top of page

The Forces That Shape Decisions in the Age of AI

 

Why drift, adversarial pressure, dependency, and legitimacy increasingly shape decisions before governance even begins

 

Boards today are not only making decisions under distortion. They are also operating inside environments that pre‑shape what becomes plausible and acceptable, often before judgment even begins.

​

The decisions are already half‑made by the time governance enters the room.
 

AI systems expand capability faster than clarity, generate outputs faster than they can be verified, and embed themselves into processes faster than governance can adapt. Yet the most important shift is less visible: decisions are increasingly influenced by forces that operate continuously in the background.
 

These forces do not appear in dashboards or reports. They are not tied to a single system or initiative. Yet they shape how signals are interpreted and how choices emerge.
 

Four forces are particularly structural in this environment: Drift, Adversarial Pressure, Dependency, and Legitimacy.

They do not distort decisions directly but shape the conditions under which distortions become likely.
 

Unlike in the AI Glass Maze (1), where boards navigate decisions under distortion, these forces operate upstream. They define the environment in which decisions take form, evolve, and sometimes drift before they are even consciously made.

​

Understanding these forces does not replace governance. It changes where governance must begin.​​

1. Drift: Temporal Distortion

 

Drift is the force of time acting on systems.

​

Description

 

Systems, data, organizations, and human behavior shift over time. Models decay, data evolves, processes adapt informally, and reliance on automation increases. No single decision fails, but the system slowly drifts.

​

Core board question

 

“How will this evolve over time?”

​

Typical signals

 

· Performance degrades without a clear trigger

​

· Teams compensate informally for system behavior

​

· Increasing reliance on outputs without revalidation

​

Example: COVID‑19 model collapse

 

“Retail and logistics models trained on pre‑2020 behavior failed when consumer patterns shifted overnight.” (2)

COVID‑19 exposed Drift instantly: models built on past patterns collapsed when those patterns no longer applied.

​

Tension: Stability vs. adaptation

​

Failure mode: Slow misalignment without clear failure signals

​

System effect: Locally rational decisions become globally outdated

​

​

Board response

​

Non‑negotiable:

​

Make the time horizon explicit. “What will this look like in 12–24 months?”

​

Challenge questions:

​

· What assumptions are most likely to change?

· When do we revalidate this system?

· Where are we becoming over‑reliant on past patterns?

​

Leadership countermeasure:

 

Establish explicit revalidation cycles and force periodic assumption resets.

2. Adversarial Pressure: Security Distortion

 

Adversarial Pressure is the force of intent acting against systems.

​

Description

 

AI systems are fallible and targetable. Manipulated inputs, poisoned data, and adversarial interactions can turn plausible signals into engineered distortions.

​

Core board question

 

“How does this behave under intentional attack?”

​

Typical signals

 

· Outputs vary unexpectedly with small input changes

​

· External inputs are integrated without verification

​

· Security treated as an extension of performance, not a distinct risk

​

Example: EchoLeak vulnerability (Microsoft Copilot)

 

“A hidden prompt injection embedded inside a document was able to extract enterprise data from Copilot.” (3)

EchoLeak showed how a single hidden instruction could redirect an AI system and exfiltrate data.

​

Tension: Performance vs. resilience


Failure mode: Engineered distortion that appears legitimate


System effect: False signals are amplified and trusted as valid

​

​

Board response

​

Non‑negotiable:

​

Evaluate systems under attack conditions. “Show how this behaves when inputs are manipulated.”

​

Challenge questions:

​

· Where can inputs be influenced or compromised?

· What happens if outputs are intentionally manipulated?

· Are we testing for performance? or for resilience?

​

Leadership countermeasure:

 

Treat every system as attackable and require red‑team validation before deployment.

3. Dependency: Structural Distortion

 

Dependency is the force of structure constraining systems.

​

Description

 

AI capabilities rely on layered, often opaque dependencies: vendors, models, data pipelines, infrastructure, and scarce talent. Many of these dependencies are not visible at the application layer and increasingly emerge from the industrial structure of AI systems themselves. These dependencies constrain reversibility and concentrate risk.

​

Core board question

 

“What are we becoming dependent on and can we exit?”

​

Typical signals

 

· Increasing reliance on a single vendor or model

​

· Limited internal understanding of critical systems

​

· Dependencies not fully mapped or documented

​

Example: Claude.ai global outage (2026)

 

“A major outage halted AI‑dependent workflows across organizations…” (4)

The outage exposed how reliance on a single provider can halt operations across entire functions.

​

Tension: Efficiency vs. autonomy


Failure mode: Loss of optionality and constrained decision space


System effect: Single points of failure with systemic impact

​

​

Board response

​

Non‑negotiable:

​

Critical dependencies must be visible and understood. “What happens if this dependency fails?”

​

Challenge questions:

​

· Where are we locked in and how do we exit?

· What is the fallback if this system becomes unavailable?

· Are we optimizing for efficiency at the cost of resilience?

​

Leadership countermeasure:

 

Maintain exit paths, fallback modes, and minimum viable internal capability.

4. Legitimacy: Societal Distortion

 

Legitimacy is the force of society judging systems.

​

Description

 

AI decisions operate under increasing public, regulatory, and stakeholder scrutiny. What is technically possible may be socially unacceptable (and what is acceptable today may not remain so).

​

Core board question

 

“Would we stand by this if it were public tomorrow?”

​

Typical signals

 

· Growing public or regulatory scrutiny

​

· Internal discomfort not formally addressed

​

· Decisions justified by legality rather than legitimacy

​

Example: Deepfake evidence in U.S. court

 

“A judge identified that a submitted witness video was AI‑generated…” (5)

The case showed how fabricated content can undermine trust in institutional processes.

​

Tension: Innovation vs. acceptability


Failure mode: Loss of trust despite technical correctness


System effect: Decisions become socially or institutionally unsustainable

​

​

Board response

​

Non‑negotiable:

​

Legitimacy must be explicitly tested. “Have we tested whether this remains acceptable over time?”

​

Challenge questions:

​

· Who could be affected beyond our immediate stakeholders?

· Are we relying on legality instead of legitimacy?

· How might this be perceived under scrutiny?

​

Leadership countermeasure:

 

Test decisions against evolving societal expectations, not just compliance thresholds.

Executive Reflection

 

The forces described here do not make decisions but shape how they emerge.

Drift, Adversarial Pressure, Dependency, and Legitimacy do not act as isolated risks. They operate continuously, often invisibly, influencing what is perceived as stable, reliable, reversible, or acceptable.

​

Governance increasingly operates under conditions where responsibility scales more slowly than the systems shaping decisions.

​

This is why decisions can appear sound while already being misaligned. This is not governance failure. It is the environment in which judgment is exercised that has shifted.

​

Boards cannot govern decisions they do not understand. And they cannot understand decisions without understanding the forces that shape them.

Governance begins upstream.

​

In an AI‑driven environment, the challenge is not only to respond to what is visible, but to recognize what is shaping visibility itself. Because by the time a decision is clearly wrong, it has often already been shaped by forces that were never made explicit ... and governance cannot correct what it never saw.

​

Igor Allinckx

AI & Governance

​

May 2026

​​

​​​​

Back to Insights

​

Part of an ongoing exploration of governance, AI, and human judgment.

bottom of page