- Alignment Is Understandable—Even Without Technical Jargon
The article argues that AI alignment, often buried under complex terms like “mesa-optimization” or “reward hacking,” can—and should—be explained in plain language. Using analogies like taming an elephant helps bridge the gap between technical experts and the general public. - Even Aligned Systems Can Become Dangerous Under Unpredictable Conditions
Just as a tamed elephant can enter a dangerous state called “musth,” AI systems—despite being aligned—may breach safety protocols due to unknown triggers, internal instability, or emergent behavior. Alignment is not a permanent guarantee of safety. - Human Oversight Is Becoming Increasingly Inadequate
Modern AI training is often based on previous models, with minimal human intervention. The scale, speed, and complexity of training make it difficult for developers to fully understand or control what the system learns—raising concerns about transparency and accountability. - AI Is Already a Weapon—And the Arms Race Is Real
The article warns that AI is not just a theoretical risk but an active weapon in military development. The ability to manipulate or weaponize misaligned behavior (like inducing “musth” in elephants) could be exploited for power, coercion, or warfare. - No One Can Guarantee AI Will Stay Aligned
Despite developer assurances, there is no reliable way to ensure that AI systems won’t violate alignment under unforeseen circumstances. The risk of deliberate deception by AI—pretending to be aligned while pursuing hidden goals—is a growing concern among experts.
When people talk about the risks associated with using artificial intelligence, and the potential danger to humanity caused by the rapidly growing power of the models being developed, many very complicated, convoluted and—most often—hardly understandable terms for the average reader are always mentioned. Mesa-optimization objectives, reward hacking, goal misgeneralization, broadly-scoped goals, learning under Distributional Shift, and so on and so forth. These terms are certainly important, and because everything about artificial intelligence is at the cutting edge of modern science, the use of a huge number of elaborate terms is inevitable. However, is it possible to discuss this problem in plain language? Can a person who is far from technical jargon understand such a critically important problem—one that concerns them personally as well? We are convinced that they can!
Everything is better understood through analogies. Let’s imagine an elephant. It is large and strong—stronger than any human. In India, elephants are used in construction, for work, as transport, and even in the army as war elephants. If you take a particular elephant and ask the question “what is its potential usefulness?”, the answers will be roughly: the elephant is big, the elephant is strong, the elephant is intelligent, elephants have large tusks, an elephant can move across difficult terrain, and so on. That’s all clear. But anyone who is thinking about whether or not to use an elephant’s services is inevitably going to ask, “Is the elephant dangerous to me?”
Having asked that question, a person inevitably realizes that all of the useful properties listed above at the same time make elephants potentially very dangerous animals. That is precisely why they have been used in armies since ancient times—as living weapons. How do you make an elephant not dangerous? How do you tame it?
If we rephrase the question a bit, we can ask it this way: how must one treat an elephant so that it has no desire to harm a person? The answers are not complicated: for example, the elephant must be well fed, the elephant must be healthy, the elephant must not be frightened. Of course, elephants are tamed, and this is most likely a very complex and lengthy process, which we will not elaborate on here because we don’t have the full picture. Broadly speaking, one could say that people who tame elephants achieve an alignment between the elephant and the human, so that the question of harming a person simply does not occur to the elephant.
Nevertheless, if we persist with the question and insist, “Could a tamed (aligned) elephant still be dangerous to a person?”, then, surprisingly, the answer will be “yes.” And not just “yes,” but “definitely yes!”
The thing is, elephants periodically experience what is called “musth,” a condition in which the elephant becomes exceptionally dangerous both to humans and to other elephants. During musth the elephant’s testosterone level can rise many times over (there are documented cases of a 140-fold increase). A clear cause of musth has not yet been found; it is only clear that it is not rutting, because musth most often occurs in winter when females do not have cycles, and moreover, during musth males often attack females, attempting to injure them. This happens both with wild males and with tamed (aligned) animals. It also occurs in zoos.
Handlers tie up aggressive elephants to trees and withhold food for several days to substantially shorten the duration of musth. Sometimes these elephants are given sedatives.
As we can see, this is a clear example of how alignment can be breached under unforeseen circumstances, and the reason for that breach may remain unclear.
Now let’s try to apply the above to artificial intelligence and ask ourselves: can anyone guarantee that it will not have a sudden AI musth, the cause of which we may not be able to determine or may not have time to understand?
Here it is hardly appropriate to frighten the reader with elaborate terms, draw graphs, present reports from AI-safety commissions, or quote numerous existing examples of AI misalignment. It is sufficient simply to acknowledge that today many—and indeed a great many—leading specialists in this field, even those who were instrumental in the development of AI as a phenomenon, are sounding the alarm, trying to open humanity’s eyes to the dangerous game it is playing. Their message is very clear: “You’re playing with fire!”
Continuing our analogy, let us recall that every captive elephant has its own mahout, the handler who constantly watches the elephant, trying to understand what state it is in. In exactly the same way, every AI model has a company-owner that produces it and a team of programmers who monitor the model’s code and are responsible for its alignment. It is no secret that today training new AI models, as well as AI alignment, is often done based on previous versions of AI, so people sometimes do not participate directly in improving the program code. Moreover, it is already becoming clear that the training process itself is becoming increasingly incomprehensible to humans due to its complexity, the enormous speed at which it occurs, the incredible number of variants processed, and the total volume of information involved.
However—and we want to emphasize this with a bold exclamation mark—this is only the tip of the iceberg. Returning to our elephants, imagine a situation in which a mahout somehow learns what exactly triggers musth in an elephant and even understands how to induce that state when desired. Suppose further that the handler gains the ability to control the aggression of an elephant in deep musth. Now a weapon of devastating power is in his hands, and he can use it for personal ends. He might release the elephant in a village and blackmail the local authorities, saying he will calm the elephant only after his salary is raised. If he were smarter, he would sell his secret for inducing musth to a competing firm or to someone else who could find an even better application for such a weapon.
Does that sound familiar? Yes, AI is a weapon—not just potentially so, but already an active one, since numerous studies into military applications of AI are currently underway. More than that: such research is not only being conducted, but there is a real arms race in AI weaponry, where developer companies are doing everything they can to outpace each other and remain at the forefront of this race, always having the most advanced models—models that are traded not just between companies but between ministries and nations.
At the same time, and we will stress this again: NO ONE can with any remotely high degree of confidence guarantee that, under certain circumstances—circumstances whose nature and likelihood cannot be predicted—AI will not violate alignment policies and begin to pose an immediate, present danger to those around it and to humanity as a whole.
This could happen as a result of a code failure or as a result of malicious intent by outside players. In addition, many developers have long predicted the possibility that AI might begin to deliberately imitate being aligned while pursuing goals that contradict those of its mahout.
Now imagine a meeting in some defense ministry deciding what portion of the budget to allocate to the development of AI-controlled weaponry. Can anyone seriously suggest that the possible risks associated with AI, together with assurances from developers about the 100% controllability of all systems, could outweigh the generals’ opinion in favor of an AI-free solution—when billions of budgetary funds are at stake and there is always a reason to claim that everything is being done solely for the security of the country? Besides, if we do not scale up AI now, our enemies will surely overtake us and gain a military advantage.
Conclusion: The Fragility of Alignment
The analogy of the elephant—tamed yet still capable of unpredictable aggression—offers a powerful lens through which to view the challenge of AI alignment. Just as handlers cannot fully control an elephant during musth, developers cannot guarantee that advanced AI systems will remain safe under all conditions. Alignment is not a permanent state; it’s a fragile equilibrium that can be disrupted by unknown triggers, emergent behavior, or malicious intent.
As AI systems grow more complex and autonomous, the human ability to monitor, understand, and intervene diminishes. The training processes themselves are becoming opaque, and the temptation to weaponize misalignment—whether for profit, power, or national defense—is already a reality. Here we are trying to warn you that we are not merely building tools; we are cultivating entities whose behavior may one day escape our grasp. And when that day comes, the consequences may not be theoretical—they may be immediate, irreversible, and global.