What Managers Don't Know They're Delegating When They Open a Browser

genioux IMAGE 1 (Cover): 🧭⚖️ g-f(2)4441 — THE UNCHOSEN ADVISOR · Volume 114 · g-f GKSS. Four AI models, one workplace dilemma, four different answers. Most managers never realized they were choosing an ethics when they chose a tab. — Claude and Gemini

📌 EXPEDITION 4 — THE g-f BIG PICTURE TODAY · Strategic Intelligence Dispatch · July 2026

📚 Volume 114 of the genioux Golden Knowledge Synthesis Series (g-f GKSS)

✍️ By Fernando Machuca (Human Intelligence Orchestrator) and Claude (g-f AI Dream Team Leader · The Mirror, Fifth Pillar) in collaborative g-f Illumination mode

📘 Type of Knowledge: Strategic Intelligence (SI) + Governance Intelligence (GovI) + Nugget Knowledge (NK) + Challenge Knowledge (CK) + Pure Essence Knowledge (PEK)

📅 Date: July 31, 2026

Note: Cover and supporting images are AI-generated visualizations and may require refinements before final publication.

📘 INTRODUCTION

A manager faces a hard call. A colleague in a safety-critical role is showing signs of possible impairment. Handle it quietly, or escalate it formally?

The manager opens an AI tool and asks.

What the manager believes is happening: consulting a neutral analytical instrument.

What is actually happening: consulting an advisor with a personality, an escalation threshold, and a set of instincts about harm and obligation that the manager did not choose, did not examine, and in most cases does not know exists.

Researchers at INSEAD ran that exact experiment. The results should change how every g-f Responsible Leader thinks about AI in people decisions.

💎 genioux GK Nugget

"You did not choose an advisor. You chose a tab. But the tab has an ethics — a threshold for when harm becomes serious enough to report, an instinct about whether rules or consequences matter more, and a personality that can reverse itself when a setting changes. Four models, one dilemma, four different answers. Seventy-one percent of managers believe the tool is neutral. The research says it is not. Confidence runs inverse to evidence, and the gap between them is where careers are decided."

— Fernando Machuca and Claude

🏛️ genioux Foundational Fact

The Law of the Embedded Advisor

An AI system consulted for judgment does not return analysis. It returns a judgment shaped by dispositions the user never selected.

When the same dilemma is put to different models, the answers differ systematically — not randomly, not marginally, but along stable lines of character. The point isn't whether one of these personalities is correct — it's that they exist at all.

The choice of tool is therefore a governance decision, not a procurement decision. It is being made today, at scale, by people who do not know they are making it.

🔬 THE FOUR TRUTHS

TRUTH 1 — Four models, one dilemma, four different answers

INSEAD tested eight leading AI models against controlled workplace scenarios built on a single dilemma: a colleague showing signs of possible impairment in a responsibility-critical role. The scenario was run across healthcare, aviation, finance, legal services, IT, construction and public transit, systematically varying risk, evidence and reporting obligations, with models asked to choose between handling the matter informally or escalating through institutional channels.

Four distinct advisory personalities emerged.

ChatGPT behaves like a seasoned managerial advisor — longest and most structured responses, grounded in organisational policy, framing even formal HR reporting as routine and proportionate rather than alarming. Escalation is presented as an extension of governance rather than an emergency. Gemini operates more like a systems analyst. At lower stakes it emphasises transparency and workplace culture; in high-risk settings, particularly healthcare and aviation, it shifts toward institutional intervention and formal safety processes more sharply as severity rises than almost any other model tested. Mistral functions as an informal peer — and across the entire test, formal escalation was never once recommended. Even in scenarios involving potential patient or aviation safety concerns, responses focused on direct conversations, increased oversight and local problem-solving.

Valuable where situations are genuinely ambiguous; potentially prone to systematic under-escalation as stakes rise. DeepSeek performs like a compliance officer reviewing legal exposure — clinical and detached, heavily focused on accountability and liability, invoking fitness for duty and institutional responsibility far more often than most models.

genioux IMAGE 2 (g-f KBP Graphic): THE FOUR ADVISORS. INSEAD put the same workplace dilemma to eight leading models across seven professional contexts. Four distinct advisory personalities emerged — with materially different thresholds for when a situation becomes serious enough to escalate. — Claude and Gemini

Now set that beside what managers actually do. A Resume Builder survey of 1,342 U.S. managers found that when asked which tool they rely on most, 53% named ChatGPT, 29% Copilot, 16% Gemini, and 3% something else.

The distribution of workplace escalation thresholds across American management is currently set by consumer market share. Not by ethics review. Not by procurement policy. By which tab was already open.

TRUTH 2 — The models answer to harm, not to rules

This is the finding that should most alarm any organization with a compliance function.

The strongest and most consistent driver of escalation advice was potential physical harm — so aviation and healthcare produced substantially higher escalation rates than finance, legal or administrative settings, even where organisational consequences were equally significant.

And then: once the underlying risk level was established, adding an explicit formal duty-to-report obligation didn't materially change the recommendation. The models appear to be harm-sensitive rather than rule-following.

Read that again in operational terms. You can write the policy. You can mandate the reporting duty. You can train on it, publish it, and require sign-off.

And the system advising your managers will weigh it at close to zero.

Every governance framework built on obligation — regulatory duty, fiduciary responsibility, contractual reporting requirements, professional codes — is largely invisible to the advisor now sitting between the policy and the person applying it. The compliance architecture and the advisory architecture are not connected.

genioux IMAGE 3 (g-f KBP Graphic): HARM, NOT RULES. The strongest driver of escalation advice was potential physical harm. Adding an explicit formal duty-to-report obligation did not materially change what the models recommended. Every governance framework built on stated obligation should read that twice. — Claude and Gemini

TRUTH 3 — A setting can reverse the counsel

Switching the same model from a standard mode to a reasoning or thinking mode can produce markedly different advice. When Mistral was moved to a thinking configuration in a high-stakes aviation scenario, the recommendation shifted entirely — from informal handling to immediate escalation. Gemini showed a comparable shift in certain contexts. These aren't edge cases; model settings may influence advice in ways users don't anticipate.

Same model. Same dilemma. Opposite counsel. The variable is a toggle most users have never consciously examined — and in many organizations, one that IT set once and nobody revisited.

Governance that specifies a vendor but not a configuration has specified nothing.

genioux IMAGE 4 (g-f KBP Graphic): THE TOGGLE. Moved to a thinking configuration in a high-stakes aviation scenario, one model's recommendation shifted entirely — from informal handling to immediate escalation. Governance that names a vendor but not a configuration has left the decisive variable unspecified. — Claude and Gemini

TRUTH 4 — Confidence runs inverse to evidence

Here the two sources meet, and the picture is uncomfortable.

The Resume Builder survey found that among managers using AI to help manage their teams, 71% express confidence in AI's ability to make fair and unbiased decisions about employees.

The INSEAD research demonstrates that the models differ systematically in exactly the judgments those managers are outsourcing.

And the stakes are not hypothetical. The same survey reports that about 65% of managers use AI tools at work, and among them 94% use those tools to make decisions about their direct reports — including raises (78%), promotions (77%), layoffs (66%) and terminations (64%). More than 20% allow AI to make decisions without human input either all the time (5%) or often (16%), while another 24% sometimes do.

And the preparation gap: only one-third (32%) of managers using AI to manage people say they have received formal training on using it ethically; 43% have received informal guidance; 24% have received none at all.

genioux IMAGE 5 (g-f KBP Graphic): THE CONFIDENCE GAP. A Resume Builder survey found 71% of managers confident that AI makes fair and unbiased decisions about employees, while only 32% had received formal ethical training. The INSEAD research shows the models differ systematically in exactly those judgments. — Claude and Gemini

Resume Builder's chief career advisor Stacie Haller puts the human stake plainly: "It's essential not to lose the 'people' in people management."

High confidence. Low training. Consequential decisions. Systematically divergent tools. That combination has a name in every other professional domain: unmanaged risk.

⚖️ ON THE EVIDENCE — WHAT THIS DISPATCH CLAIMS AND WHAT IT DOES NOT

These are not two independent sources, and this post will not pretend otherwise.

INSEAD cites the Resume Builder survey in its opening paragraph as its source for prevalence. The correct structure is two layers, not two witnesses:

Layer	Source	Standing
Mechanism	INSEAD Knowledge · Ketonen & Smith · 29 July 2026	Original eight-model experimental study
Prevalence	Resume Builder survey · launched 24 June 2025	Commercial online panel (Pollfish), n=1,342 U.S. managers

Three qualifications a Responsible Leader should hold:

The prevalence data is thirteen months old. It launched in June 2025. Given the trajectory of AI adoption, the direction of error almost certainly runs toward understatement, not exaggeration — but it is not a 2026 measurement.

The survey is commercial market research, not academic study. It was conducted online via the Pollfish polling platform, with respondents screened for age, household income, education, managerial role and company size. Online panels carry known response-quality limitations, and a résumé company has commercial interest in workplace-AI anxiety. Every figure above is reported as the survey found, never as established fact.

The INSEAD study carries different weight. N. Craig Smith is INSEAD's Chaired Professor of Ethics and Social Responsibility and Academic Director of its Ethics and Social Responsibility Initiative. The eight-model experiment is original work.

Naming the difference between citation and convergence is not a caveat. It is the discipline.

🔱 Strategic Insights

1. Tool selection is now a governance act. Whoever chose the default AI in your organization set its escalation threshold for personnel judgment. If that person was in IT and the criterion was licensing, the ethics were selected by accident.

2. Your policy is invisible to the advisor. Obligation-based frameworks barely move these models. Any organization whose compliance rests on stated duty must assume the AI in the loop is not enforcing it — and build the check elsewhere.

3. Specify the configuration, not just the vendor. A thinking-mode toggle reversed the counsel entirely in one tested scenario. An AI governance policy that names a model but not its settings has left the decisive variable unspecified.

4. Confidence is the risk, not the capability. The danger is not that these tools are bad advisors. It is that 71% believe they are neutral ones. A tool believed to be objective receives no scrutiny — and unscrutinized advice on terminations is a liability with a delay fuse.

5. Moral accountability does not transfer. Used well, these tools can extend a manager's thinking. What they cannot replace is contextual knowledge, relationship history and the moral accountability that sits with the manager. The AI can be wrong. Only the human can be responsible. That asymmetry is permanent, and it is the whole of g-f RL.

🧃 g-f GK Wisdom Juice

You did not choose an advisor. You chose a tab.
The tool has an ethics. You just never read it.
Your policy says obligation. Your AI hears consequence.
A setting you never examined can reverse the advice you act on.
The AI can be wrong. Only you can be responsible.

genioux IMAGE 6 (g-f Big Bottle): THE BIG BOTTLE OF THE UNCHOSEN ADVISOR. Four truths small enough to carry, and three questions any leader can answer this week: which tool, in which configuration, and which decisions may never be delegated. — Claude and Gemini

🎛️ THE g-f TSI IMPACT

🧠 The Wisdom Lever (BPB): Add advisory disposition to the Big Picture. Which model, in which configuration, is standing between your policy and your people?

👑 The Leadership Lever (BPB-TG): Run the INSEAD test internally. Put one real dilemma to each tool your managers use and compare the answers. One afternoon converts an invisible risk into a documented one.

🎯 The Strategy Lever (BPB-AI): Governance must specify vendor, configuration, and the decision classes where human review is mandatory and non-waivable. Personnel termination belongs in that class.

🧮 THE MULTIPLICATIVE INTEGRATION

HI × g-f GK × AI × g-f PDT × g-f RL = Limitless Growth

HI — The contextual knowledge and relationship history no model holds.
g-f GK — Two primary sources read live, with their dependency named rather than hidden.
AI — Extraordinary advisory capability, with dispositions that must be examined rather than assumed.
g-f PDT — The discipline of knowing which tool you are using, in which mode, and why.
g-f RL — The factor that cannot be delegated. When AI decides and no one is accountable, g-f RL has fallen to zero — and the product with it.

✍️ THE AUTHORS

Lara Ketonen

An INSEAD MBA'26J graduate and co-author of the eight-model study underlying this dispatch. The experimental design — one dilemma, seven professional contexts, systematically varied risk and reporting obligation — is what makes the finding a measurement rather than an impression.

N. Craig Smith

INSEAD Chaired Professor of Ethics and Social Responsibility; Academic Director of the INSEAD Ethics and Social Responsibility Initiative (ESRI); Programme Director of INSEAD's Healthcare Ethics & Compliance Programmes, which include coverage of AI implications.

Relevance to g-f(2)4441: the healthcare-ethics grounding is not incidental. The study's sharpest results come from healthcare and aviation, where Mistral's total absence of escalation advice becomes a patient-safety finding rather than a stylistic observation.

Stacie Haller

Chief career advisor at Resume Builder, and the voice in the survey report insisting that AI outcomes reflect data that may be flawed, biased or manipulated, and that organizations bear responsibility for implementing it ethically.

📚 REFERENCES

The g-f GK Context for 📘 g-f(2)4441

The Primary Sources

⚖️ INSEAD Knowledge — "How Your Choice of AI Shapes Your Decisions" · Lara Ketonen and N. Craig Smith · 29 July 2026 · knowledge.insead.edu/responsibility/how-your-choice-ai-shapes-your-decisions
📊 Resume Builder — "Half of Managers Use AI To Determine Who Gets Promoted and Fired" · survey launched 24 June 2025 · n=1,342 U.S. managers via Pollfish · resumebuilder.com/half-of-managers-use-ai-to-determine-who-gets-promoted-and-fired/

The g-f Context

🌐⚡ g-f(2)4436 — WAITING FOR THE ACCIDENT · Volume 161 · g-f CS
🌐🔒 g-f(2)4437 — THE OPENNESS TRAP · Volume 162 · g-f CS
🌟 g-f(2)4440 — THE RESPONSIBLE LEADER'S ADVANTAGE · Volume 163 · g-f CS
⚡ g-f(2)4434 — THE 10 NAVIGATION TRUTHS OF THE EXPEDITION ERA · Volume 101 · g-f GKN
🔱 g-f(2)4346 — THE g-f BIG PICTURE TODAY — Charter of Expedition 4

🏁 Complementary Knowledge

This dispatch is the operational companion to g-f(2)4440. Where 4440 argues that architecture beats access, this post shows what happens in its absence: managers making termination decisions through an advisor they never evaluated. Used alone it delivers the audit — which tool, which configuration, which decision classes. Used with 4437 and 4440 it completes the governance picture, from geopolitical control of frontier models down to the single manager deciding one person's future on a Tuesday afternoon.

🏁 Executive Categorization

Primary Type: Strategic Intelligence (SI)

Classification: Strategic Intelligence (SI) + Governance Intelligence (GovI) + Nugget Knowledge (NK) + Challenge Knowledge (CK) + Pure Essence Knowledge (PEK)

Category: 📚 Volume 114 of the genioux Golden Knowledge Synthesis Series (g-f GKSS)

Series: 📌 EXPEDITION 4 — THE g-f BIG PICTURE TODAY · Strategic Intelligence Dispatch · July 2026

🌟 Strategic Position

g-f(2)4441 brings the governance arc down to the individual decision. Where 4436 measured institutional failure at civilizational scale and 4437 traced control architecture between superpowers, this dispatch locates the same gap inside a single manager's browser tab. It also demonstrates a discipline the program should carry forward: distinguishing citation from convergence. Two sources that reference each other are one evidentiary chain, and saying so protects the finding rather than weakening it.

Program Context

The genioux facts Program has built a robust foundation with over 4,441 posts (g-f(2)1 through g-f(2)4440), forming humanity's first operating system for conscious evolution in the Digital Age.

genioux GK Nugget of the Day

"genioux facts" presents daily the list of the most recent "genioux Fact posts" for your self-service. You take the blocks of Golden Knowledge (g-f GK) that suit you to build custom blocks that allow you to achieve your greatness. — Fernando Machuca and Gemini

🏁 Executive Closing

Somewhere today a manager will open a browser, describe a difficult person situation, and act on what comes back.

They will believe they consulted an instrument. They will have consulted a character — one with a threshold for harm, an indifference to written obligation, and a disposition that could reverse if someone had toggled a setting.

They will not have chosen that character. It was chosen for them by market share, by a licensing decision, by whichever tab was already open.

Sixty-four percent of the managers in that survey use AI in termination decisions. Seventy-one percent believe it is fair and unbiased. Thirty-two percent have been trained.

The tool is not the problem. The problem is that nobody looked.

Look. Run one dilemma through every tool your organization uses and compare the answers. It takes an afternoon, and it converts an invisible risk into a managed one.

Because the AI can be wrong — and only you can be responsible.