of differentiation granularity weakens the specific trust relationships that complex coordination depends on.
The compound effect is not a vicious cycle in the Putnam sense — it is not defection breeding distrust. It is something more disorienting: noisy oscillation. Trust decisions are being made, revised, overridden, and recalibrated in rapid cycles. The organisation is not frozen between equilibria — it is churning through them, never settling long enough for the stable patterns that efficient collaboration depends on to form.
What does this look like in practice? You delegate a critical decision to someone whose last three outputs were excellent — but you no longer know whether that excellence was theirs or their agent’s. You second-guess yourself. You add a review step you wouldn’t have added last year. The review itself is informed by your own AI-assisted assessment, which you’re not entirely sure you calibrated correctly. You make the trust decision anyway — you have to, the deadline doesn’t wait for epistemic certainty — but the decision carries less conviction than it would have in a stable signal environment. Next week, new information arrives that shifts your assessment again. The cycle repeats.
This is the perpetual cognitive load. Not the difficulty of any single trust decision, but the accumulation of decisions that never fully resolve. Each one consumes attention. Each one leaves a residue of uncertainty. The instinct-driven trust judgements that used to operate in the background — the automatic, low-cost assessments that free up attention for the actual work — now demand conscious processing. You are always on the lookout. Always updating your assumptions. Always aware that the basis for your last assessment may already be stale.
Signal quality is not only degrading. New signals are forming alongside the old ones — prompt traces, reasoning logs, tool-usage patterns, reproducibility across runs, latency-quality tradeoffs, consistency across agent workflows. These are real, they did not exist five years ago, and they are becoming legible to the people who learn to read them. But these signals share a property that the older ones did not: each is a technological artefact, generated by tools that are themselves non-stationary. A prompt trace from last quarter's model carries different information than a prompt trace from this quarter's, even for the same task. Reproducibility metrics depend on the specific version. Tool-usage patterns shift as the tools shift. The signal generation process itself is moving, and moving faster than the evaluators can learn to read it. The gain in signal availability is offset by a loss in signal stability. This is not a transition from one evaluation regime to another. It is sustained non-stationarity, where the rate at which new evaluation signals appear and obsolete themselves is itself increasing. The new signals do not solve the trust-evaluation problem. They multiply the surfaces on which it must run, each of which is moving.
Consider what a team looks like in practice. In previous work [9] I described the emerging unit of engineering production as two humans and four AI agents — two people directing multiple specialised AI systems for analysis, coding, validation, and execution. This is not a theoretical projection. It is how work increasingly happens. But it changes the trust dynamics fundamentally.
In a traditional team of eight engineers, trust forms across a dense network of relationships. Each person observes the others’ work, builds competence beliefs through repeated interaction, and develops the disposition assessments that enable delegation without oversight. The community sustains itself through this network density.
In the 2+4 model, two humans interact primarily with their AI agents, not with each other. Ideas that were previously bounced between colleagues are now bounced off AI. Analysis that was previously peer-reviewed is now AI-validated. The human-to-human interaction surface shrinks dramatically. Each person operates in a semi-autonomous loop with their own agent stack, producing outputs that arrive fully formed to the other.
This is where the trust formation problem becomes acute. The two humans still need to trust each other — to delegate, to divide responsibility, to make joint decisions under uncertainty. But the evidence they need to form those trust beliefs is increasingly mediated through AI. They see each other’s outputs, not each other’s thinking. The process that produced the output — the reasoning, the dead ends, the judgement calls — is invisible, absorbed into the human-agent interaction.
The risk is structural isolation. Not the deliberate information hoarding of a low-trust organisation, but the emergent siloisation of a workflow where each person’s primary collaborator is an AI that doesn’t share context laterally. Two people can work in adjacent AI-mediated loops for weeks, producing excellent individual output, while their trust relationship — the beliefs about each other’s competence, disposition, and reliability — remains unformed or stale.
But the answer is not to retreat to the old way. Deep peer interaction — slow deliberation, synchronous scrutiny, face-to-face challenge — produces higher-quality trust signals. It also does not scale and cannot match the pace that AI-augmented work demands. The AI-mediated loop is sustainable and productive but generates thinner trust signals. Neither mode wins on its own. Both are true simultaneously.
The emerging competence — and the one that will increasingly define the outliers — is calibration: knowing when the AI loop is sufficient and when genuine human scrutiny is required. The person who moves fast through routine work in their agent loop and slows down deliberately for the high-stakes architectural decision, the ambiguous requirement, the interpersonal tension that no AI can navigate — that person is exercising a judgement that is almost invisible from the outside. The output looks the same either way. But the quality of the trust decisions embedded in that output is fundamentally different.
This is what makes the identification problem from the previous section acute. The outlier you need to find is not the person who produces the best output — AI compresses that signal. It is the person who exercises the best judgement about when to trust the AI loop and when not to. The difference is real but time-delayed. Two people produce equally impressive deliverables. Six months later, one person’s work holds up — the strategy anticipated a risk nobody else saw, the decision accounted for a dynamic nobody else considered, the design proved resilient when conditions changed. The other’s did not. The calibration shows up in the resilience of the outcome, not the polish of the output.
The difficulty is that organisations make trust allocation decisions now, based on the immediate signal, not six months from now when the outcomes diverge. By the time the difference becomes visible, authority and resources have already been allocated based on the surface — which looked identical.
This raises a harder question that the article does not resolve. If old signals compress and new signals are themselves non-stationary, the only candidate evaluation surface left is long-term reputation accumulated over time. But sprint velocity is precisely what removes the runway reputation needs to stabilise. Humans do not stop allocating leadership and competence in the absence of clean signal — social systems fill these gaps regardless, because someone has to be trusted with the next decision. So reputation will form. What is unclear is whether the reputation that forms in this environment tracks reality or drifts from it. In stable signal environments, reputation converges toward truth over time because outcome feedback corrects early misallocations. In non-stationary environments with compressed signals, reputation may form just as quickly but without the corrective loop that aligns it with actual competence. The result is reputation networks that are confidently held and self-reinforcing — once someone is treated as the trusted authority on a domain, that treatment itself becomes a signal others use — but which may be decoupled from the underlying truth they are supposed to track. Whether the reputation networks that emerge in AI-augmented organisations will be more or less accurate than those they replace is an open empirical question. The article cannot answer it. It can only flag that the question is now open in a way it was not before.
This is more disorienting than a straightforward trust collapse. In a low-trust equilibrium, at least the rules are clear. In noisy oscillation, the rules keep shifting. You cannot develop a stable strategy because the evidence base for your trust judgements churns faster than you can act on it — and the people around you, working in their own AI-mediated loops, are equally unable to provide the stable signals you need.
V. A design problem, not a behavioural one
The instinctive response to this dynamic is behavioural prescription: communicate more, be more transparent, build psychological safety. These operate at individual speed in an environment moving at technological speed. By the time a leader has demonstrated consistent trustworthy behaviour through a full cycle of AI-driven change, the next cycle has already begun.
There is a stronger version of this argument that the article has so far understated. Perpetual cognitive load is not only a felt phenomenon — it is bounded by biology. Established research on working memory, attention residue, and the costs of context-switching places hard ceilings on how much trust recalibration any individual can sustain before judgement quality degrades. The rate of environmental change has no equivalent ceiling. This means the mismatch between AI evolution and human adaptation is not a gap that closes through individual effort or training. It is a gap that grows, because one side has a limit and the other does not. Behavioural prescription assumes the gap is closeable. It is not. The cognitive load problem is real, biological, and increasing — and any response that does not account for the biological ceiling is asking individuals to absorb costs that compound until they collapse.
The response must be structural, and it must address the specific dynamics described above.
If the problem is that outliers are buried by convergence to the mean, the design response is to create conditions where exceptional use of AI is visible and attributable — not through surveillance, but through contexts where the work is seen. Showcases, not scorecards. Attribution that distinguishes what the human contributed from what the tool provided, framed as quality infrastructure rather than monitoring. Identifying outliers in a compressed distribution is itself a new organisational competence — one that most organisations have not yet developed and that existing performance management systems are not designed for.
If the problem is that the tool outpaces mental models, the design response is to reward adaptation speed rather than current competence. The person who was excellent last quarter with last quarter’s tools may be average now. What matters is the rate at which someone integrates new capabilities and produces novel output — and most organisations do not measure this.
If the problem is that differentiation compresses, the design response is evolutionary rather than prescriptive. Let the outliers demonstrate what is possible. Make their methods visible. Let social learning raise the floor. This is Putnam’s virtuous cycle applied to capability growth — one person’s exceptional use raises expectations, which raises everyone’s floor, which creates space for the next outlier to push further.
But none of this works in a low-trust environment. In a low-trust organisation, outliers are threats. Their methods are hoarded rather than shared. Adaptation is punished because it disrupts established hierarchies. The trust infrastructure from the previous article is the prerequisite. You need the high-trust equilibrium — or at least a trajectory toward it — before you can build the adaptation mechanism on top of it.
These are design hypotheses, not proven interventions. Whether outlier visibility, adaptation reward, or evolutionary adoption actually move the needle on trust belief convergence in AI-augmented organisations is an empirical question — and the ATDP framework, developed for AI agent design, may provide a method for answering it.
VI. A framework for testing
In recent work [2], I proposed the AI Trust Design Process — a framework that models design choices as actions in a Markov Decision Process, where the state is a population’s trust belief distribution and the reward is social capital change. The framework was developed for AI agent design: how do you design an AI agent so that the population’s trust beliefs move in a direction that builds social capital rather than eroding it?
The proposition I want to advance — carefully, as a hypothesis rather than a claim — is that the same framework generalises to organisational environment design. The design variable is not an AI agent’s behaviour. It is the organisational environment’s properties: visibility mechanisms, reward structures, adaptation cadences, attribution systems. The population is the team or organisation. The trust belief distribution is the same Castelfranchi decomposition. The reward function is the same social capital metric.
If this generalisation holds, then ATDP provides not a prescription but a method — a way to empirically discover which design properties of an organisational environment move the trust belief distribution toward convergence even as AI accelerates change around it. The specific parameters described in the previous section — outlier visibility, adaptation reward, evolutionary adoption — become testable hypotheses within that framework, not intuitions defended by argument.
This is a proposed generalisation, not a validated one. The ATDP framework has been developed for the context of AI agent design and is currently being tested experimentally. Applying it to organisational environments is a new claim that expands its scope significantly. The machinery is compatible — the state space, the action space, and the reward function all have natural organisational analogues. But compatibility is not validation. The weight matrix that maps design properties to belief distribution changes would need to be learned in an organisational context, where the design space is vastly larger and the experimental controls are weaker than in a laboratory setting.
What I am confident of is this: the problem described in this article — noisy trust allocation driven by AI-accelerated environmental change — is a design problem, not a behavioural one. The tools for addressing design problems empirically exist. Whether the specific framework I have proposed is the right tool for this particular design problem is an open question. But the question itself is the right one to ask.
VII. Conclusions
Three claims this article advances:
One. AI changes how trust decisions are made in organisations through three specific mechanisms: it compresses competence signals toward the mean (Galton regression applied to AI-augmented work); it shifts the competence target faster than assessment can track (perpetual adjustment); and it restructures teams around human-AI collaboration loops that reduce the human-to-human interaction surface where trust evidence is generated.
Two. The combined effect is not a trust collapse but a noisier trust allocation — decisions made at full speed on degraded inputs, producing what one practitioner aptly named perpetual cognitive load. The differentiating skill in this environment is calibration: knowing when the AI loop is sufficient and when human scrutiny is required. This skill is real, it produces outcomes that diverge meaningfully over time, but it is almost invisible from the outside because the surface output looks identical to less considered work.
Three. This is a design problem, not a behavioural one. The instinct to prescribe better individual behaviour fails because behavioural interventions operate at individual speed in an environment moving at technological speed. Structural responses — designing for outlier visibility, rewarding adaptation, making the calibration skill legible — operate faster but require frameworks that can be tested empirically. The ATDP framework, originally developed for AI agent design, may provide such a method when generalised to organisational environments. That generalisation is a hypothesis, not a validated approach.
The argument is meant to make the trust dynamics in AI-augmented organisations legible enough that they can be designed for, rather than left to drift.
VIII. Known limits and future work
The compression-versus-stretching dynamics are asserted, not measured. This article argues that AI compresses the middle of the competence distribution while stretching the tails. This is directionally consistent with the Galton regression literature and with the recent application to LLM outputs [12]. But the specific claim that organisational competence distributions respond this way to AI adoption has not been empirically measured. Longitudinal studies tracking competence signal variance within teams before and after AI adoption would be the right test. The distribution shape matters — if AI uniformly compresses rather than differentially affecting the middle and tails, the identification problem I describe is less severe than argued.
The 2+4 team model is emerging, not established. The shift from human-dense teams to small human teams directing AI agent stacks is observable in some engineering contexts but is not yet widespread enough to generalise from. Whether the trust dynamics I describe — emergent siloisation, process invisibility, the calibration skill as differentiator — apply equally in non-engineering domains is an open question. The argument is grounded in mechanism (reduced human-to-human interaction surface) which should generalise, but the specific manifestations may differ substantially across domains.
The calibration skill has not been formally defined. I argue that the emerging outlier competence is knowing when to trust the AI loop and when to apply human scrutiny. This is intuitively recognisable but theoretically imprecise. What exactly constitutes calibration skill? How would you measure it? How does it relate to existing constructs like metacognition or professional judgement? These questions need answers before the concept can support empirical work. At present it is an observation, not a construct.
The ATDP generalisation to organisational environments is untested. The proposition that the ATDP framework extends from AI agent design to organisational environment design is grounded in structural compatibility — the state space, action space, and reward function all have natural analogues. But the design space for organisational environments is vastly larger, the experimental controls are weaker, and the feedback loops are slower than in AI agent experiments. Two specific tensions deserve acknowledgement. First, the Markov property — the assumption that the future state depends only on the current state and the action taken — is a simplification when applied to human trust. Two teams with identical current trust distributions but different histories (one recovering from a failure, one declining from a betrayal) may respond very differently to the same design intervention. The current state representation may encode enough history through the belief values themselves to preserve the property, but this is an assumption that needs testing, not a given. Second, there is a speed paradox: structural interventions operate faster than behavioural ones, but still slower than AI evolution. Organisational design moves at institutional speed — quarters at best. AI tooling moves at weeks. If the environment shifts three times between designing an intervention and measuring its effect, the framework risks solving yesterday’s problem. The design method must account for this latency or it becomes retrospective rather than adaptive.
The macro-level social capital trajectory is undecided. This article deliberately avoids claiming that AI will increase or decrease social capital. Technology has broadly raised human wellbeing, and the established correlation between social capital and economic outcomes suggests an upward trajectory. But corrosion is visible — in the erosion of shared epistemic foundations [9], in the fragmentation of communities into echo chambers, in the exploitation of trust defaults at scale. The macro trajectory — whether the broad upward trend in wellbeing and social capital continues or whether the visible corrosion [9] accelerates — is shaped by forces well beyond any single organisation’s design choices. But design choices are what we have. This article is an attempt to make the ones that matter legible.
References
[1] Prifti, Y. (2026). Trust at Organisational Speed. Weighted Thoughts.
[2] Prifti, Y. (2026). A Markov Framework for AI Trust and Societal Outcomes. SSRN. doi:10.2139/ssrn.6390618
[3] Castelfranchi, C. & Falcone, R. (2010). Trust Theory: A Socio-Cognitive and Computational Model. Wiley.
[4] Putnam, R.D. (1993). Making Democracy Work: Civic Traditions in Modern Italy. Princeton University Press.
[5] Axelrod, R. (1984). The Evolution of Cooperation. Basic Books.
[6] Levine, T.R. (2019). Duped: Truth-Default Theory and the Social Science of Lying and Deception. University of Alabama Press.
[7] De Meo, P., Prifti, Y. & Provetti, A. (2025). Trust Models Go to the Web. ACM Transactions on the Web, 19(2). doi:10.1145/3715882
[8] Edmondson, A.C. (2018). The Fearless Organization: Creating Psychological Safety in the Workplace. Wiley.
[9] Prifti, Y. (2026). The Falling Cost of Lying and the Active Role of AI Agents. Weighted Thoughts.
[10] Galton, F. (1886). Regression Towards Mediocrity in Hereditary Stature. Journal of the Anthropological Institute of Great Britain and Ireland, 15, 246–263.
[11] Secrist, H. (1933). The Triumph of Mediocrity in Business. Northwestern University.
[12] Krashniak, A. et al. (2025). Galton’s Law of Mediocrity: Why Large Language Models Regress to the Mean and Fail at Creativity in Advertising. arXiv:2509.25767.
[13] Pluchino, A., Rapisarda, A. & Garofalo, C. (2010). The Peter Principle Revisited: A Computational Study. Physica A, 389(3), 467–472. doi:10.1016/j.physa.2009.09.045
[14] Prifti, Y. (2023). The Emergence of Interpersonal and Social Trust in Online Interactions. PhD thesis, Birkbeck, University of London.
[15] Sharma, M. et al. (2024). Towards Understanding Sycophancy in Language Models. ICLR 2024.