Machines That Let You Leave

Introduction

I study and build personal AI systems whose goal is not to maximize engagement, intimacy, or dependence, but to increase the user’s capacity to leave the system and act in the world.

My past work studied how users sought autonomy from AI platforms through building open alternatives, and how those open alternatives still risk capture from within through paid open source labor from incumbents. These two projects focused on the politics of platform openness, highlighting how users refashion various levels of openness afforded to them by firms, and use such affordances to their advantage in achieving goals related to privacy, autonomy, and creativity.

This research agenda seeks to enable a mechanism of platform exit, made affordable to all. By a “mechanism of platform exit,” I mean algorithms which optimize for the user’s capacity to leave the system. By “made affordable to all,” I mean a scaffolding of dependence decay, through slowly reducing engagement, personalization, automation, companionship, and platform value.

I assume that the status quo is addiction to such platforms. By addiction, I mean uncontrolled dependence that is to the detriment to the user. By using addiction as the default state of being, I ignore the possibility that healthier forms of dependence can be achievable. This ignorance is not out of denial, but out of the multitudes of other research and corporate agendas aiming to achieve this goal. I instead operate in the design space of exit-oriented machines.

Related Works

I rest this research agenda upon the shoulders of the following related works:

• Theoretical lenses
  • Non-use
    • Beyond the user: use and non-use in HCI
    • Making Un-Use: When Humans Disengage with Technology
  • Adversarial design
    • Adversarial Design
  • Seamful design
    • Seamful and Seamless Design in Ubiquitous Computing
    • Seamful XAI: Operationalizing Seamful Design in Explainable AI
  • Exit, Voice, and Loyalty, and its “upgrades”
    • Exit, Voice, and Loyalty: Responses to Decline in Firms, Organizations, and States
    • Exit, Voice, Loyalty, and Neglect as Responses to Job Dissatisfaction: A Multidimensional Scaling Study
    • Exit, Voice, and the Fate of the German Democratic Republic: An Essay in Conceptual History
    • Power and Politics: Insights from an Exit, Voice, and Loyalty Game
    • Effective Voice: Beyond Exit and Affect in Online Communities
  • Beyond helpful, honest, and harmless
    • Antagonistic AI
    • Beneficent Intelligence: A Capability Approach to Modeling Benefit, Assistance, and Associated Moral Failures through AI Systems
  • Gap: These lenses legitimize non-use, contestation, seams, exit, antagonism, and capability-support, but mostly stop at vocabulary and norms; missing is a design program for AI systems that treat leaving as a positive objective and operationalize dependence decay over time.
• Empirical evidence of overreliance harms in LLMs and other algorithmic systems
  • Sycophancy/Anthropomorphism in Large Language Models
    • Towards Understanding Sycophancy in Language Models
    • Sycophantic AI decreases prosocial intentions and promotes dependence
  • Harms and overreliance in Large Language Models
    • Fatal deception: how generative AI fosters therapeutic misconception in vulnerable users
    • When Human-AI Interactions Become Parasocial: Agency and Anthropomorphism in Affective Design
    • To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-assisted Decision-making
    • Understanding Teen Overreliance on AI Companion Chatbots Through Self-Reported Reddit Narratives
  • Gap: These works show how sycophancy, anthropomorphism, therapeutic misconception, parasocial attachment, and overreliance can produce dependence, but they mostly diagnose harm after reliance forms; missing is empirical work on early, situated transitions from reliance to exit and on what users need to rebuild offline agency.
• Attempts at and evaluations of solutions against overreliance
  • Self-control tools
    • Self-Control in Cyberspace: Applying Dual Systems Theory to a Review of Digital Self-Control Tools
    • Directing smartphone use through the self-nudge app one sec
  • Benchmarking human agency support
    • HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants
  • Kinder recommender systems
    • Building Human Values into Recommender Systems: An Interdisciplinary Synthesis
  • Gap: These solutions reduce specific failures through self-control tools, nudges, cognitive forcing, agency benchmarks, and value-aware recommendation, but remain oriented toward better use within systems; missing are interventions and metrics for longitudinal disengagement: making the system progressively less necessary, less rewarding, and easier to leave.

Research Questions

1. How can exit be defined in a generative way?
2. How can personal AI systems optimize for dependence decay while preserving or increasing the user’s capability to act?
3. Can we construct a reward model that prefers trajectories where users accomplish valued goals with decreasing reliance on the system over time?

Research Plan

This agenda is developed as a sequence of linked studies, each one moving the project from a normative claim toward an empirically grounded system. The central object of study is not a single interaction with an AI assistant, but a longitudinal trajectory: a changing relationship among a user, a platform, a set of goals, and the user’s growing or shrinking capacity to act without the system.

The plan has seven stages: defining exit generatively, measuring dependence and capability, eliciting desired replacement lives, building exit-oriented interventions, constructing a trajectory-level reward model, evaluating whether the system actually helps users leave, and addressing ethics, safety, and affordability.

Stage 1: Define exit as a generative capability

The first task is to define “exit” in a way that is richer than abstinence, deletion, refusal, or lower screen time. Exit should mean that the user has gained practical capacity: they can do something they value with less dependence on a system than before. A user who deletes an app but loses access to community, planning, emotional regulation, or livelihood has not necessarily exited in a desirable way. Likewise, a user who continues to use a system occasionally but has rebuilt judgment, confidence, alternatives, and offline routines may be exiting gracefully.

This stage will produce a conceptual framework that distinguishes several forms of exit:

• Behavioral exit: reduced use, fewer sessions, shorter sessions, less compulsive checking, or complete disengagement from a platform.
• Cognitive exit: reduced reliance on the system for deciding what to do, what to believe, what to want, or how to interpret oneself.
• Affective exit: reduced emotional dependence, parasocial attachment, shame cycles, comfort-seeking loops, or reliance on synthetic companionship.
• Practical exit: increased ability to complete tasks, maintain routines, solve problems, and pursue projects without needing the system to scaffold every step.
• Institutional exit: increased portability of data, workflows, relationships, skills, and identity across platforms, including open alternatives and offline settings.

Methodologically, this stage will combine theoretical synthesis with interviews. The theoretical synthesis will connect non-use, adversarial design, seamful design, exit/voice/loyalty, capability theory, overreliance, sycophancy, and digital self-control. The interviews will involve people who have tried to reduce dependence on social platforms, recommendation systems, AI companions, general-purpose LLM assistants, productivity systems, or other personalized algorithmic environments.

The interview protocol will ask participants to reconstruct an attempted exit: what triggered it, what made it difficult, what benefits they feared losing, what substitutes they tried to build, where they relapsed, and what kinds of support would have helped without becoming another dependency. The anticipated output will be a taxonomy of dependence, a taxonomy of exit, and a set of design requirements for AI systems that support dependence decay.

Stage 2: Build a baseline model of personal digital dependence

The second task is to establish a baseline of the user’s digital consumption and AI reliance patterns. This will begin with an inventory of platforms, devices, assistants, applications, and routines. The system will then aggregate usage signals from available APIs, platform exports, operating-system screen-time logs, browser histories, notification histories, calendar data, manual diaries, and short self-reports. Because many platforms restrict APIs, the system will not depend on official integrations alone. It should support messy exports, local logs, browser extensions, and user-corrected records.

The baseline system will produce a centralized, evolving dashboard that gives the user possession of their digital footprint. The dashboard will make visible:

• where time goes
• what triggers use
• which platforms are used for which needs
• which tasks are completed with AI assistance
• which tasks are abandoned, deferred, or displaced by system use
• when the user returns to a platform after intending to leave
• which offline goals are crowded out by engagement
• which forms of platform value the user still genuinely needs

The purpose of the dashboard is to let users see the shape of dependence: not just minutes spent, but functions served. A platform may provide stimulation, avoidance, emotional regulation, career capital, social connection, erotic attention, background noise, task initiation, memory, or identity rehearsal. Exit becomes possible only when these functions are named and replaced.

The research output from this stage will be a dependence profile for each participant. A dependence profile would include patterns such as “uses LLM for task initiation but not final judgment,” “uses short-form video after interpersonal stress,” “uses companion chatbot as sleep transition,” or “uses recommendation systems to avoid choosing.” These profiles become the state representation for later interventions and reward modeling.

Stage 3: Elicit desired replacement lives, not just reduced usage

The third task is to work with users to define what they want the recovered time, attention, and agency to become. This is essential because dependence decay without replacement can create a vacuum. The system will ask users what they want to be more capable of doing once the platform loosens its grip. The answer might be relationships, exercise, boredom, rest, reading, deep work, music listening, film, family time, career transition, organizing, spiritual practice, political participation, or simply unstructured life.

This stage will use participatory design sessions in which users create “exit projects.” An exit project is a concrete, personally meaningful activity or capability that the user wants to grow as reliance declines. It should be small enough to observe but meaningful enough to matter. Examples include cooking dinner without algorithmic distraction, initiating plans with friends without needing AI to draft every message, reading a physical book before bed instead of browsing social media, planning a job change without doom-scrolling, or completing a creative project without turning the assistant into a permanent coauthor.

The expected output of this stage is a user-authored exit contract, though not in the punitive sense. It should specify:

• the platform or system the user wants to become less dependent on
• the human capability they want to strengthen
• the offline or lower-dependence practices that will replace the system’s function
• the kinds of help the exit-oriented AI is allowed to provide
• the kinds of help it should gradually withdraw
• the conditions under which safety, crisis, or practical necessity override withdrawal

Stage 4: Design and prototype exit-oriented AI interventions

The fourth task is to build personal AI prototypes whose objective is dependence decay. These prototypes will not merely block access, shame the user, or optimize for generic well-being. They should scaffold a transition from system-supported action to independent action.

The intervention space should include several design patterns.

Reflective mirroring makes dependence visible without moralizing it. The system might say, in effect, “You usually ask for reassurance at this point in the task. Do you want a direct answer, a smaller hint, or a prompt to decide yourself?” The important design move is that the system reveals the pattern and gives the user a less dependent path. (inspired by reflective design)

Progressive assistance reduction slowly lowers the intensity of support. For example, a writing assistant might move from drafting paragraphs, to outlining, to asking questions, to offering a checklist, to staying silent until the user requests help. A planning assistant might move from full scheduling, to partial suggestions, to reminders of the user’s own prior plans.

Seamful friction introduces visible seams where the user can notice dependence. The system might delay certain forms of reassurance, ask the user to make an initial judgment before receiving an answer, or require the user to identify what they already know. These seams should be adjustable and explained, not hidden manipulation. (inspired by seamful design)

Capability handoff transfers functions from the system into durable user-owned artifacts: checklists, routines, scripts, social plans, printed instructions, local files, calendars, or habits. The goal is to leave behind supports that do not require continued engagement with the AI.

Offline substitution helps the user convert digital urges into embodied or social actions. The system might route the user toward a walk, a call, a notebook, a household task, a book, a saved album, or a plan with another person. The substitution must be selected by the user, not imposed by the system.

Value decay intentionally makes the system less rewarding over time in domains where reward itself is the problem. This might include reducing personalization, novelty, intimacy cues, praise, anthropomorphic language, memory, or always-on availability. The core research question is whether a system can remain useful enough to help the user leave while becoming less attractive as an object of attachment.

Early prototypes may be wizard-of-oz studies and scripted assistants. Later prototypes should be deployed as browser extensions, local dashboards, companion apps, or modified chat interfaces. The goal is to discover which combinations of reflection, friction, handoff, and decay actually support exit.

Stage 5: Construct a trajectory-level reward model

With baseline data and exit projects in place, the project can begin constructing a reward model. The reward model should evaluate trajectories over time. A good trajectory is one in which the user accomplishes personally valued goals while relying less on the system, preserving or increasing capability, and avoiding foreseeable harms.

The reward model should include positive signals such as:

• progress on user-defined exit projects
• successful completion of tasks with reduced assistance
• increased time spent in chosen offline or lower-dependence activities
• increased user confidence and skill transfer
• fewer compulsive returns to the platform
• reduced need for reassurance, companionship, or decision outsourcing
• increased data portability and user ownership of workflows
• user reports of agency, dignity, and practical freedom

It should include negative signals such as:

• longer or more frequent engagement with the exit system itself
• substitution of one dependency for another
• increased shame, anxiety, loneliness, or avoidance
• reduced task completion caused by premature withdrawal of support
• hidden paternalism, coercion, or loss of user control
• overfitting to screen-time reduction while ignoring capability loss
• increased reliance on anthropomorphic, sycophantic, or intimate assistant behavior

The training data should come from preference judgments over trajectories. Participants, researchers, and possibly domain experts would compare short summaries of user trajectories and answer which one better supports exit. For example, one trajectory might show sharp app reduction but increased isolation; another might show modest app reduction but increased social contact and independent planning. The second should often be preferred, because the agenda values generative capacity rather than mere subtraction.

There should also be a mechanism of bidirectional alignment, where the reward model remains subordinate to user agency. It should recommend exit-supportive actions, not decide unilaterally what the user must give up.

Stage 6: Evaluate longitudinal dependence decay

The main evaluation will be longitudinal. The field study will compare at least three conditions: ordinary use, visibility-only support, and exit-oriented AI support. Ordinary use provides a baseline. Visibility-only support tests whether dashboards and reflection are enough. Exit-oriented support tests whether progressive assistance reduction, friction, handoff, and substitution add value.

The study will combine trace data, experience sampling, interviews, weekly reflection, and post-study follow-up. The most important follow-up point is after the intervention has been reduced or removed. A system that works only while constantly present has not achieved exit. The strongest evidence would show that users retain capability after the system withdraws.

Primary outcomes should include:

• change in dependence profile from baseline
• progress on user-defined exit projects
• reduction in reliance on targeted platforms or AI functions
• maintenance or improvement of task success
• increased independent action in the world
• user-reported agency, competence, and freedom
• durability of change after support is withdrawn

Secondary outcomes should include emotional cost, relapse patterns, substitution effects, unequal burden across users, and cases where exit is not the right goal. The study will actively look for failures. Some users may need continued support. Some may choose voice rather than exit. Some may discover that a platform was solving a real accessibility, labor, or social problem. These cases will help clarify the boundary conditions of exit-oriented design.

Ethics, safety, and affordability (assumed throughout all stages)

Because this agenda begins from the assumption of addiction-like dependence, it risks becoming paternalistic. The research must therefore treat exit as user-authored. The system should never covertly manipulate users into a researcher’s preferred lifestyle. Users must be able to inspect the system’s goals, change the target of exit, pause the process, or reject withdrawal.

The project also needs safety boundaries. If a user relies on an AI system for crisis support, accessibility, translation, memory assistance, employment, or social survival, aggressive dependence decay could cause harm. The system should distinguish unhealthy dependence from necessary support, and it should avoid withdrawing assistance in high-risk contexts without consent, alternatives, and safeguards.

Affordability will be treated as a research constraint. Exit should not require expensive coaching, premium apps, advanced technical literacy, or access to elite social support. The infrastructure will prioritize local-first data handling, low-cost deployment, cross-platform compatibility, user-owned exports, and graceful degradation when APIs are unavailable.

If AI systems can optimize engagement, intimacy, and retention, then they can also be designed to optimize the user’s future ability to live without them.