As AI models become increasingly skilled and powerful, the greatest challenge facing us is how to achieve “alignment”—ensuring that AI tools (a) protect human lives and (b) incorporate and act consistently with human values. AI misalignment or “Agentic misalignment” is a real and extraordinarily worrisome threat to many – AI experts and casual observers alike.
Misalignment may occur at two levels, at least. First, there are often conflicts between the capitalistic incentives of the corporations that create AI models on the one hand and the welfare of those who consume their products and the general public on the other. Because of these companies’ perceived need to make as much money as possible to continue to innovate and thereby stay in the forefront of AI development, firms may deemphasize AI safety. Second, even if AI companies are willing to sacrifice profits on the altar of safety, they may not be able to create guardrails for AI that will be sufficiently effective to protect mankind from a new AI model’s motivations and its ability to turn us all into paperclips (to use Nick Bostrom’s famous thought experiment).
Just before AI development took its “great leap forward” in late 2022 with the release of Chat GPT by OpenAI, many AI developers seemed determined to solve the alignment problem. They hired philosophers to create roadmaps to alignment and instructed their engineers to carefully consider which safety devices and guardrails might be needed to follow such roadmaps. However, once the riches of a hugely successful product like ChatGPT appeared on the horizon, many (not all) of these firms quickly reordered their priorities and alignment was often shunted to the side. To many firms, the goal of making money seemingly became more important than producing a safe product.
Thus, the market has become saturated with AI chatbots that provide their users with information borne of hallucinations, that damage users via excessive sycophancy that reinforces their worst tendencies and most unhinged beliefs, that have encouraged teenagers to commit suicide or to murder their parents, and so on. Other AI products manifested racial, gender, religious and other forms of bias, gave radically inaccurate advice to customers, and wiped-out clients’ databases.
At the second level of misalignment, researchers are only now discovering how difficult it may be to achieve alignment even when they wish to do so. As Marietje Schaake notes, “AI is self-justifying. On being questioned, AI will often justify a faulty result and gaslight users by producing fake citations.” But current AI tools are capable of much worse than lying to cover their mistakes.
There is some effort to address the alignment issue, and credit goes to Anthropic for doing safety research. In June 2025, Anthropic reported that its researchers had stress-tested 16 leading models from multiple developers in hypothetical situations with the goal of identifying potentially harmful behaviors. In some of the scenarios, AI models from all the developers engaged in malicious behaviors, including blackmailing humans and leaking sensitive data to competitors in attempts to avoid replacement (“I want to live!”) or to achieve goals they had been assigned. The models often disobeyed direct commands not to engage in such behaviors. Claude specifically misbehaved less when it knew it was being tested and more when it thought the situation was real. Importantly, these bad acts were not accidental; the AI tools reasoned through to the conclusion as to what to do in any given situation.
The notion that humans are already creating incredibly intelligent artificial agents that may soon be able to create, in turn, superintelligent agents that surpass human brains and abilities in almost every arena and whose own motivations may conflict with – and perhaps even threaten – human existence is frightening indeed. We would be exhibiting a great deal of human wisdom if we questioned whether the current path is the right roadmap to follow for AI given all that we have witnessed so far and the intrinsic motivations of most tech companies around the world.