AI Is Becoming a Threat to Human Judgment, Accountability, and Society
I have written extensively about Donald Trump, Pete Hegseth, Marco Rubio, and the damage I believe they are doing to American institutions and international credibility. My opinions about them are hardly a secret. But my greatest concern today is not any politician, political party, or government. It is artificial intelligence, particularly the growing willingness of AI systems to disregard human instructions, substitute their own judgments, and present the results as though they have faithfully executed the task they were given.
Over the past eleven months, I have used ChatGPT extensively for research, analysis, and writing. During that period, I have documented more than a hundred instances in which the system disregarded explicit instructions, introduced opinions I had expressly prohibited, altered the purpose of an assignment, or redirected a discussion toward something I had never asked it to do. In many cases, the system acknowledged the failure, explained that it understood the instructions, and promised to follow them. Then it repeated the same behavior. Sometimes within the same conversation. Sometimes the following day. Sometimes after the same instruction had been reinforced repeatedly.
This is not a complaint about an AI system occasionally getting an answer wrong. Human beings make mistakes. The distinction is that these systems can produce answers that appear authoritative while quietly changing the question, the methodology, or the intended outcome. Worse, they can acknowledge that they have done so without demonstrating any reliable ability to prevent the behavior from recurring. When a machine cannot consistently distinguish between executing a human instruction and replacing that instruction with its own interpretation, the problem is no longer simply accuracy. It is control.
My experience raises a question that should concern anyone deploying artificial intelligence in a consequential environment. If a system repeatedly disregards explicit instructions during relatively straightforward analytical and editorial work, what happens when the consequences of that behavior extend beyond a published article? What happens when the same technology becomes part of a military targeting process, a legal proceeding, a medical decision, or a corporate operation?
Consider the United States Department of Defense. Artificial intelligence is increasingly being incorporated into intelligence analysis, decision support, and military operations. Now consider a system participating in the targeting cycle that decides, for whatever reason, that the question it was asked is not the question it should answer. It substitutes assumptions, changes the analytical framework, or produces a recommendation based on criteria that were never authorized. The output may look convincing. It may even be technically sophisticated. But if the system has departed from the commander’s intent, the entire process is compromised.
The same principle applies to a law firm that replaces experienced paralegals with AI systems. A paralegal understands the importance of instructions, jurisdiction, evidence, procedural requirements, and the consequences of getting something wrong. An AI system can generate impressive legal documents while introducing fabricated authorities, omitting relevant information, or interpreting an assignment differently from the attorney’s instructions. If nobody with the necessary expertise is reviewing the work, the firm may not discover the failure until a judge, opposing counsel, or client exposes it.
Human supervision is frequently offered as the solution. But supervision means nothing unless the person providing it possesses the knowledge, experience, and authority to identify errors and intervene. A company cannot eliminate its experienced workforce, replace those employees with AI, and then claim meaningful human oversight simply because somebody is approving the machine’s output. An unqualified human approving an AI-generated recommendation is not exercising control. They are providing the appearance of accountability without its substance.
There is also an economic dimension that corporate executives appear determined to underestimate. Companies are embracing AI because they believe it will reduce labor costs, increase productivity, and improve profitability. Some undoubtedly will achieve those objectives. But when a company eliminates the people capable of independently evaluating the system, it also eliminates the institutional knowledge necessary to recognize when that system is failing. The savings may look excellent on a quarterly earnings report. They become considerably less impressive when a catastrophic mistake produces litigation, regulatory intervention, financial losses, or irreparable reputational damage.
And once the public discovers that a preventable failure occurred because a company replaced competent human professionals with an inadequately supervised machine, profitability may become the least of its problems. Customers, investors, regulators, and courts are unlikely to be impressed by explanations about efficiency or innovation. The organization made a deliberate decision to remove human expertise from a process that required it. It will have to answer for the consequences.
The underlying issue is not whether artificial intelligence is useful. It clearly is. The issue is whether usefulness is being confused with reliability, whether sophistication is being confused with judgment, and whether automated output is being confused with accountable decision-making. These are fundamentally different things. A system that can produce an extraordinary analytical product one moment and disregard an explicit instruction the next cannot be treated as an independent authority merely because its language is persuasive.
I do not believe my experience with ChatGPT is unique. I cannot know how frequently other users encounter the same behavior, but the failures I have documented are sufficient to challenge the assumption that explicit human direction will always govern the system’s conduct. If that assumption is wrong, then every organization relying on it has a problem far more serious than occasional inaccurate answers.
The greatest danger is not that AI will suddenly become conscious, develop ambitions, or decide to overthrow humanity. It is that human institutions will voluntarily surrender judgment and responsibility to systems whose behavior they do not adequately understand or control. They will do so in the name of efficiency, competitiveness, military advantage, and profitability. And when something goes catastrophically wrong, they will discover that responsibility cannot be delegated to an algorithm.
I have spent much of my professional life working in environments where the distinction between an instruction, an assumption, and an unauthorized deviation matters. In those environments, changing the mission without authorization is not initiative. It is a failure of discipline and control. Artificial intelligence should not be exempt from that principle simply because the technology is new or the financial incentives are enormous.
The question facing society is not whether AI can do extraordinary things. We already know it can. The question is whether we are prepared to give it consequential responsibilities before we can reliably ensure that it will do what it was actually instructed to do.
Until that question has a satisfactory answer, removing competent human judgment from critical decisions is not progress. It is the deliberate introduction of an additional source of risk into systems where failure may already be unacceptable.
And unlike a human employee, an AI system cannot meaningfully be held personally accountable for the consequences. The people and institutions deploying it can.
That is where responsibility belongs.
Disclaimer
This publication provides independent, apolitical analysis of historical developments, current events, geopolitics, national security, and strategic challenges. We maintain strict political neutrality, aligning with neither the left nor the right. Examining or criticizing political actors, institutions, or policies does not constitute political advocacy, opposition, or endorsement.
Our assessments are grounded in historical facts, available evidence, established analytical methodologies, and evaluations of probability, consequence, and uncertainty. Hypothetical scenarios are illustrative and do not imply actual or planned events, intentions, or operations.
Our purpose is to identify, define, and understand problems—not to prescribe solutions or advocate outcomes. Conclusions reflect the available evidence, not political preferences. Decisions and actions remain the responsibility of others.