{
  "title": "Measuring Reversibility in AI Systems: An Empirical Program for Affordable Variation and Exit",
  "short_title": "Measuring Reversibility in AI Systems",
  "subtitle": "An Empirical Program for Affordable Variation and Exit",
  "author": "Raynor Eissens",
  "series": "Ambient Era Research",
  "publication_type": "Empirical Program",
  "date_published": "2026-09-26",
  "format": "independent HTML article",
  "url": "https://research.madeby.re/measuring-reversibility/",
  "has_zenodo_record": false,
  "doi": null,
  "peer_reviewed": false,
  "abstract": "Artificial intelligence is making generation, inference, and\nsoftware-mediated action dramatically cheaper. Yet falling execution\ncosts do not imply that technological systems become easier to leave,\nreplace, modify, or recombine. A system may offer nearly free generation\nwhile simultaneously accumulating identity, memory, workflow state,\ncredentials, social dependencies, and institutional commitments that are\nexpensive to move. This paper proposes reversibility as a measurable\nproperty of AI systems: the degree to which a person or organization can\nchange, abandon, replace, or recombine an AI-mediated arrangement while\npreserving important state and avoiding disproportionate irreversible\nloss. It develops an empirical framework around five measurable dimensions: exit cost, state portability, attentional burden, irreversible\nresidue, and variation cost . Together they operationalize two\nconcepts from the Ambient Era framework: reversible\nchoice , the ability to change technological arrangements\nwithout prohibitive loss, and affordable variation , the\nability to generate and test meaningfully different viable alternatives\nwithout transferring the cost of complexity back onto human attention.\nThese concepts were previously formulated as part of a broader\nhypothesis in which scarcity migrates from execution toward attention,\ncoherence, trust, and exit as intelligence becomes abundant. The paper proposes controlled migration experiments, longitudinal\nfield studies, protocol-level measurements, and failure-recovery tests\ncapable of comparing agent-bound systems with more portable,\nprotocol-mediated architectures. It does not assume that ambient or\ninfrastructural AI is preferable or inevitable. The central empirical\nquestion is narrower: as AI becomes cheaper to use, does\nmeaningful technological choice become cheaper to reverse? The answer is not currently known.",
  "central_question": "As AI becomes cheaper to use, does meaningful technological choice become cheaper to reverse?",
  "dimensions": [
    "exit cost",
    "state portability",
    "attentional burden",
    "irreversible residue",
    "variation cost"
  ],
  "key_proposition": "Cheap generation and cheap exit are independent variables.",
  "sections": [
    {
      "heading": "Abstract",
      "text": "Artificial intelligence is making generation, inference, and\nsoftware-mediated action dramatically cheaper. Yet falling execution\ncosts do not imply that technological systems become easier to leave,\nreplace, modify, or recombine. A system may offer nearly free generation\nwhile simultaneously accumulating identity, memory, workflow state,\ncredentials, social dependencies, and institutional commitments that are\nexpensive to move.\n\nThis paper proposes reversibility as a measurable\nproperty of AI systems: the degree to which a person or organization can\nchange, abandon, replace, or recombine an AI-mediated arrangement while\npreserving important state and avoiding disproportionate irreversible\nloss.\n\nIt develops an empirical framework around five measurable dimensions: exit cost, state portability, attentional burden, irreversible\nresidue, and variation cost . Together they operationalize two\nconcepts from the Ambient Era framework: reversible\nchoice , the ability to change technological arrangements\nwithout prohibitive loss, and affordable variation , the\nability to generate and test meaningfully different viable alternatives\nwithout transferring the cost of complexity back onto human attention.\nThese concepts were previously formulated as part of a broader\nhypothesis in which scarcity migrates from execution toward attention,\ncoherence, trust, and exit as intelligence becomes abundant.\n\nThe paper proposes controlled migration experiments, longitudinal\nfield studies, protocol-level measurements, and failure-recovery tests\ncapable of comparing agent-bound systems with more portable,\nprotocol-mediated architectures. It does not assume that ambient or\ninfrastructural AI is preferable or inevitable. The central empirical\nquestion is narrower: as AI becomes cheaper to use, does\nmeaningful technological choice become cheaper to reverse?\n\nThe answer is not currently known."
    },
    {
      "heading": "1. Introduction",
      "text": "The economics of artificial intelligence are changing rapidly.\n\nThe Stanford AI Index reported that the cost of querying a model\nperforming at roughly GPT-3.5 level on MMLU fell from approximately $20\nper million tokens in November 2022 to $0.07 by October 2024, a decline\nof more than 280-fold in roughly eighteen months. Depending on task and\ncapability level, inference prices have fallen by orders of\nmagnitude.\n\nThis development encourages a straightforward intuition: if\nintelligence becomes cheaper, technological choice should become\neasier.\n\nThat conclusion does not necessarily follow.\n\nAn AI service may be inexpensive to invoke while becoming expensive\nto leave. A user may accumulate years of conversational memory in one\nprovider, tool permissions in another, workflow automations in a third,\ncredentials in a fourth, and persistent agent state in a fifth. An\norganization may discover that generating an alternative workflow costs\npennies while migrating its accumulated context, audit trails,\npermissions, evaluation history, institutional knowledge, and human\nhabits costs months.\n\nThe economically significant transition may therefore not be from\nexpensive intelligence to free intelligence. It may be from expensive execution to cheap execution while other forms of\nscarcity remain or intensify .\n\nThe Ambient Era framework describes one version of this transition.\nAs execution becomes abundant, bottlenecks may migrate toward attention,\nselection, trust, coherence, differentiation, and the ability to exit\ntechnological arrangements without disproportionate loss. Within that\nframework, affordable variation means more than\nproducing many alternatives. Alternatives must be meaningfully viable,\ninexpensive to test, and sufficiently reversible that humans do not\nbecome permanent compensators for the complexity created by abundant\ngeneration.\n\nThis paper isolates that claim from the broader theory and makes it\ntestable.\n\nThe key proposition is:\n\nA system can score highly on one and poorly on the other.\n\nThat distinction creates an empirical research program."
    },
    {
      "heading": "2. From Abundance to\nReversibility",
      "text": "Most measures of AI progress concern capability.\n\nThey measure benchmark performance, latency, inference price,\nparameter efficiency, energy consumption, task completion, or autonomous\noperating duration. These are important variables, but they primarily\nmeasure what a system can do and what it costs to make\nit do it.\n\nThey rarely measure what happens when the user wants to stop.\n\nConsider two hypothetical systems.\n\nSystem A can generate a working application for €0.20. Over a year it\naccumulates proprietary memory, undocumented workflow state,\nprovider-specific tool integrations, and credentials that cannot be\nexported intact.\n\nSystem B costs €2.00 to perform the same generation. Its memory,\nauthorization relationships, workflow definitions, provenance, and\nartifacts can be exported in interoperable formats and reconstructed\nelsewhere within minutes.\n\nBy an execution-cost metric, A is ten times cheaper.\n\nBy an exit-cost metric, B may be dramatically cheaper.\n\nThe distinction recalls Albert Hirschman’s separation between exit and voice as responses to\ndeteriorating organizations and products. Exit is meaningful only when\nleaving is a practical option rather than a nominal one. The concept was\noriginally institutional and economic rather than computational, but the\nunderlying problem maps naturally onto increasingly persistent AI\nsystems.\n\nThe empirical question is therefore not merely:\n\nHow inexpensive is intelligence?\n\nIt is also:\n\nHow inexpensive is it to change your mind about the system\nproviding it?"
    },
    {
      "heading": "3. Definition of Reversibility",
      "text": "For the purposes of this research program:\n\nReversibility is not binary.\n\nDeleting an application is easy. Reconstructing the state accumulated\nthrough that application may not be.\n\nLikewise, data export alone does not establish reversibility. A\nfifty-gigabyte archive that no competing system can interpret provides\ntechnical extraction without practical exit.\n\nReversibility therefore has at least five dimensions:\n\nExit Cost State Portability Attentional Burden Irreversible Residue Variation Cost\n\nThese dimensions should initially be measured separately rather than\ncollapsed prematurely into a single universal score."
    },
    {
      "heading": "4. Exit Cost",
      "text": "4.1 Definition\n\nExit Cost (EC) is the total resource expenditure\nrequired to move from one AI-mediated arrangement to a viable\nalternative.\n\nIt includes more than monetary switching fees.\n\nA first operational model is:\n\nE C = C m + C t + C s + C i + C l + C h EC = C_m + C_t + C_s + C_i + C_l + C_h\n\nwhere:\n\nC m C_m = direct monetary migration cost C t C_t = time required for migration C s C_s = state reconstruction cost C i C_i = identity and authorization reconstruction cost C l C_l = capability or functionality loss C h C_h = required human intervention\n\nBecause these components have different units, empirical studies\nshould report them individually and use normalized composite indices\nonly after domain-specific calibration.\n\n4.2 Example measurement\n\nA participant uses AI System A for eight weeks.\n\nAt a randomly assigned point, the participant must move to System\nB.\n\nResearchers measure:\n\nelapsed migration time; number of manual operations; percentage of relevant history successfully transferred; percentage of integrations restored; credentials requiring reauthorization; workflows requiring reconstruction; direct cost; task-performance degradation following migration; subjective difficulty.\n\nThe experiment produces an observable exit cost rather than an\nabstract claim about openness."
    },
    {
      "heading": "5. State Portability",
      "text": "Persistent AI increasingly depends on state.\n\nState may include:\n\nconversation history; long-term memories; preferences; files; embeddings; tool connections; workflow definitions; agent plans; task histories; provenance; authorization relationships; evaluation records; learned procedures.\n\nThe importance of portable identity and credentials is already\nvisible outside consumer AI. The W3C Verifiable Credentials 2.0\nspecifications formalize machine-verifiable credential exchange between\nissuers, holders, and verifiers. SPIFFE similarly specifies\ninteroperable workload identities designed to operate across\nheterogeneous computing environments.\n\nThese technologies do not prove that AI systems are becoming\nreversible. They demonstrate that some components traditionally bound to\nparticular applications or machines can be separated into\ninfrastructure-level representations .\n\n5.1 State Portability Ratio\n\nA simple experimental measure is:\n\nS P R = usable critical state after migration critical state before migration SPR = \\frac{\\text{usable critical state after migration}}\n{\\text{critical state before migration}}\n\nThe numerator should measure functional recovery ,\nnot merely exported bytes.\n\nIf a system exports 100 percent of its memory but only 30 percent can\nbe meaningfully reconstructed by the destination system, its effective\nportability is 0.30 rather than 1.00.\n\nResearchers should ask participants before migration which state they\nconsider essential. This prevents the provider or researcher from\ndefining success solely around what happened to be technically\nexportable."
    },
    {
      "heading": "6. Irreversible Residue",
      "text": "AI systems do not merely store data. They generate consequences.\n\nAn automated action may:\n\ncreate accounts; send messages; make purchases; change permissions; modify production systems; train downstream models; create social commitments; generate public artifacts; alter organizational procedures.\n\nSome consequences cannot be undone simply by switching providers.\n\nThis paper calls the accumulation of such consequences irreversible residue .\n\nAn initial measure could record, during a fixed evaluation\nperiod:\n\nnumber of externally consequential actions; proportion automatically reversible; proportion manually reversible; proportion effectively irreversible; expected cost of remediation; number of downstream dependencies created.\n\nThis distinction is increasingly important as agents acquire external\ntool access. Interoperability protocols such as MCP explicitly connect\nAI systems to external tools and data, while A2A is designed to enable\nindependent agents from different vendors or frameworks to collaborate.\nThese protocols increase composability, but composability alone does not\nguarantee reversibility.\n\nA system may become easier to connect while simultaneously becoming\ncapable of creating more difficult-to-reverse consequences.\n\nThat tension is empirically testable."
    },
    {
      "heading": "7. Attentional Burden",
      "text": "A system should not count as highly reversible if maintaining\nreversibility requires continuous human vigilance.\n\nSuppose System A automatically preserves exportable state but forces\nthe user to classify every interaction, approve dozens of\nsynchronization requests, maintain compatibility files, and continuously\nverify replicas.\n\nTechnically, the system is portable.\n\nPractically, humans are carrying the complexity.\n\nThis motivates Attentional Burden (AB) .\n\nPossible measures include:\n\nnumber of required user interventions; verification actions per task; interruption frequency; migration decisions requiring manual resolution; total active attention time; error-correction workload; subjective cognitive-load measures.\n\nThe Ambient Era framework places particular emphasis on bounded human\nattention because cheap generation can otherwise externalize\ncomputational abundance into human selection work.\n\nThe relevant design question is therefore not simply whether a system\ncan produce alternatives.\n\nIt is whether people can make use of those alternatives without becoming the integration layer themselves ."
    },
    {
      "heading": "8. Affordable Variation",
      "text": "Cheap variation is not the same as affordable variation.\n\nGenerative systems can already produce vast numbers of options. The\ncost of producing another paragraph, interface, image, workflow,\nsoftware component, or agent plan may approach negligible levels.\n\nBut an alternative is economically meaningful only if it can be\nevaluated and adopted.\n\nDefine a viable variation as an alternative\nthat:\n\ndiffers materially from the baseline; satisfies a predefined minimum performance threshold; can be evaluated within the experimental environment; does not require disproportionate irreversible commitment merely to\ntest.\n\nAn empirical Affordable Variation Rate (AVR) could\nthen be estimated as:\n\nA V R = number of viable alternatives tested total monetary + human evaluation cost AVR =\n\\frac{\\text{number of viable alternatives tested}}\n{\\text{total monetary + human evaluation cost}}\n\nA stronger formulation would penalize irreversible commitment:\n\nA V R R = N v C g + C e + C h + C r AVR_R =\n\\frac{N_v}\n{C_g + C_e + C_h + C_r}\n\nwhere:\n\nN v N_v = viable alternatives evaluated; C g C_g = generation cost; C e C_e = evaluation cost; C h C_h = human attention cost; C r C_r = expected cost of reversing commitments created during testing.\n\nThe important conceptual move is simple:"
    },
    {
      "heading": "9. Architecture as an\nIndependent Variable",
      "text": "The proposed empirical program compares different architectures\nrather than presuming that one architecture is superior.\n\nAt minimum, researchers should distinguish:\n\n9.1 Agent-bound architecture\n\nImportant state and functionality remain closely coupled to a\npersistent agent, application, or vendor environment.\n\n9.2 Protocol-mediated\narchitecture\n\nTools, credentials, data, state, or execution relationships are\nexposed through standardized or independently replaceable\ninterfaces.\n\n9.3\nInfrastructure-mediated architecture\n\nContinuity, authorization, storage, orchestration, or identity\nincreasingly exist outside the currently active agent instance.\n\nThese categories are ideal types rather than mutually exclusive\nproducts.\n\nModern systems increasingly combine them.\n\nMCP was introduced as an open protocol for connecting AI systems to\nexternal data and tools rather than requiring a custom integration for\nevery source. Google introduced A2A to support collaboration between\nagents built by different vendors and frameworks, and later transferred\nthe protocol into independent Linux Foundation governance.\n\nThese developments provide a natural experimental setting.\n\nIf protocols actually increase reversibility, migrations between\nimplementations using common interfaces should exhibit lower\nreconstruction costs than equivalent migrations between proprietary\nintegrations.\n\nIf they do not, protocol interoperability may improve connectivity\nwithout materially reducing lock-in.\n\nEither result is informative."
    },
    {
      "heading": "10. Research Questions",
      "text": "This framework produces several direct research questions.\n\nRQ1\n\nDo protocol-mediated AI architectures exhibit lower exit costs than\notherwise comparable agent-bound architectures?\n\nRQ2\n\nDoes separating memory, identity, authorization, and tool\nrelationships from the active model or agent increase effective state\nportability?\n\nRQ3\n\nDoes falling inference cost correlate with falling switching cost, or\ncan the two diverge?\n\nRQ4\n\nDoes increased agent autonomy increase irreversible residue?\n\nRQ5\n\nDoes interoperability reduce human attentional burden during\nmigration, or merely relocate integration work onto the user?\n\nRQ6\n\nCan systems support greater affordable variation without\nproportionally increasing evaluation burden?\n\nRQ7\n\nDoes coordination-layer concentration reduce practical reversibility\neven when technical interoperability exists?\n\nThe last question is particularly important. Farrell and Newman show\nhow asymmetric networks can generate power through central hubs and\nchokepoints. A technically distributed AI ecosystem could therefore\nremain difficult to exit if identity, standards, payment, discovery,\nmodel access, or protocol governance concentrate around a small number\nof nodes."
    },
    {
      "heading": "11. Testable Hypotheses",
      "text": "The research questions can be converted into preregisterable\nhypotheses.\n\nH1: Portability Hypothesis\n\nSystems that externalize critical state through interoperable\nrepresentations will exhibit higher effective state portability during\nprovider migration than systems in which state remains\napplication-bound.\n\nH2: Exit-Cost Hypothesis\n\nHolding task complexity constant, protocol-mediated architectures\nwill require less time and human intervention to migrate between\nproviders than proprietary point-to-point integrations.\n\nH3: Abundance–Exit\nIndependence Hypothesis\n\nDeclining inference cost will not necessarily predict declining exit\ncost.\n\nThis is a core falsifiable proposition of the paper.\n\nIf inference prices fall and switching costs reliably fall by a\nsimilar magnitude, reversibility may simply be a downstream consequence\nof technological abundance rather than a distinct economic variable.\n\nH4: Autonomous Residue\nHypothesis\n\nIncreasing autonomous action capability will increase irreversible\nresidue unless accompanied by explicit reversibility mechanisms such as\ntransactions, staged execution, provenance, undo semantics, or bounded\nauthorization.\n\nH5: Affordable Variation\nHypothesis\n\nSystems that combine cheap generation with low switching and\nevaluation costs will enable users to test more viable alternatives per\nunit of human attention than systems optimized only for generation\ncost.\n\nH6: Coordination-Layer\nHypothesis\n\nNominal interoperability will produce weaker reductions in exit cost\nwhen identity, authorization, state formats, discovery, or execution\nremain concentrated under a single controlling platform."
    },
    {
      "heading": "12. Study I: Controlled\nMigration Experiment",
      "text": "The first experiment should be deliberately mundane.\n\nParticipants perform the same multi-session knowledge-work task using\ntwo system architectures.\n\nPossible domains include:\n\nsoftware development; research synthesis; personal productivity; content production; small-business workflow automation.\n\nParticipants work long enough for meaningful state to accumulate.\n\nAt an unannounced but ethically disclosed migration point, the active\nAI provider or agent implementation is replaced.\n\nConditions\n\nCondition A: application-bound state and proprietary\nintegrations.\n\nCondition B: externally represented state and\nprotocol-mediated integrations.\n\nThe underlying model capability should be held as constant as\npractically possible.\n\nMeasurements\n\nResearchers record:\n\nmigration duration; monetary cost; number of user operations; state recovery percentage; lost memories; integrations requiring manual reconstruction; authorization rework; human attention time; post-migration task performance; number of irreversible losses; participant confidence that migration succeeded.\n\nA counterbalanced within-subject design could reduce variance caused\nby individual technical competence.\n\nThe experiment succeeds scientifically regardless of which\narchitecture wins.\n\nIf no meaningful difference appears, the hypothesis that\narchitectural externalization improves reversibility is weakened."
    },
    {
      "heading": "13. Study II:\nAffordable Variation Experiment",
      "text": "Participants receive a design or planning problem for which multiple\nlegitimate solutions exist.\n\nExamples might include:\n\ndesigning a software architecture; producing a communications strategy; selecting a workflow; constructing a user interface; developing an organizational process.\n\nDifferent systems generate comparable numbers of alternatives.\n\nThe experiment then measures the entire variation\ncycle , not merely generation:\n\ng e n e r a t i o n → e v a l u a t i o n → t r i a l → a b a n d o n m e n t / a d o p t i o n generation \\rightarrow evaluation \\rightarrow trial \\rightarrow abandonment/adoption\n\nResearchers measure:\n\ncompute cost per alternative; human evaluation time; number of viable alternatives identified; time required to prototype each; cost of abandoning a tested alternative; state or dependencies left behind; final outcome quality.\n\nIf generation becomes nearly free while evaluation dominates total\ncost, the experiment would provide direct evidence for scarcity\nmigration.\n\nIf evaluation costs decline proportionally with generation, that part\nof the theory would be weakened."
    },
    {
      "heading": "14. Study III: Longitudinal\nExit Study",
      "text": "Laboratory migrations cannot reproduce all forms of lock-in.\n\nA complementary longitudinal study should therefore follow real users\nor organizations over six to twenty-four months.\n\nResearchers would periodically measure:\n\nnumber of AI providers in active use; accumulated memories and state; connected external services; workflow dependencies; provider-specific artifacts; credential relationships; switching attempts; abandoned switching attempts; successful migrations; reasons for remaining with a provider.\n\nAn important dependent variable is latent exit cost :\nusers may never attempt to switch precisely because they expect\nswitching to be costly.\n\nSurveys should therefore distinguish:\n\nfrom:\n\nThose are economically very different forms of persistence."
    },
    {
      "heading": "15. Study IV: Failure-Recovery\nTest",
      "text": "Reversibility matters most when something goes wrong.\n\nResearchers could deliberately introduce controlled failures such\nas:\n\nagent misconfiguration; corrupted memory; revoked credentials; failed provider; incompatible model update; incorrect automation; tool compromise.\n\nThe study then measures whether the user can return to a known-good\nstate.\n\nPossible metrics include:\n\nMean Time to Reversal\n\nM T T R v = ∑ time required to restore acceptable state N MTTR_v =\n\\frac{\\sum \\text{time required to restore acceptable state}}\n{N}\n\nRecovery Completeness\n\nPercentage of pre-failure functionality and state restored.\n\nResidual Damage\n\nExternal consequences that remain after nominal recovery.\n\nThis extends conventional reliability thinking. A system is not\nmerely reliable when failure is rare. It is also resilient when failure\nremains cheap to undo ."
    },
    {
      "heading": "16. Coordination-Layer\nMeasurement",
      "text": "Technical portability alone cannot establish practical\nreversibility.\n\nThe institutional structure surrounding a protocol matters.\n\nA longitudinal dataset could therefore track:\n\nnumber of independent implementations; provider concentration; governance structure; specification licensing; credential-provider concentration; identity portability; dependency on proprietary registries; default-provider effects; export/import compatibility; switching fees; API restrictions.\n\nThis is where the empirical program intersects with the broader\nAmbient Power hypothesis.\n\nA distributed system can still contain highly centralized\nchokepoints. Farrell and Newman’s work on weaponized interdependence\nprovides a useful warning against equating network distribution with\ndistributed power.\n\nReversibility therefore needs to be measured at both:\n\nthe application layer and the coordination layer ."
    },
    {
      "heading": "17.\nReversibility Profiles Rather Than a Universal Score",
      "text": "There is a temptation to combine all variables into one number:\n\nR = f ( E C , S P R , A B , I R , A V R ) R = f(EC, SPR, AB, IR, AVR)\n\nThis paper recommends resisting that temptation initially.\n\nThe dimensions are not obviously commensurable.\n\nA medical AI system may rationally sacrifice some portability for\nstringent audit requirements. A temporary creative tool may require\nalmost no persistent state. A financial agent may need deliberate\nirreversibility for settlement finality.\n\nTherefore the initial empirical output should be a Reversibility Profile :\n\nOnly after multiple domains have been studied should researchers\ndetermine whether a general composite index is useful."
    },
    {
      "heading": "18.\nDistinguishing Reversibility from Related Concepts",
      "text": "Portability is not\nreversibility\n\nA dataset can be portable while the surrounding workflow is\nimpossible to reconstruct.\n\nInteroperability is not\nreversibility\n\nSystems can communicate while remaining expensive to replace.\n\nReliability is not\nreversibility\n\nA highly reliable system can still produce catastrophic lock-in.\n\nPrivacy is not reversibility\n\nStrong privacy protections do not necessarily make exit\ninexpensive.\n\nDecentralization is not\nreversibility\n\nA decentralized architecture can contain formats, governance\nstructures, or economic dependencies that make switching difficult.\n\nCheap generation is not\nreversibility\n\nThis is the central distinction."
    },
    {
      "heading": "19. Relation to\nUbiquitous and Ambient Computing",
      "text": "The idea that computation may retreat from focal attention has a\nsubstantial intellectual history.\n\nMark Weiser’s 1991 account of ubiquitous computing imagined\ncomputational elements becoming sufficiently embedded in everyday\nenvironments that their presence ceased to dominate conscious\nattention.\n\nThe present framework does not claim that disappearance itself\nconstitutes progress.\n\nAn invisible system can be extraordinarily difficult to escape.\n\nIndeed, reduced interface salience could make dependencies less\nlegible.\n\nThe contribution of reversibility as a measurement program is\ntherefore to ask a question that classic ubiquitous-computing visions do\nnot answer:\n\nAmbientness without reversibility could produce invisible\nlock-in.\n\nAmbientness with reversibility would be a substantially different\narchitecture."
    },
    {
      "heading": "20. What Would\nFalsify the Program’s Stronger Claims?",
      "text": "A useful theory must risk failure.\n\nSeveral findings would weaken the underlying Ambient Era\ninterpretation.\n\n20.1 Exit costs fall\nautomatically\n\nIf falling AI execution costs consistently produce proportional\nreductions in migration and switching costs, there may be little need to\ntreat reversibility as an independent economic primitive.\n\n20.2\nProprietary agent systems remain highly reversible\n\nIf tightly integrated persistent agents can achieve low migration\ncosts without externalized state or open protocols, the architectural\nhypothesis would be weakened.\n\n20.3\nProtocolization does not improve portability\n\nIf MCP-, A2A-, credential-, or other protocol-mediated architectures\nshow no measurable improvement in provider substitution or state\nrecovery, interoperability may primarily improve connectivity rather\nthan exit.\n\n20.4 Users do not value\nreversibility\n\nIf users consistently choose irreversible systems even when\nequivalent reversible alternatives exist and fully understand the\ndifference, reversibility may be less economically important than\nproposed.\n\n20.5 Variation remains easy\nto evaluate\n\nIf evaluation and selection costs fall as rapidly as generation\ncosts, the proposed migration of scarcity toward human attention would\nbe substantially weaker.\n\nThese are not edge cases to explain away.\n\nThey are valid outcomes."
    },
    {
      "heading": "21. What Would Constitute\nStrong Evidence?",
      "text": "Conversely, the framework would gain support if repeated studies\nfound that:\n\ninference costs continue to fall while exit costs remain stubbornly\nhigh; accumulated AI state predicts provider lock-in; protocol-mediated state significantly lowers migration cost; identity and authorization externalization improves provider\nsubstitution; autonomous systems increase irreversible residue without explicit\nreversal primitives; users can generate vast numbers of alternatives but are constrained\nprimarily by evaluation and integration; architectures designed for reversibility enable more experimentation\nat equal or lower attentional cost.\n\nThe strongest evidence would not be another conceptual analogy.\n\nIt would be a reproducible dataset showing that reversibility\npredicts meaningful technological freedom independently of raw AI\ncapability or price ."
    },
    {
      "heading": "22. Implications for AI\nEngineering",
      "text": "If reversibility proves measurable and consequential, it becomes an\nengineering property.\n\nSystems could then be designed around explicit reversal\nmechanisms:\n\nexportable memory; state schemas independent of model provider; portable credentials; provenance records; bounded delegation; reversible transactions; staged execution; undoable tool actions; provider-neutral workflow definitions; independent audit trails; model-substitution tests; standardized migration interfaces.\n\nThis would alter how agent systems are evaluated.\n\nA benchmark might ask not only:\n\nCan the agent complete the task?\n\nbut also:\n\nCan another agent inherit the task tomorrow without\nreconstructing the world?\n\nThat second question becomes increasingly important as AI systems\naccumulate persistent responsibility."
    },
    {
      "heading": "23. Implications for AI\nGovernance",
      "text": "Reversibility also offers a governance lens that does not require\nagreement about the eventual capabilities of artificial\nintelligence.\n\nGovernments, organizations, and standards bodies could measure:\n\npractical switching costs; portability obligations; continuity rights; credential independence; dependency concentration; availability of interoperable alternatives.\n\nThis reframes some debates around AI competition.\n\nThe relevant question is not merely how many AI providers exist.\n\nTen nominal providers do not constitute meaningful variation if\nleaving one requires abandoning accumulated identity, history,\npermissions, and workflows.\n\nCompetition requires usable exit ."
    },
    {
      "heading": "24. The Agent as an\nEmpirical Question",
      "text": "The wider Ambient Era framework proposes that the explicit agent may\nbe a transitional offload mechanism rather than the final architectural\nform of AI. The current evidence does not establish this. Existing\ntrends support only the weaker observation that some context, identity,\nauthorization, and coordination functions can move outside individual\nagent processes.\n\nThis paper therefore does not assume an agent-to-ambient\ntransition.\n\nInstead, it offers a way to detect one.\n\nIf over time:\n\nstate increasingly survives agent replacement; identity becomes independent of model instance; tools become accessible through shared protocols; switching between agents becomes inexpensive; users interact less with persistent branded agent identities; orchestration moves into shared infrastructure;\n\nthen agent salience may indeed decline.\n\nIf instead persistent agents accumulate identity, history,\nrelationships, proprietary state, and ecosystem lock-in, agents may\nbecome more durable rather than transitional.\n\nBoth futures are compatible with cheap intelligence.\n\nReversibility measurements can help distinguish them."
    },
    {
      "heading": "25. A Minimal Empirical\nProgram",
      "text": "The research agenda can begin without large institutional\nresources.\n\nA minimal first study would require only:\n\ntwo AI architectures; one multi-session task; controlled accumulation of state; a forced migration; measurement of migration cost and state recovery.\n\nThe smallest credible experiment could compare:\n\nA. an agent whose instructions, memories, files,\ntool state, and credentials are maintained inside one application;\n\nagainst\n\nB. an architecture where as much of that state as\npossible is maintained in provider-independent formats and external\nservices.\n\nAfter several sessions, swap the underlying agent or provider.\n\nMeasure what survives.\n\nThat experiment alone would produce more empirical information about\nreversible AI architectures than another layer of abstract\nterminology."
    },
    {
      "heading": "26. Discussion",
      "text": "AI abundance is often described in terms of what people will be able\nto create once cognition becomes inexpensive.\n\nThat framing is incomplete.\n\nA civilization of cheap creation could still be characterized by\nexpensive departure.\n\nUsers might generate software instantly while remaining locked into\nidentity systems. Organizations might deploy interchangeable models\nwhile being unable to migrate their operational history. Agents might\ncommunicate through open protocols while depending on concentrated\nauthorization infrastructure. Infinite apparent variation could coexist\nwith remarkably little practical freedom.\n\nThe relevant measure of abundance may therefore not be the number of\npossibilities a system can generate.\n\nIt may be the number of possibilities a person can actually\ntry without becoming trapped by each one .\n\nThis is the economic intuition behind affordable variation.\n\nReversibility gives that intuition an empirical form."
    },
    {
      "heading": "27. Conclusion",
      "text": "Falling inference costs are transforming the economics of\ncomputation. They do not tell us what happens to switching costs,\naccumulated state, identity dependencies, human attention, or\nirreversible commitments.\n\nThose variables need to be measured separately.\n\nThis paper proposes reversibility as a research\nobject for AI systems and decomposes it into:\n\nexit cost; state portability; attentional burden; irreversible residue; affordable variation.\n\nThe framework makes no claim that protocol-mediated, ambient, or\ninfrastructural AI will inevitably replace persistent agents.\n\nInstead, it supplies experiments capable of determining whether such\na transition is occurring.\n\nThe central question is deliberately simple:\n\nIf the answer is yes, technological abundance may broaden meaningful\nchoice.\n\nIf the answer is no, AI may create a peculiar economic condition in\nwhich alternatives become almost free to generate while existing\narrangements become increasingly expensive to escape.\n\nThat distinction may prove more important than abundance itself."
    },
    {
      "heading": "References",
      "text": "Anthropic. (2024). Introducing the Model Context Protocol .\nMCP was introduced as an open standard for connecting AI systems with\nexternal data sources and tools.\n\nFarrell, H., & Newman, A. L. (2019). Weaponized\nInterdependence: How Global Economic Networks Shape State Coercion .\nInternational Security, 44(1), 42–79.\n\nGoogle Cloud. (2025). Announcing the Agent2Agent Protocol\n(A2A) .\n\nGoogle Cloud. (2025). Google Cloud donates A2A to Linux\nFoundation .\n\nHirschman, A. O. (1970). Exit, Voice, and Loyalty: Responses to\nDecline in Firms, Organizations, and States . Harvard University\nPress.\n\nMaslej, N., et al. (2025). Artificial Intelligence Index Report\n2025 . Stanford Institute for Human-Centered Artificial\nIntelligence.\n\nSPIFFE. SPIFFE Workload API and SPIFFE Standard .\nSpecifications for portable and interoperable workload identity.\n\nW3C. (2025). Verifiable Credentials Data Model v2.0 . W3C\nRecommendation.\n\nWeiser, M. (1991). The Computer for the 21st Century .\nScientific American, 265(3)."
    }
  ]
}