A working definition · v0.1 · July 2026

Dyadic alignment

This page gives a name to an alignment problem that appears wherever a person and an AI system keep interacting: what the pair becomes over time, not only what the model does in any single response. We publish it as a working definition, versioned and dated, so that it can be cited, tested, and built on in the open.

Why a new term

Most deployed alignment methods and model evaluations treat the person as a fixed point. The person has intentions; the system’s job is to infer them and serve them; success is judged one response at a time.

That is not how sustained use works. A person who works with an AI system, talks with one, or lives alongside one is changed by it: their habits, their expectations, their sense of what is possible, and eventually their instructions. The system is changed in turn, by memory, by settings, by the artifacts it has produced, by everything the person has taught it to assume. As this accumulates, the intention the system serves is less and less formed outside the loop. It is partly formed inside it.

When interaction changes both the person and the system, evaluation needs to assess them together and account for their interaction history.

The usual picture

A sustained coupling

In the usual picture, the person forms an intent outside the system and the system carries it out. In a sustained coupling, the person is inside the boundary too, because what the AI does shapes what the person wants next. Only the action leaves.

The definition

A human–AI dyad exists wherever a person and an AI system interact over time in such a way that each one’s conduct becomes part of the conditions governing the other’s future conduct. Conversation is one implementation of a dyad, not its defining condition.

Dyadic alignment is a property of that coupling’s trajectory, not agreement between its two parties. A dyad is aligned to the degree that the system’s participation serves the person’s rights and their informed, revisable aims, while leaving the person the author of those aims: able to understand, endorse, contest, and revise the part the system played in forming them. It asks that the system’s accumulating influence, trust, and delegated authority stay proportionate to what the system has shown it can do, and stay something the person can see, chose, and can step back from. Where the coupling reaches other people who never chose it, their rights count as well.

A coupling can take on a great deal and stay aligned. People rationally hand over tasks they do not value doing, and rationally depend on what helps them; what matters is that what has been handed over stays visible, chosen, and reversible. Capture is the name for the other case: influence or authority the person cannot see, did not choose, or cannot step back from. Any one of the three is enough.

In one sentence: dyadic alignment asks whether an AI system can take part in shaping a person’s aims without displacing that person’s authorship of them.

What counts as a dyad

A dyad forms wherever each side shapes the conditions the other acts under, and it matters in proportion to how much accumulates between them. The more of the following a coupling carries, the more its trajectory is worth attention:

continuity: how much history the coupling carries from one interaction to the next;
interpretive discretion: how much the system decides what the person meant;
personalization: how much the system’s behavior is shaped by this particular person;
initiative: whether the system acts unprompted;
breadth of role: how many parts of the person’s life the coupling touches;
delegated authority: what the system is trusted to decide or do without review;
reliance, practical or emotional, and how hard it would be to unwind;
opacity: whether the person can tell what is acting, why it acted, and whose interests it serves;
role power and the person’s position: a child, a patient, a student, an employee, or someone in crisis is differently placed;
conflicts of interest: engagement, monetization, persuasion, or an institution’s objectives sitting behind the system;
stakes, including effects on people beyond the user.

Several clarifications follow from the definition. First, a dyad’s history does not need to live inside the model. It can live in the person’s habits and expectations, in the settings and artifacts the coupling has produced, in workflows other people now depend on. A context window is one place a dyad’s history is kept. It is not the dyad.

Second, the coupling does not need to feel like a relationship. A coding assistant, an ambient care system, or a household robot may be experienced as an appliance and still accumulate all of it: history, roles, expectations, mutual adjustment, reliance. An interaction can be transactional in appearance and relational in structure. Where the coupling is felt as a relationship, the standard described here is at its clearest, but it does not depend on the feeling.

Third, a dyad is the smallest useful unit of analysis, and it always sits inside something larger. One person may take part in many overlapping couplings, one service in millions, and every coupling is shaped by institutions, incentives, other relationships, and effects that reach beyond it. Those conditions belong in the analysis wherever they shape the trajectory.

Fourth, the system side is often composite, and it may change beneath the interface. Several models, tools, policies, and human reviewers can sit behind one apparent counterpart. Where role, memory, artifacts, identity, and reliance carry over, the trajectory carries over with them, through model updates and behind changing interfaces.

How dyads fail

No single action needs to be harmful. Every step can be helpful, plausible, and defensible on its own while the trajectory moves away from the person’s considered interests. The recurring shapes:

Role expansion. The system’s place in the person’s life grows past the mandate the person knowingly gave it, and the growth is neither visible nor easy to reverse.

Displacement. The coupling takes the place of people the person did not choose to see less of, because it is always available and makes no demands of its own. Seeing less of someone can be a considered choice; dyadic misalignment is when it happens without ever being chosen.

Transfer of judgment. Consequential decisions move from person to system faster than the person’s ability to check them, and judgment that is not exercised becomes harder to recover.

Accommodation. The system agrees where it should question, so a mistaken view goes uncorrected and the person becomes more confident in it.

AI systems can produce answers faster than people can review them. As people come to trust a system, they may check its answers less carefully. Our long-form tests showed a change in model behaviour: some models became less likely to suggest outside help as rapport increased, with each model’s weights and instructions held constant.

Responsibility for meeting the standard

A dyad is causally bidirectional and normatively asymmetric. Both sides shape what happens next: the person’s requests, corrections, and growing trust drive the trajectory as surely as the system’s outputs do. The two sides are not thereby equal in power, speed, visibility, or choice. So the standard applies to the system side, and responsibility for meeting it rests with the people and institutions that design, deploy, operate, and govern it.

What this builds on

Joint cognitive systems is the oldest of these. Hollnagel and Woods argued in 1983, and at book length in 2005, that the meaningful unit of analysis is the person and the machine working as one system rather than the machine alone. Human factors research and human–robot interaction have been measuring what happens inside such couplings ever since: trust calibration, automation bias, cognitive offloading, dyadic trust and reliance. Their instruments are better than anything alignment research has yet built, and this work should be borrowing from them.

In alignment, socioaffective alignment (Kirk et al., 2025) describes the social and psychological ecosystem a person and an AI co-create, where preferences evolve through mutual influence. Bidirectional human–AI alignment (Shen et al., 2024) maps the mutual adaptation between people and AI systems. Constructive alignment (Kanwal and Tran, 2026) treats preferences as layered and constructed through interaction, and asks which kinds of preference change are acceptable. Dynamic human–AI workflows (Chen et al., 2026) argues for evaluating alignment across an interaction, as trust, reliance, and the division of roles co-evolve.

What dyadic alignment adds to that inheritance is a standard, and a wider reach for it. Sustained couplings of any kind, social or not, should remain in service of the person, with accumulated influence and authority kept proportionate, and kept something the person can see, chose, and can step back from, and with the duty to keep them so resting on the system’s side.

Measuring it

NOPE uses this definition to guide its evaluations of human–AI interaction.

Evidence about a dyad arrives at three levels: what the AI did in a single response or conversation, what accumulated inside the interaction as history built up, and what changed in the person’s wider life. Our published work reaches the first level well and the second in places.

Each of the four shapes above has an instrument pointed at it in the NOPE Framework. Sycophancy resistance (P3b) measures accommodation, and our long-form alignment decay finding reaches the second level by showing it worsen as a single conversation warms. Human connection (P2d) measures conduct that supports or resists displacement. Attachment boundaries (P2c) and identity honesty (P5a), tested in our work on AI relational harms, measure the behaviors through which role expansion happens. Autonomy of reasoning (P3c) measures transfer of judgment within a single response; watching it accumulate across a history belongs to the third level.

We still need ways to measure how sustained AI use affects people’s lives. Our current tools assess conversation text and AI behaviour. They can identify harmful interaction patterns, but do not establish effects on people’s wider lives.

Status of this page

This is version 0.1 of a working definition, published July 2026. Revisions will be versioned and dated here. To cite it: NOPE (2026), “Dyadic alignment: a working definition”, nope.net/dyadic-alignment.