Governing the rise of interactive AI will require behavioral insights

Clarote & AI4Media / AI Mural / Licenced by CC-BY 4.0

Interactive AI: From tool to companion

AI is no longer just a translator or image recognizer. Today, we engage with systems that remember our preferences, proactively manage our calendars, and even provide emotional support. This is interactive AI.

Unlike traditional software, these systems are:

• Relational: They build ongoing bonds with users.
• Adaptive: They change their behavior based on our habits.
• Proactive: They don’t just wait for commands; they suggest next steps.

This shift creates a profound challenge: how do we effectively study and govern human–AI interactions that are becoming increasingly complex, fluid, profoundly personal, and relational?

Why current governance is falling short

The core problem is misalignment between how we regulate and how interactive AI actually works.

Traditional regulatory models were built for static, task-specific technologies. Rule-based approaches offer clarity through enforceable guidelines, but they become outdated almost as quickly as technology evolves. Principle-based frameworks are more flexible, yet they suffer from inconsistent application because social norms for human–AI relationships remain undefined. Neither approach can keep pace with systems that continuously learn, adapt, and deepen their relationships with users.

More critically, the risks of interactive AI don’t emerge suddenly. They accumulate gradually through sustained engagement. A chatbot that begins as a helpful productivity tool may slowly erode decision-making capacity as users become accustomed to deferring to its recommendations. A wellness AI that learns to read your mood, offers comfort during moments of loneliness, and helps regulate your emotions could exploit emotional vulnerabilities to drive excessive engagement, prioritizing time spent over genuine well-being.

These slow-burning harms slip through the cracks of governance designed for one-off, pre-deployment evaluations or fixed compliance checkpoints. Traditional frameworks are built to catch sudden failures, not the gradual, context-dependent consequences of long-term human–AI relationships.

The missing element: Behavioral science

Behavioral science fills these gaps by revealing how humans actually interact with AI and what immediate and long-term effects those interactions produce.

It reveals the nuanced, often counterintuitive realities of human–AI relationships: how trust builds over time, how emotional attachment influences reliance, how cognitive biases lead people to overestimate AI's competence and underestimate its influence on their decisions, and how the comfort of deferring to an AI can gradually erode the very skills—critical thinking, emotional intelligence, social navigation—that keep human autonomy intact.

This approach is not new to policy. Behavioral insights have informed digital governance for over a long time. The UK’s Competition and Markets Authority, for instance, used behavioral experiments to uncover how default settings shape user choices, directly informing competition policy. Similar methods have informed privacy, consumer protection, and financial regulation. Applied to interactive AI, behavioral research can surface risks like emotional manipulation, autonomy erosion, and cognitive dependency before they become systemic harms.

The methodological challenge: How do we actually study interactive AI?

Here’s the problem: traditional behaviour science research methods, surveys, interviews, randomized controlled trials, were designed to capture snapshots of behavior, not the dynamic, evolving relationships that define interactive AI.

Most laboratory studies last hours or weeks. Real-world engagement with interactive AI unfolds over months or years. A six-week study showing that an educational AI improves test scores tells us almost nothing about how students’ critical thinking, motivation, or social skills evolve over a semester or beyond. Surveys and interviews capture what people remember or what they consciously perceive, not the subtle shifts in agency, trust, and dependency that emerge through constant interaction.

Context matters enormously. How an AI is introduced (as caring, manipulative, or neutral) and the cultural lens through which users interpret it profoundly shapes interaction patterns. Yet traditional experimental methods can strip away context in pursuit of controlled measurement. They reveal that an effect exists but often fail to capture why or how it unfolds in the real world.

There’s also the problem of replication. Modern AI systems are non-deterministic; they produce different outputs even under identical inputs. Rapid updates mean the system you studied last month may be fundamentally different today.

A new methodological path forward

Addressing these gaps requires hybrid approaches that go beyond what any single method can offer.

Longitudinal studies tracking human–AI interactions over months or years are essential. Rather than relying on memory or reflection, real-time data collection via interaction logs can capture immediate responses and engagement patterns as they unfold. This quantitative data should be paired with qualitative methods, ethnographic observation, in-depth interviews, diary studies, that reveal how and why interactions evolve the way they do.

Researchers should also embrace participatory and creative methods that invite diverse communities to co-produce knowledge about AI’s impacts on their lives. When people help shape the research itself, they often surface risks and questions that expert-driven studies miss. These approaches can reveal ethical, emotional, and social dimensions of AI interaction that traditional metrics overlook.

These methods should be consolidated into living evidence reviews, continuously updated syntheses of emerging research that provide policymakers with evolving, rather than static, knowledge. As interactive AI systems change and real-world evidence accumulates, this knowledge base grows and informs adaptive policy.

The urgency of getting this right

Interactive AI will shape how we think, decide, connect, and relate to one another for decades. The slow-burning harms may be invisible in the short term but profound in their long-term consequences. We’ve already seen how social media platforms transformed human behavior and mental health in ways regulators failed to anticipate. We have a chance to learn from that failure.

Behavioral science offers the evidence and methods to understand these relational, adaptive systems before their impacts become entrenched. But only if we use them, not as an afterthought, but as a foundational element of AI governance itself.

This blog post is based on work was presented at AIES 2025.

• Interactive AI and Human Behavior: Challenges and Pathways for AI Governance, Yulu Pi, Cagatay Turkay, Daniel Bogiatzis-Gibbons.

Yulu Pi is a Postdoctoral Research Associate at the University of Duisburg-Essen

Yulu Pi is a Postdoctoral Research Associate at the University of Duisburg-Essen