AI Alignment Is a Process, and IS Has Studied That Process Before
IS alignment research spent thirty years moving from alignment as a state to alignment as continual adjustment. AI alignment research is now making the same move, and the older field has lessons to offer.
The word alignment had a long life in IS before it reached AI. Benbya, Leidner and Preston (2019) open their research curation with a sentence that could sit in an AI safety paper with one word changed. "The alignment of Information Systems (IS) with the business (i.e., hereafter IS alignment) has been a top managerial concern for over 30 years and remains an ongoing research stream of key interest to the IS discipline." I wrote before about why IT-business alignment is still unsolved. This note asks what that history can teach the newer debate on AI alignment.
The first lesson is that alignment stopped being treated as a state. Benbya and colleagues describe how "researchers questioned the conception of IS alignment as a linear relationship in which alignment links a set of antecedents to organizational consequences". The static view, they write, "can lead to excessive rigidity and conditions of misalignment, because it does not substantially account for the possibility of a complex, dynamic and unpredictable competitive environment". The field moved toward a view of alignment that "represents an emergent process of dynamic interactions and continual adjustments between business and IS across multiple organizational dimensions (e.g., strategic, operational and social)".
AI alignment research is now making a similar move. Shen et al. (2025) write that "Existing alignment research has been primarily approached as an AI-centered, static, and unidirectional process." They argue for "a shift toward a bidirectional, interconnected mode of human-AI alignment", and their alternative will sound familiar to IS readers. "Rather than a one-time process or static goal, we should consider alignment as a continuous, evolving engagement between humans and AI, requiring constant reassessment and recalibration." The formal work points the same way. Zhuang and Hadfield-Menell (2020) give "necessary and sufficient conditions under which indefinitely optimizing for any incomplete proxy objective leads to arbitrarily low overall utility". They conclude that "we should view the design of reward functions as an interactive and dynamic process".
The second lesson is about incompleteness. Hadfield-Menell and Hadfield (2018) compare reward design with contracts. "Contracts in human relationships are usually, and maybe necessarily, incomplete." Human contracts work anyway, because "human contracting is supported by substantial amounts of external structure, such as generally available institutions (culture, law) that can supply implied terms to fill the gaps in incomplete contracts." Their advice is that "perfect reward specification is routinely not possible, so AI researchers should be focused on designing optimally in the face of irredeemable divergence between AI and human utility." The way I read the IS literature, it reached a similar place by a different road. A plan is never complete, so alignment is kept through adjustment and through the social side of the organization, not through a better plan. Benbya and colleagues report that researchers converged on three dimensions, "a) strategic, b) operational, and c) social", and in my reading the third is the one that fills the gaps that no plan can fill.
The third lesson is about meaning. Benbya and colleagues warn about "the established plurality of meanings embedded in the term 'alignment'". AI alignment has the same problem. Ji et al. (2023) use the word to mean making "AI systems behave in line with human intentions and values." Grote, Parker and Crowston (2026) use it for "proper alignment of control and accountability for the stakeholders involved in AI development and use." Ghasemaghaei (2026) studies "organizational-AI misalignment", the gap between what an organization says it values and what its AI system actually recommends. These are three different relationships under one word. The first is between a model and human values, the second is between control and accountability, and the third is between an organization's words and its system's outputs.
I do not think this is a problem to be solved by picking a winner. IS alignment research learned to live with plurality by naming the dimensions it meant. AI alignment research would benefit from doing the same, so that a paper on reward models and a paper on accountability in organizations do not talk past each other while using the same word.
The last lesson is the one I am least sure about. Grote and colleagues describe AI as "the first technology that is fundamentally reshaped by its use", because the systems keep learning from data produced while they are in use. If that is right, then alignment as continual adjustment is not just a good idea borrowed from IS. The technology forces it. Every deployment changes the system that the next round of alignment has to work with. IS research on alignment as a process has methods for studying exactly that kind of change over time, from longitudinal case studies to process theories that pay attention to the social side. I think those methods are more useful to AI alignment than our field usually gives itself credit for.
There is a connection here to the note on who the model is aligned to. If alignment is a process, then the stakeholders are not fixed either. The people whose values matter at training time are not the only people whose values matter once the system is in use.