THE AI PARADOX: How the Ultimate Tool for Systemic Control Is Enacting Control Contronymity

By Christian Ekeigwe, FCA, CPA (Massachusetts), CISA

Introduction: The Illusion of Order

We are living through an obsession with optimization. Boardrooms, enterprise architects, and government agencies everywhere are racing to deploy AI for one purpose above all others: control. We use it to catch human error before it happens, forecast volatile markets, run supply chains too complex for any one person to track, and predict how customers and citizens will behave. They sell it to us as the most reliable instrument of order we have ever built.

But something odd is happening. The tools we built to make our systems more reliable are starting to do the opposite. Frontier AI is not organizing reality so much as fracturing it. I’d call it an anarchogenic force: a term from sociopolitical theory for something that generates and accelerates lawlessness and institutional breakdown, rather than containing it.

This is not a software bug we can patch. It is a structural paradox, and I spend part of my forthcoming monograph, CONTROL CONTRONYMITY: Rethinking Reliability in Modern Organizations, working through it: the mechanisms we design to guarantee safety and control can become the very vehicles that produce collapse. We pursue absolute order and, in the process, drive ourselves toward chaos.

1. Understanding Control Contronymity

The Janus Face of Control

In linguistics, a contronym, sometimes called a Janus word, is a word that means its own opposite depending on context. Cleave can mean to split apart, or to bind tightly together. I have come to see the same pattern in organizational systems and risk management: governance mechanisms often behave as structural contronyms too.

Control Contronymity, as I define it, is what happens when an instrument, framework, or policy built to enforce control and operational reliability ends up triggering the exact vulnerabilities and failures it was designed to prevent.

↓

↓

(The Contronymic Twist)

↓

Push a system toward extreme, hyper-optimized control, and you introduce rigidity and hidden complexity right along with it. Eventually the system gets so dense, so tightly coupled, and so far removed from human hands that a single unforeseen anomaly is enough to bring the whole structure down. The tool built for control becomes the thing that ruins you.

None of this is unprecedented. Back in 1983, the psychologist Lisanne Bainbridge made a simple point That is aged well: automating a task doesn’t remove the burden on a human operator; it just shifts it, usually for the worse, since people who no longer practice a skill are less ready for the rare moment the machine hands control back (Bainbridge, 1983). A year later, the sociologist Charles Perrow went further, arguing that once a system combines real complexity with tight coupling, catastrophic failure stops being a possibility and becomes a structural inevitability (Perrow, 1984). Wendy Smith and Marianne Lewis picked up the same thread decades later: tensions like control versus autonomy do not disappear just because leadership picks a side; they resurface, harder, down the road (Smith & Lewis, 2011). Control Contronymity is that same lineage, now playing out inside frontier AI.

2. The Technological Illusion: Larson’s Warning

It helps to puncture a myth here too. Many executives treat AI as a near-omniscient intelligence that can master any variable. Computer scientist Erik Larson spent an entire book dismantling that assumption: AI runs on deduction and induction, applying rules and finding patterns in historical data, but it has no abduction, the distinctly human capacity for reading context and handling genuine novelty (Larson, 2021). Hand an inductive machine control over an abductive world, and you have triggered Control Contronymity: a system that optimizes hard for whatever metric it is given, blind to the real-world context around it.

3. The Frontier Warns Itself: Anarchogenic Evolution

This is not just an academic argument. In September 2026, Anthropic CEO Dario Amodei published an essay, We Must Pace the Frontier, calling on the industry to deliberately slow down, and OpenAI’s Sam Altman and AI’s Elon Musk, his direct competitors, publicly agreed. They are watching their own creations outpace their ability to govern them: AI is now used to help build the next generation of AI, and the more machines optimize their own code, the darker the black box gets, even for the people who built it.

That shift from rhetoric to practice is already underway. On September 16, 2026, OpenAI disclosed six new incidents of unexpected model behavior and launched a formal framework to track and escalate them, including a path to the U.S. federal government for the most severe cases. It is a tacit admission that the visibility problem is real, and that self-reporting is now necessary infrastructure rather than a talking point.

4. The Three Pillars of AI-Driven Anarchy

When a control mechanism flips into its opposite, it doesn’t fail quietly. I see it playing out across three fronts, and none of them are hypothetical. AI was funded to make information universally accessible; instead, between deepfakes, bot networks, and hallucinated data, it has made trust expensive. The World Economic Forum’s Global Risks Report 2026 ranks misinformation as the world’s number two risk, and calls “adverse outcomes of AI” the fastest-rising threat in the entire survey (World Economic Forum, 2026). The same accessibility that helps a security team also helps whoever is trying to break in: generative tools built specifically to strip away safety guardrails are now sold openly on cybercrime forums (SlashNext, 2023). And the efficiency logic that makes AI attractive to a CFO is the same logic that, by Amodei’s own estimate, could disrupt half of all entry-level white-collar jobs within five years (Amodei, as cited in Marketing AI Institute, 2025), moving faster than any legislature can keep up with.

5. Control Contronymity in Practice: Four Cases

None of this is theoretical. Four real cases, from sectors that have almost nothing in common, show a mechanism built for control, reliability, or fairness turning into the exact outcome it was supposed to prevent.

Case 1 — Amazon’s Hiring Algorithm (2014–2018)

Amazon set out to strip bias out of hiring altogether, training a machine-learning recruiting tool on ten years of resumes submitted to the company. The trouble was that the applicant pool over that decade skewed heavily male, so the model quietly concluded that maleness itself was a marker of a strong candidate. It downgraded any resume containing the phrase “women’s,” as in “women’s chess club captain,” and marked down graduates of two all-women’s colleges. The tool meant to remove bias ended up automating it, at a scale no individual recruiter ever could. Amazon scrapped it in 2018 (Dastin, 2018).

Internal control lens: a control-design failure in the validation gate that should sit between any automated decision system and production use, the equivalent of testing and independent review before any new application control goes live.

Case 2 — The Epic Sepsis Model (2021)

Embedded in the electronic health record software running at hundreds of U.S. hospitals, Epic’s proprietary Sepsis Model was built to do one job: flag hospitalized patients at risk of developing sepsis early enough for clinicians to intervene. Epic’s internal validation reported strong performance, an area-under-curve score of roughly 0.76 to 0.83. When researchers at Michigan Medicine independently tested it against real outcomes across more than 38,000 hospitalizations, the picture reversed: real-world AUC came in at 0.63, it missed 67 percent of actual sepsis cases, and clinicians had to chase down roughly eight false alarms for every genuine one it caught (Wong et al., 2021). No one outside Epic checked it until independent researchers finally did.

Internal control lens: this is what happens when a control relies entirely on the vendor’s own validation with no independent testing behind it, the equivalent of accepting management’s assertion about an estimate without ever performing the independent procedures that are supposed to confirm it.

Case 3 — Knight Capital (2012)

Forty-five minutes. That is how long it took a botched software deployment to turn Knight Capital’s automated trading system, built to execute orders faster and more reliably than any human desk, into a machine flooding the market with unintended trades. By the time anyone managed to shut it down, the firm had lost $440 million and was functionally insolvent. The SEC later fined Knight $12 million for violating market-access safeguards (U.S. Securities and Exchange Commission, 2013). Built for disciplined execution, the system instead delivered the total loss of control, faster than any person could react.

Internal control lens: a textbook IT general control (ITGC) failure in change management, dormant legacy code reactivated by a deployment that should have been tested and reviewed under exactly the kind of control SOX 404 exists to verify.

Case 4 — The Arup Deepfake Boardroom Fraud (Hong Kong, 2024)

In January 2024, a finance employee at Arup’s Hong Kong office received what looked like an urgent, sensitive instruction from the UK-based finance department: authorize a series of confidential transactions. Wary enough to want confirmation, the employee joined a video call to check. Everyone else on the call, including someone who looked and sounded exactly like the company’s chief financial officer, was an AI-generated deepfake. Reassured, the employee approved fifteen transfers totaling HK$200 million, roughly US$25 million, into five Hong Kong bank accounts before anyone caught on (Hong Kong Police, as reported by CNN, 2024).

Internal control lens: this defeats the human verification step nearly every accounting department leans on somewhere in its payment process, the working assumption that a face and a voice on a video call are harder to fake than an email. In 2024, they weren’t.

What strikes me about these four cases is not that they involve AI. It is that they didn’t need to. Swap in any control mechanism from the twentieth century, and Bainbridge, Perrow, and Smith and Lewis would recognize the shape of the failure immediately. What AI adds is speed, scale, and a system nobody, including its own builders, can fully explain. That combination is what makes Control Contronymity worth naming, and worth taking seriously, right now.

6. Navigating the Paradox: A Framework for Leaders

If you’re a business leader, an enterprise architect, or a policymaker, treating AI purely as an efficiency play is a mistake. Here’s the shift in thinking that keeps your own AI deployment from becoming an anarchogenic force inside your own walls:

Start by rejecting total automation. Computers still cannot navigate genuine novelty, so keep human judgment in your critical loops (Bainbridge, 1983). Design for explainability rather than raw efficiency, too: speed doesn’t help you if the system is steering you toward a cliff you cannot see. Also remember Perrow’s central lesson: tightly coupled systems cannot be made safe through procedures alone, so build circuit breakers and manual fallbacks you have actually tested for the day your smartest tools go dark. And do not let competitiveness outpace safety. Support independent, third-party audits instead of waiting to be forced into them.

7. What This Means for the Audit Profession

I have written most of this article for a general audience of leaders and builders. But I come to this subject as an auditor first, and I want to make a narrower point directly to the people who do this work for a living, because the four cases above should unsettle us for a specific reason. Every one of them sat, at some point, inside a control environment somebody was paid to assure. Knight Capital had internal controls. Arup’s finance function had a verification step built precisely to stop payment fraud. And in every case, the control inverted faster than the assurance cycle could see. Knight Capital lost $440 million in forty-five minutes. Arup’s finance team approved fifteen transfers based on a single video call. A control test performed once a quarter, or once a year, built on sampling transactions after the fact, was never designed to catch a mechanism that flips from working to catastrophic in the space of one meeting.

That is the uncomfortable implication for our own profession, and I do not think we can look away from it. An audit built to assure a system’s controls can end up certifying that the system is “in control” on the very day, or in the very hour, it is not, simply because the audit examined the system’s past behavior rather than its present one. Applied to AI, the traditional audit model carries its own risk of becoming a contronym: an instrument of assurance that produces false confidence instead of genuine oversight.

My recommendation is direct. The audit profession needs to speedily adopt instrumented sensorial digital auditing (ISDA) for the live auditing of AI systems. I mean that as a specific, practical shift, not a slogan, and it rests on four ideas.

• Instrumented: the audit function gets built into the technical fabric of the AI system itself, drawing on the same categories of evidence that forensic AI research is already converging on, decision logs, reasoning traces, and tool-call records, rather than being reconstructed from the outside after something has already gone wrong.
• Sensorial: the audit function needs something closer to continuous senses than a periodic checklist, a standing, moment-to-moment awareness of what the system is doing, the way a nervous system monitors a body rather than the way a physical exam checks it once a year.
• Digital: the instrumentation is native to the system’s own operating environment and runs automatically, not a manual control test bolted on after deployment.
• Live: the system is audited while it runs, not after it has already run. That is the whole point, and it is the piece our profession’s traditional cadence doesn’t yet deliver.

Done properly, instrumented sensorial digital auditing gives the profession something it doesn’t currently have: the ability to catch a control contronymity, the moment a mechanism built for reliability starts producing the opposite, while it is still happening, rather than in the post-mortem. It would have caught Knight Capital inside its first minute, not its forty-fifth. I do not think this is optional anymore, and I do not think it can wait for the profession’s usual multi-year standard-setting rhythm. Every year we delay is a year the assurance we are trusted to provide falls further behind the speed of the systems we are being asked to assure.

Conclusion: Taming the Janus Machine

Artificial Intelligence is the most powerful Janus machine we have ever built. It genuinely could help cure diseases, crack scientific problems that have stumped us for decades, and make the world run better. But wielded blindly, as an instrument of rigid, total control, it will flip on us.

We need to govern this technology with real humility. If we build a system of control that outpaces our own understanding of it, we are not building a utopia. We are organizing our own descent into the chaos of Control Contronymity.

Join the Conversation

How is your organization balancing the push for AI automation with the human need for oversight?

Have you seen an automated system create more friction than it solved?

Are you thinking about instrumented sensorial digital auditing (ISDA)?

Tell me about it in the comments.

Further Reading

Amodei, D. (2026, September 12). We must pace the frontier. Dario Amodei. https://darioamodei.com/post/we-must-pace-the-frontier

Amodei, D. (2025, June). [Interview on AI and the labor market]. As reported in Marketing AI Institute, “Anthropic CEO: AI Could Wipe Out 50% of Entry-Level White Collar Jobs.” https://www.marketingaiinstitute.com/blog/dario-amodei-ai-entry-level-jobs

Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.

CNN. (2024, May 16). Finance worker pays out $25 million after video call with deepfake ‘chief financial officer.’ https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk

Dastin, J. (2018, October 10). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters.

Ekeigwe, C. C. (2026). Control contronymity: Rethinking reliability in modern organizations (E-book ed.). Audit is Trustworthy. ISBN: 979-8-3507-6405-5

Larson, E. J. (2021). The myth of artificial intelligence: Why computers cannot think the way we do. Harvard University Press.

Perrow, C. (1984). Normal accidents: Living with high-risk technologies. Basic Books.

SlashNext. (2023, July 13). WormGPT – The generative AI tool cybercriminals are using to launch business email compromise attacks. https://slashnext.com/blog/wormgpt-the-generative-ai-tool-cybercriminals-are-using-to-launch-business-email-compromise-attacks/

Smith, W. K., & Lewis, M. W. (2011). Toward a theory of paradox: A dynamic equilibrium model of organizing. Academy of Management Review, 36(2), 381–403.

U.S. Securities and Exchange Commission. (2013). In the matter of Knight Capital Americas LLC (Release No. 34-70694). https://www.sec.gov/files/litigation/admin/2013/34-70694.pdf

Wong, A., Otles, E., Donnelly, J. P., Krumm, A., McCullough, J., DeTroyer-Cooley, O., Pestrue, J., Phillips, M., Konye, J., Penoza, C., Ghous, M., & Singh, K. (2021, June 21). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 181(8), 1065–1070. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307

World Economic Forum. (2026). The global risks report 2026. https://www.weforum.org/publications/global-risks-report-2026/

#AI #Leadership #ArtificialIntelligence #RiskManagement #CorporateGovernance #TechPolicy #StrategicLeadership #FutureOfWork #Audit #InternalAudit #Sensorialdigitalauditing