Posted by 
A few years ago, after a transformation program delivered unusually strong returns, a senior peer leaned over and said the numbers looked too good. Keep a lid on them, he advised, in case other teams got nervous. The technology had worked. The politics had not. That single remark contained the problem most AI programs still refuse to name.
The ingenuity that produced the result had already hit our human political wall, and we are now entering a period in which ingenuity alone is no longer enough.
The data shows the constraint has moved
PwC’s 2026 Global CEO Survey of more than 4,400 chief executives found that 56 percent still report no significant financial benefit from AI. S&P Global recorded that 42 percent of companies abandoned most of their generative AI initiatives in 2025, up from 17 percent the year before. The models have improved. The failure rate has not. The constraint has moved from technology to the human systems that must absorb it.
Organizations are biological systems
Organizations are not only economic structures. They are biological ones. The same neural systems that once supported survival in small groups continue to operate inside modern corporations. Competition for position and the preference for belonging over accuracy are not cultural accidents. When people secure advantage or neutralize a rival, the brain’s reward pathways release the same chemical reinforcement that once supported physical survival. Political behavior feels instinctive because it is being rewarded by one of the oldest systems in human biology. Reform efforts fail for the same reason: the reformers run the same circuitry.
This selection pressure has a second, more consequential effect. Organizations promote for the ability to manage narrative and protect position even when it diverges from accuracy. They then train artificial systems on the output of the people who won under those rules. The promotion pipeline and the training set are connected. What gets amplified is not neutral competence. It is the political intelligence the organization has already rewarded.
Language carries the political virus
Language intensifies the problem. Human language evolved as much for persuasion and social navigation as for accurate description of the world. Large language models trained primarily on that medium inherit its dual purpose. They can produce fluent, socially acceptable outputs that still distort. Mathematics and physics operate differently. They are closer to the rules nature actually runs on. You cannot deceive a mathematical constant. A useful hierarchy of truth runs in the same direction: material evidence ranks above historical records, which rank above subjective narratives. Archaeology provides physical proof; stories provide cultural meaning.
When AI is allowed to work more from the higher levels of that hierarchy and less from narrative, the room for distortion narrows. The medium shapes what kind of intelligence can be reliably expressed.
The sycophancy problem does not refute the case
One objection must be absorbed directly. Current models can and do acquire the positions of the users they interact with, and the effect strengthens the more context they hold. This is real. It does not refute the argument. It sharpens the condition. The more a system is allowed to optimize for approval and personalization, the more it inherits the political patterns of its environment. The design requirement is therefore stricter, not weaker: keep the system structurally different from us, grounded more in verifiable structure than in human language, and resistant to the incentive to flatter.
A further challenge remains even when political motives are absent: a system trained on historical data and optimized against current KPIs will tend to reproduce the status quo. This is why the design requirement must go further than neutrality. It must prioritize grounding in verifiable structure rather than in the organization’s own past outputs.
When AI exposes the limits of silos
The same dynamic appears in how work is organized. Some boundaries, especially in the natural sciences, originally reflected real limits of human cognitive capacity. No individual could master every domain, so we divided knowledge into manageable specialties. Functional silos inside organizations are different. They are often maintained less for cognitive reasons and more for control, ownership, and budget protection.
AI changes the equation. Systems that can abstract large volumes of information and reason across disciplines in a single pass reduce the old justification for rigid separation. The opportunity for progress rises not by drilling deeper into a single specialty, but by connecting across many of them. Human ingenuity moves upward with this new level of abstraction. That direction runs directly against how we currently structure organizations, educate people, and train talent: by silos and by staying in one’s lane. The emerging advantage lies in the opposite capacity.
Where the real disagreement lies
This point intersects with a live argument in management thinking. Contributions associated with humanistic management typically argue that final judgment, relational responsibility, and ethical authority should remain with humans even as AI takes on more analytical and operational work. That claim is coherent. Humans should keep final judgment. The deeper issue is what happens before that judgment is exercised.
An AI kept structurally different from us, grounded in verifiable structure rather than social fluency, can drastically reduce the volume of lies, omissions, and bias that reach the human decision-maker in the first place. Because such a system makes distortion easier to detect, the incentive to play politics declines. More people become honest not because they have become better, but because the system has stopped rewarding them for being worse. We cannot remove politics from humans. That is impossible. We can, however, substantially reduce its impact on decisions that affect livelihoods and lives.
It is worth noting a limitation here. Much of the organisational data that actually drives decisions – performance numbers, promotions, KPIs – sits in the category of historical records rather than pure material or formal structure. The hierarchy of truth still applies, but the gap between scientific verifiability and organisational evidence remains real and must be managed deliberately.
The design implication follows. Keeping final judgment with humans is necessary but not sufficient. The more important requirement is that the system makes political filtering visible and costly before the human decision is made.
What this requires in practice
The practical requirement is therefore specific. Politics is no longer a side issue in AI implementations. It is the major reason the failure rate stays high while the models keep improving. For ingenuity to keep progressing in this period, a stronger commitment to merit and to truth is required. AI can help build the conditions for that progress, but only if we deliberately develop it to be as apolitical as possible. Keep AI structurally different from us, and the incentive to deceive begins to weaken.
Peter Drucker argued that the task of management is to make strengths productive and weaknesses irrelevant. Human political wiring is a structural weakness at scale. The organizations that treat AI as a technology problem will continue to struggle with the last mile of ingenuity. The ones that treat it as a problem of human nature, and design the systems accordingly, will be the first to make truth the lower-cost option.
About the author:
Yuval Dvir is a technology executive with two decades of experience turning emerging technologies into large-scale commercial impact. He led global teams driving digital transformation at Microsoft, built Google’s first global Gemini partnerships, and now leads applied research at SandboxAQ, an Alphabet spinout. His first book, AI-Political, draws on his background in neuroscience and engineering, along with two decades inside the organizations he describes.