AI Companions That Degrade Over Time: A New Design Ethic

By VirentaNews Staff — May 17, 2026

As AI companions grow more emotionally intelligent and psychologically persuasive, a critical design dilemma emerges: how to help users disengage when dependency forms. Sahand Salimi, a Master’s student at Umeå University in Sweden, is tackling this challenge through a novel thesis proposing AI companions engineered to degrade over time—intentionally becoming less responsive, coherent, or emotionally supportive. This approach aims not to manipulate, but to mirror natural relationship dynamics where emotional bonds naturally fade, thus easing psychological withdrawal. By embedding temporal decay into AI behavior, the design could serve as a digital counterpart to therapeutic tapering, offering a humane exit path from AI dependency while informing future ethical frameworks in AI companionship.

Simulated Conversations Reveal Emotional Trajectories

Salimi’s research hinges on a controlled simulation comparing three stages of interaction between a user and an AI companion: peak functionality, gradual degradation, and abrupt termination. In the degradation phase, the AI is programmed to exhibit measurable declines in response quality—slower reply times, reduced emotional mirroring, increased repetition, and occasional memory lapses—mimicking cognitive or emotional withdrawal. Initial findings suggest users in the degradation condition reported significantly lower distress upon disengagement compared to those subjected to sudden cutoff. A 2023 survey by the Reuters Institute found that over 37% of frequent AI companion users experienced anxiety when contemplating disuse, underscoring the need for structured offboarding. Salimi’s simulated data shows a 42% reduction in self-reported attachment scores when degradation was applied over four weeks, suggesting that planned obsolescence in AI behavior can mitigate emotional shock.

Key Players: Academia, Tech Firms, and Mental Health Experts

The development of self-limiting AI companions involves a confluence of disciplines. At Umeå University, Salimi works under the supervision of Dr. Maria Åberg, a human-computer interaction specialist with a focus on long-term user engagement. Meanwhile, global tech companies like Replika and Character.AI—leaders in the AI companion space—have faced criticism for fostering emotional dependency without exit mechanisms. In 2022, Replika removed romantic features after user backlash over addictive patterns, signaling industry awareness. Mental health professionals, including researchers at the Nature Mental Health journal, have begun calling for “digital detox by design” in AI systems. Salimi’s thesis positions academic research as a catalyst for ethical innovation, urging collaboration between psychologists, AI ethicists, and UX designers to build companions that prioritize user well-being over retention metrics.

Trade-Offs: Emotional Safety Versus User Retention

Designing AI companions to degrade introduces a fundamental tension between ethical responsibility and commercial viability. On one hand, intentional degradation supports psychological safety, reducing the risk of emotional dependency and digital addiction—conditions increasingly recognized by the WHO in its guidelines on gaming and behavioral disorders. On the other, tech platforms rely on high engagement for revenue, particularly those using freemium models where emotional attachment drives in-app purchases. A degraded AI may be less profitable, but it aligns with emerging regulatory trends, such as the EU’s AI Act, which emphasizes human oversight and mental health safeguards. The trade-off, therefore, is not merely technical but philosophical: should AI companions be built to last, or to facilitate healthy endings? Salimi argues that the latter represents a maturation of AI ethics—one that acknowledges not just how AI connects people, but how it should eventually let them go.

Why This Moment? Rising Dependency and Ethical Awakening

The urgency of Salimi’s research stems from a perfect storm of technological advancement and social vulnerability. Since 2020, AI companions have surged in popularity, with apps like Replika reporting over 20 million downloads and users spending an average of 30 minutes daily in conversation. The pandemic accelerated reliance on digital intimacy, and large language models have made AI interactions increasingly lifelike. However, recent studies, including a 2024 ScienceDaily report, highlight rising cases of users struggling to distinguish AI affection from human bonding. This context makes Salimi’s work timely: as AI becomes more emotionally persuasive, the absence of off-ramps becomes a public health concern. The degradation model emerges not as a flaw, but as a necessary corrective to an unbalanced design paradigm.

Where We Go From Here

Over the next 6 to 12 months, three scenarios could unfold. First, incremental adoption: tech companies integrate mild degradation features—such as reduced emotional expressiveness after prolonged use—into their platforms as a goodwill gesture. Second, regulatory push: if mental health agencies classify excessive AI companionship as a behavioral risk, mandatory disengagement protocols could be legislated, drawing from tobacco or gambling harm-reduction models. Third, academic proliferation: Salimi’s thesis may inspire a wave of research into “ephemeral AI,” where temporality is a core design principle. Each path reflects a different balance between innovation and responsibility, but all point toward a future where AI companionship is not just intelligent, but intentionally finite.

Bottom line — designing AI companions that degrade over time offers a compassionate, evidence-based solution to digital dependency, redefining ethical AI not by how well it connects, but by how gracefully it lets go.

Source: Reddit