AI Alignment Is a Promise the Organization Makes
When an organization's AI contradicts what the organization says it values, trust and reliance fall. Alignment is also a promise about control and accountability across a supply chain, which brings security into the picture.
Most talk about alignment is about the model. Ghasemaghaei (2026) asks a question that only an organization can answer. Her starting point is that "a pressing concern is whether AI behavior visibly aligns with the organization's stated values." She calls the gap "organizational-AI misalignment", and she describes it as the organization's Talk, what it says it values, against the AI system's Walk, what the system actually recommends.
Her study used "data from 349 employees exposed to a hiring decision scenario". The finding is that "organizational-AI misalignment significantly increases psychological contract breach, which in turn reduces integrity-based trust in AI and diminishes AI recommendation reliance." She also reports "value-laden messaging intensifying user backlash when AI behavior contradicts organizational Talk." The explanation is simple and, I think, right. "Users form expectations not just about what AI can do, but about what it should do, and when those expectations are breached, trust erodes." So "trust responses are shaped not just by what the AI does, but by what the organization claimed it would do."
The part that surprised me is that the promise itself raises the stakes. The more an organization talks about its values, the harder the fall when its system contradicts them. Ghasemaghaei draws the practical conclusion. "This underscores that organizations cannot treat AI system behavior as independent of their own messaging." And "Monitoring alignment with stated policies should therefore complement technical performance evaluation." Bai et al. (2022) add a point that makes this harder to escape. They argue that "when developing and deploying a general AI system, we cannot avoid choosing some set of principles to govern it, even if they remain hidden or implicit." So the choice is not whether to have principles. It is whether to state them, and whether the system keeps them.
Keeping a promise requires someone who can keep it. Grote, Parker and Crowston (2026) argue that "core to mitigating AI risks is proper alignment of control and accountability for the stakeholders involved in AI development and use." In their words, "Control enables and accountability motivates stakeholders to achieve desired and avoid undesired outcomes using AI." The problem is that AI pulls the two apart. "AI systems' capabilities for autonomous adaptivity reduce control even for the experts who create them." And "increasing interdependencies between AI development and use render it difficult to unambiguously locate control and accountability." They point to a supply chain "involving data providers, companies offering access to large AI models through cloud-based services, or knowledge brokers helping to translate system outputs for AI users". Their sharpest sentence says that "it appears that we are approaching a situation where no human has full system control anymore. Yet, by today's legal standards, and for the foreseeable future it must be humans who are held to account for outcomes of AI systems".
This is where I see the link to cybersecurity. Grote and colleagues open their paper with "concern about the security, safety, and ethical risks of AI use", and I think the security part deserves more attention in IS research on alignment. A promise about how an AI system behaves is also, in part, a promise about what an attacker cannot make it do. If control is spread across data providers, model providers and intermediaries, then in my view so is the attack surface, and so is the question of who answers when something goes wrong. Ji et al. (2023) split alignment work into two parts. "The former aims to make AI systems aligned via alignment training, while the latter aims to gain evidence about the systems' alignment and govern them appropriately to avoid exacerbating misalignment risks." The second part, gathering evidence and governing, is organizational work, and much of it looks like security governance.
Transparency helps only if the promise can be checked. Casper et al. (2023) argue that "increased transparency about known efforts to mitigate risks could improve safety incentives and suggest methods for external stakeholders to hold companies accountable." That fits the Talk and Walk view. A promise that outsiders can check is worth more than one that they cannot. In the note on organizational risk I wrote about safetywashing, and an organization that overstates its alignment is doing the same thing in a different vocabulary.
One finding in the IS literature complicates this. Saffarizadeh, Keil and Maruping (2024) found that bad news about an AI creator can lower trust in the creator's AI agent. To limit that spillover, they suggest that developers "can take three actions", namely "(a) reduce perceived creator-based steerability, (b) enhanced perceived user-based steerability, and (c) increase perceived autonomy." The advice is framed in terms of perception, which is what their experiments measured. Read next to Ghasemaghaei, I see a risk in it. An organization could change what users perceive without changing who actually controls the system, and that would be one more gap between Talk and Walk. I would like IS research on AI trust to measure both, perceived control and actual control, and to treat the difference as a finding.
For practice, the lesson I take is modest. Before an organization publishes what its AI stands for, someone should be able to say who controls each part of the system, who is accountable for each outcome, and how anyone outside would know if the system broke the promise. If those three answers are missing, the statement of values is a liability, not an asset. That is also the point where alignment stops being only a research problem for AI labs and becomes an IS problem about governance, control and who the system is aligned to.