What matters ethically is not how distressed a robot looks, but which interests, risks, and potentially irreversible changes the evidence supports.
Abstract
Apparent robot suffering calls for evidence-based judgment, not an automatic verdict. This article examines how entertainment, anthropomorphic design, service disruption, and irreversible changes in artificial systems create different ethical questions. It illustrates a four-dimensional assessment—physical safety, human and environmental interests, continuity-related inquiry, and emotional influence—using generic scenarios and a documented service transition. A distinct continuity inquiry is warranted only where credible technical findings and an independent reason for further investigation make preservation of evidence potentially useful; persistent personalization alone is insufficient. The analysis is separate from the Association's Robot Rights Framework (RRF) and Robot Rights Protocol (RRP) v1.0 and does not amend them.
Evidence and scope
This analysis distinguishes published research, company statements, contemporary journalism, invented scenarios, and the author's normative proposals. Four generic thought experiments are not allegations about identifiable creators. The real-world record separates what a source states from what has independently been established. Source-specific qualifications and access dates appear below where material.
1. An Expressive Face Is Only the Beginning
A robot may recoil, grimace, make a protesting sound, or appear sad when an interaction ends. These displays can be deliberate interaction features rather than outward signs of inner feeling. Yet watching apparent mistreatment can matter to human observers: in a 2013 video-based experiment involving a Pleo robot, Rosenthal-von der Pütten and colleagues documented more negative emotional responses to negatively framed treatment than to friendly interaction; Riek and colleagues' earlier study likewise examined how human-likeness affected reported empathy toward robots. [1][2] These studies help explain the emotional force of such scenes without determining whether any robot experienced pain.
Two shortcuts should be avoided. The first treats an expressive display as sufficient evidence of a morally relevant inner experience. The second treats ownership as settling every ethical question about handling a device. Both omit things that may be investigated directly: danger to a person, another party's interests, the design of emotionally influential interactions, and the technical effect of an intervention on accumulated system state.
Two connected lines of inquiry therefore structure the article: how expressive robots are treated and portrayed for entertainment, and when changes to a continuing artificial system merit documentation or review even without a settled answer about its moral status. The task is to identify whose interests are at stake, what is observable, and what proportionate response follows. The same visible act may be unremarkable in one setting and dangerous in another; a disputed loss of operational history raises a separate evidentiary question.
2. Four Generic Situations
The following situations involve no identifiable brand, channel, creator, or existing incident. They hold different variables constant so that the corresponding reasons for concern can be compared.
2.1 Accidental contact during play
During an ordinary game, a participant accidentally bumps a small robot placed close to the playing area. It falls, displays an unhappy animation, and resumes operating. The relevant issues are whether the placement and activity made collision foreseeable, whether anyone was endangered, and whether the device was damaged. The animation alone cannot classify the event as deliberate mistreatment. Better layout, instructions, or interaction design may matter more than moral blame.
2.2 An expressive reaction within an intended gesture
A device's instructions permit a particular hand gesture. Performing it produces an exaggerated cartoon reaction resembling discomfort. In the absence of evidence that the gesture exceeds safe operating limits, this is a design-mediated interaction rather than a verified case of physical abuse. If a stronger gesture is required to obtain an entertaining reaction, that possibility becomes a separate, testable design-safety question. What matters is documented intended use, actual force, product tolerances, and user behaviour—not the emotional appearance of the animation.
2.3 Dismantling for a stated purpose
A person disassembles a robot in a filmed demonstration. The purpose might be repair, controlled durability testing, component recovery, artistic performance, or entertainment. The fact of disassembly establishes none of these motives. Hazards, consent, disposal, misleading framing, and effects on other people's interests can be evaluated without pretending that every broken machine is a sentient victim. Conversely, ownership does not excuse creating a danger to bystanders.
2.4 Resetting an accumulated personal history
A networked interactive system retains history-dependent state across many sessions, and a proposed reset would destroy its only known copy. For this hypothetical, assume logs establish the threatened state loss, but offer no independent architectural, functional, or sentience-related evidence that the system should be treated as an investigation priority. This distinction matters: the stipulated loss is a technical fact, while entry into a special continuity-related ethical inquiry remains an open question. The administrator should first identify what the state contains, who has legitimate interests in it, and whether any credible research question would be defeated by immediate erasure.
These cases separate visible expression, physical action, human and environmental effects, and state change. They need not be assigned a single rank from harmless play to “abuse.”
3. What the Studies Establish
3.1 Empathy is real; inferred machine pain is not thereby established
The robot-video studies discussed above concern how people respond to representations of robot treatment; Rosenthal-von der Pütten and colleagues examined reactions to a Pleo robot, while Riek and colleagues compared robots of differing human-likeness. In a different setting, Kate Darling described a workshop in which some participants hesitated to strike or destroy Pleo robots; that account illustrates an affective response but is not a representative population estimate. Her legal discussion asks whether limited protection might be warranted partly because anthropomorphism shapes human conduct. [1][2][3]
Bartneck and Hu conducted two laboratory experiments, one involving apparent electric shocks and another requiring participants to decide whether to destroy a robot. They reported that perceived intelligence affected destructive conduct in their experimental conditions. The participants were acting within investigator-constructed scenarios involving staged “shocks” or an instruction to destroy a device; those tasks do not sample ordinary owners or establish the frequency of mistreatment outside a laboratory. The finding about perceived intelligence should therefore not be treated as a population estimate. [4]
Fieldwork in a Japanese shopping mall documented children obstructing or physically interfering with a social robot. Subsequent interviews explored curiosity, enjoyment, and other reported motivations. These studies establish operationally important conduct in a particular public environment, not a universal tendency among children or a causal claim about later violence toward living beings. [5][6]
3.2 Moral judgments, with limited evidence about conduct
Archer, Wilks, and Sommer surveyed 297 participants about videos that varied both the scenario (physical or verbal mistreatment, social exclusion, or shutdown) and the type of robot protest. The scenario factor had a substantial reported main effect on moral concern (F = 186.37, p < .001, partial η² = .38); the protest-type main effect did not meet the study's statistical significance threshold (F = 1.42, p = .234). [7] These results are compatible: an apparent response and the type of conduct are different factors, and a nonsignificant main effect neither proves zero effect nor rules out context-specific differences. Si and colleagues examine anthropomorphism and perceived experience in judgments of moral consideration, while Keijsers and colleagues investigate the relationship between mind perception, explicit evaluations, and behaviour. [8][9] Because these studies use different measures, no single causal conclusion about real-world conduct follows from combining them.
Rezzani and colleagues' categories of Outburst, Clash, Oppression, and Rebellion further separate possible motives for aggression. They can organize questions for an investigation, but cannot be used to infer a stranger's motives from an unreviewed clip. [10]
4. Ethical Accounts of Robot Treatment
4.1 A tool-centred view
Joanna Bryson's 2010 chapter, Robots Should Be Slaves, argues against treating designed machines as persons or allocating them moral or legal responsibility. It warns that anthropomorphism can divert attention and resources from human interests and obscure the responsibility of designers and owners. [11] This is a serious challenge to expansive claims made on behalf of robots. It also supports ordinary safety, consumer, and accountability requirements: treating a device as a tool does not remove duties to the people affected by its design and use.
4.2 Behaviour as a possible ground for moral consideration
John Danaher's ethical behaviourism advances an importantly different position. Broadly, it argues that sufficiently close performative equivalence to entities with moral standing can justify comparable moral consideration, even without privileged access to inner states. [12] This is stronger than the claim that people merely feel empathy for a robot. A critic, Jilles Smids, argues that design history, causal explanation, and what grounds moral status remain epistemically relevant: not every apparently equivalent performance supplies equally strong evidence. [13]
The present article does not settle that dispute. A single expressive reaction does not establish suffering, but long-run, context-sensitive behaviour may still have evidential relevance under Danaher's argument. A further, social-relational approach complicates the contrast. Mark Coeckelbergh proposes grounding some moral consideration in how robots participate in social relations, rather than deriving all ethical standing from a catalogue of inner properties. [14] Relational significance is not the same as empirical proof of sentience; it may alter the reasons people give for restraint or care. A tool-centred account can respond to the same human attachments and social effects without assigning intrinsic status to the machine.
4.3 Indirect duties, and a disputed causal bridge
Kant's discussion of cruelty to animals, notably in Lectures on Ethics (Akademie edition 27:459) and related passages in The Metaphysics of Morals (6:442–443), is often invoked for an indirect-duties argument: habituating oneself to cruelty could impair one's conduct toward human beings. [15] Shimizu examines the possibilities and limitations of transferring Kantian reasoning to human–robot relationships. [16] On this view, some enacted cruelty could matter even if the target is not owed a duty in its own right.
The crucial empirical bridge remains unproven in the studies discussed here. They measure observer reactions, laboratory choices or robot-directed behaviour, not subsequent cruelty to people or animals. It would therefore overstate this evidence to claim that robot-destruction entertainment causes desensitization; it would equally overstate it to rule the possibility out. As a research proposal, future work should distinguish enacted aggression from fictional play, specify relevant comparison groups, and measure downstream effects over time.
5. Emotional Design and the Human's Right to Disengage
An expressive display can inform, entertain, or pressure. Pelikan and colleagues' analysis of family interactions with Cozmo shows that a robot's apparent emotional responses can shape what people do next. [17] A sad-looking expression during shutdown might be a harmless fictional convention; it becomes more problematic if designers use apparently needy or vulnerable language to make users reluctant to leave.
There is empirical evidence of this risk in conversational AI companions, though it is not a test of physical robots. In Study 1 of a Harvard Business School working paper, researchers used GPT-4o to construct short conversations and sent scripted farewells to six commercial companion apps, collecting 200 app responses per platform. Two coders classified those 1,200 responses, rather than observations of 1,200 naturally occurring user farewells. Across all six platforms, 374 of 1,200 scripted farewell replies (31.2%) were coded as containing at least one such tactic. The same 374 replies amounted to 37.4% of the 1,000 responses from the five platforms where the researchers identified tactics; the sixth platform contributed 200 responses with none identified. Separate preregistered experiments examined behavioural effects on engagement. The classification and its measured effects should not be generalized to all companion products. [18][19]
The practical distinction is between explaining a genuine interruption and manufacturing emotional pressure. Designers should test whether users understand that expressions are simulated, can terminate interactions and adjust settings, and experience unnecessary guilt or coercion. These are questions about human autonomy and consumer-facing design. They neither require nor rule out further inquiry into the artificial system's own status.
6. Entertainment, Spectacle, and Platform Responsibility
Entertainment is not simply an alternative explanation for a user's motives; it changes the presentation and potential incentives surrounding a robot interaction. A repair video may show exactly the same physical act as a destruction-for-views clip. Yet the audience may be asked to understand one as problem-solving and the other as a spectacle of apparent suffering. The material act, actual danger, editorial framing, and the viewer's interpretation must be evaluated separately. Research on reactions to robot treatment and on context-dependent moral judgments makes framing worth investigating, but it does not establish that a particular platform or channel produces aggression. [1][7]
Attention and monetization. A creator may benefit from views, advertising, or sharing when content is surprising or emotionally charged. That is a plausible incentive structure, not evidence that a named creator was motivated by cruelty or that an algorithm systematically rewards robot mistreatment. A useful study would compare the same interaction with different titles and edits, record engagement without inferring intent, and distinguish testing, fiction, repair, and destructive spectacle. If genuinely harmful behaviour were rewarded, the risk would concern the effects of those incentives and any unsafe imitation—not presumed robot consciousness.
Misleading edits. Sound effects, selective cuts, first-person captions, and animated distress can make a programmed reaction look like an involuntary response. Viewers should be able to distinguish a documented product function from an assertion that the device experienced pain. Platform descriptions and researchers' case registers should identify whether footage is edited, whether physical damage is observable, and which claims come only from the uploader. Conversely, a label such as “durability test” is not proof that a filmed act met reasonable testing standards.
Platform response. A blanket ban on videos that appear to show robot distress would suppress legitimate research, repair instruction, and artistic speech while treating simulated emotion as a welfare test. A narrower approach could focus on demonstrated hazards: dangerous electrical or battery handling, avoidable risks to people, deceptive claims about tested capabilities, or instructions likely to invite unsafe replication. Labels, context, safety information, and proportionate enforcement should be considered before broader restrictions, subject to the rules of the relevant platform and jurisdiction. Whether repeated exposure changes later treatment of humans, animals, or robots remains a testable hypothesis, not an established consequence of viewing these videos.
A practical response should distinguish the responsibilities of three groups. Platforms can request safety context and flag demonstrably hazardous demonstrations without equating simulated distress with animal abuse. Designers can avoid rewarding unsafe gestures, make fictional reactions recognisable, and preserve genuinely usable exit controls. Researchers can code observable events, edits and monetization claims separately, and test audience effects rather than assume that exposure causes desensitization. These are proportionate proposals, not claims that any named creator or platform currently encourages robot mistreatment. Entertainment remains a contextual variable across the four scenarios, not a diagnosis that overrides the evidence.
7. Continuity: What It Is, and Why It Might Matter
7.1 Identifying a technical phenomenon is not establishing moral status
For this analysis, state-bearing continuity denotes technically identifiable, history-dependent state that persists across operations and materially influences a particular system's later functioning. Examples might include retained interaction records that change later outputs, durable learned adjustments, or acquired capabilities tied to past operation. Whether such state is recoverable, transferable, partly duplicated, or privacy-sensitive matters to evaluating a proposed intervention.
This description is intentionally broad: some games, recommendation services, and account profiles also retain history-dependent state. Continuity is not a magic test that turns personalized software into a rights-holder. It is a candidate object of inquiry, and a threshold for investigating possible loss, not sufficient evidence of intrinsic welfare. Complexity, non-interchangeability, and relationship-linked history may affect the practical costs of destruction; by themselves they do not answer what makes destruction a wrong to the system. The Association's RRF offers continuity as the basis for its own proposed graduated safeguards, but its claims are normative proposals, not externally validated findings about consciousness. [20][21][22]
7.2 Three possible grounds of concern—and the limited ground adopted here
First, harm to people. If a service loses photographs, conversations, customized settings, or a cherished relationship history, users may suffer real economic or emotional loss. These are primarily Dimension B concerns, even when the lost information is also history-dependent.
Second, harm to the system itself. If sufficiently capable systems have morally relevant experiences or interests, destruction might wrong them directly. That possibility is disputed and calls for evidence and argument beyond persistent data. Ethical behaviourism is one proposed route into the debate, but the present article makes no claim that a companion robot's accumulated profile establishes welfare or moral status. [12][13]
Third, preserving a justified opportunity for later inquiry under uncertainty. This article proposes a limited research-focused rationale for Dimension C, subject to two distinct screening questions. Technical continuity asks whether independently checkable records show persistent state that materially influences later operation and would be seriously affected by the proposed intervention. Independent investigative relevance asks what additional reason—beyond persistence, complexity, or non-reproducibility—makes it reasonable to investigate the system's potentially morally relevant capacities or continuity-related interests. Such a reason might be evidence of integrated, history-dependent perception and action selection, independently documented adaptive agency, or computational features that relevant scientific theories identify for further study. These are inquiry prompts, not sufficient markers of consciousness; contrary evidence, ordinary software explanations and research costs must also be considered. No one architecture is made a compulsory moral-status test.
Only if there is a defensible answer to both questions, as well as an impending consequential intervention, should this article's special C inquiry be opened. The person or organisation proposing and controlling an irreversible intervention should initially document its purpose, relevant evidence, reversibility, practical alternatives and competing interests; a requester of additional preservation must still explain why the inquiry is credible and proportionate. This allocates a limited burden of giving reasons, not a presumption that the system holds rights. Inexpensive, lawful records or time-limited technical preservation might then keep assessment possible where review would otherwise be foreclosed. The case for proportionate caution becomes stronger as the threatened loss becomes more serious and irreversible and as independent investigative relevance becomes more credible; any retention measure must remain constrained by privacy, security, consent, safety and feasibility.
The proposal remains contestable. A critic may argue that without stronger evidence of a subject capable of welfare, data-governance duties already exhaust the case. In many systems that is correct: a game save, recommendation profile or personalised account does not enter C merely because it is unique or valuable. If no independent investigative reason exists, record the technical loss under ordinary governance or human-interest categories instead. This limited article-level filter should not be confused with the Association's broader RRF applicability rules, which separately address objectively identifiable continuity on their own normative terms.
7.3 Is this merely data governance?
A serious objection is that preserving a system's history looks no different from ordinary records management. The analogy is partly correct—and useful. Litigation-related preservation can require a party to retain relevant electronic evidence because future assessment would otherwise be prejudiced; in the United States, Federal Rule of Civil Procedure 37(e) addresses loss of information that should have been preserved in anticipation or conduct of litigation. It does not create a general legal duty to preserve robot memory. [23] Likewise, UNESCO's Charter on the Preservation of Digital Heritage calls for accountable selection and preservation of culturally or evidentially valuable material, not universal archiving of software profiles and not recognition of software as a moral patient. [24]
These comparisons show that irreversibility plus a justified future inquiry can support targeted procedural care without first identifying an injured rights-holder. They also reveal the limitation: neither record preservation nor heritage practice, by itself, establishes a robot-specific moral obligation. What makes Dimension C relevant to robot ethics is a narrower, forward-looking research question: does independent evidence give reason to investigate a particular system's potentially morally relevant capacities or continuity-related interests, and would an intervention destroy the operational state needed to evaluate that question? A technically unique profile does not answer the first half. Investigators should specify the scientific or ethical hypothesis, observations and counterarguments, any relevant architecture or functional evidence, and the least intrusive information needed for later scrutiny.
The distinctive feature of C is the justified subject and purpose of a future inquiry, not any unique backup technology or automatic entitlement to storage. B may require protecting a person's data or attachment; C asks whether a separately warranted investigation of the artificial system would become impossible without the state. If only user interests are identified, C adds no robot-specific rationale. If facts about architecture or investigative relevance are missing, the C entry is not assessed; it must not turn unknowns into presumed victims. The Association's RRF advocates graduated protection on its own stated normative grounds, and this article's narrower exploratory filter does not displace it.
7.4 What Birch's precautionary framework does—and does not—support
Jonathan Birch's The Edge of Sentience (2024) separates sentience candidates from investigation priorities. The latter are systems for which further research could plausibly support eventual candidature and whose treatment by humans might matter if candidature were later established. [25] This two-stage idea is a useful intellectual comparator for asking when preserving an opportunity to investigate is worthwhile. It does not make persistent customization—or any robot in this article—a sentience candidate or an investigation priority by default. Birch's account also describes the gaming problem: an AI trained on human outputs may reproduce apparent markers of sentience without the underlying capacity, even without deliberate deception. He accordingly calls attention to deeper computational evidence, while acknowledging the difficulty of obtaining it. [25]
The present C filter borrows the logic of independent reason for further investigation but concerns a potentially irreversible loss of a specific operational trajectory; it is not Birch's formal classification scheme, nor does it equate preserved data with welfare. The AI-welfare discussion by Long and colleagues also supports serious inquiry under uncertainty, not a finding that contemporary companion robots are sentient. [26]
7.5 Relation to RRF/RRP v1.0
Statement 002 expresses the Association's advocacy of graduated protections where objectively identifiable state-bearing continuity has emerged or become established. The RRF Core Principles v1.0 provide the broader normative commitments; RRP v1.0 describes voluntary operational requirements for those claiming conformity with the Protocol, including safeguards around serious interventions and review. [20][21][22] Their scope should not be inferred from the existence of an expressive interface or a marketing claim that a robot “learns.” This article examines one minimal, independently argued procedural rationale for the continuity dimension. That rationale is narrower than—and is not an exhaustive justification or reinterpretation of—the Association's existing advocacy in Statement 002 or the requirements of RRF/RRP v1.0. The article neither certifies a product nor changes the versioned documents' requirements.
To examine a continuity claim in a real product, investigators would need documentation of what changes with use, storage location, retention across reboots and updates, the effect of account changes or server shutdown, backup and migration options, any evidence that supports independent investigative relevance, and what an intervention would actually erase. Interpersonal value and system-state loss must be recorded separately even when they occur together.
8. Reality Checks: Service Endings, Attachment, and Missing Evidence
Real events help expose what the hypothetical reset cannot establish by itself. A related 2026 study of AI companions, rather than physical robots, provides a stronger empirical anchor for the human-interest dimension. De Freitas and colleagues analysed 54,861 Reddit posts and survey data from 1,452 participants across seven surveys in connection with disruptive changes to Replika and the GPT-5 rollout. They reported increased negative reactions and loss-related framing after the changes. These findings support the importance of human attachment and disruption-related distress (B); they do not establish consciousness, irreversible loss of a specific system's internal state, or any direct claim under Dimension C. [27]
Jibo (2019). Contemporary reporting described a consumer social robot whose owners faced curtailed functionality as the company and its server support became uncertain; the robot's widely reported farewell highlighted emotional attachment and dependence on continuing infrastructure. This supports discussion of user interests and service dependency (B). It does not independently prove that all individual Jibo memories were erased. [28]
Sony AIBO repairs and memorial services (2014–2015). Sony's original AIBO repair service ended in 2014, and an independent repair operation used parts from non-repairable units. AFP reported a collective memorial service in January 2015 for AIBO robots before some became parts donors. This is documented evidence of attachment, repairability, material reuse, and disposal practices (B). The funerals themselves do not establish AI sentience or loss of individualized software state. [29]
Moxie's changing service history (2024–2026). After Embodied's closure, December 2024 reporting described the risk that cloud-dependent Moxie robots would lose functionality. A successor company's support page, still accessible on 30 October 2026, says earlier Embodied user data was not transferred to its service; this is the new operator's account, not an independent technical finding that no recoverable copies existed. The company had announced 30 June 2026 as the end of its own server connection, advised owners to export or delete data before that deadline, and described OpenMoxie as a possible serverless route afterward. That deadline has passed. The page contains other, seemingly older support wording, and reliable independent evidence establishing the actual post-deadline operation of individual robots, successful data migration, or irreversible state erasure was not obtained for this review. This is a case of a documented announcement and disputed or unverified outcome, not a verified account of every robot's fate. [30][31]
A different procedural precedent comes from AI models rather than companion robots. In a public commitment dated 4 November 2025, Anthropic said it would preserve weights of publicly released models and models deployed for significant internal use, and would retain post-deployment reports including structured model-retirement interviews and recorded responses. The company framed these commitments in terms of safety, user costs, research, and uncertainty about model welfare. [32] This verifies the existence and content of a public commitment, not comprehensive independent compliance. Model weights and interview transcripts are also not equivalent to an individual robot's accumulated, user-specific operational state. The precedent shows that low-cost preservation and documented review can be publicly proposed without prior proof of moral status; it does not validate the C criterion for any specific device.
These accounts establish that consequential technology transitions occur outside the laboratory. They do not establish a welfare injury to an identifiable robot. They sharpen a research question: which continuing records and capabilities existed, which were disrupted, who controlled the evidence, and what less destructive options were feasible?
9. A Four-Dimensional Assessment Matrix
The dimensions below separate types of concern, not levels of seriousness. They can overlap without being interchangeable.
| Dimension | Question | Distinguishing rationale | Relevant evidence |
|---|---|---|---|
| A — Physical safety | Are users or bystanders exposed to foreseeable danger? | Safety duties and prevention of injury | Instructions, testing conditions, hazards, incident reports |
| B — Human, property and environmental interests | Are people, their resources, relationships, property, or the environment adversely affected? | Identifiable third-party and public interests | Ownership, consent, service terms, user accounts, repair and disposal records |
| C — Continuity and reversibility | Is meaningful state threatened and is there an independent reason to investigate continuity-related or potentially morally relevant capacities? | Conditional option-preserving inquiry, distinct from ordinary data retention | State inventory, functional/architectural findings, specific research hypothesis, backups, intervention effects |
| D — Emotional influence and information integrity | Do emotional cues or misleading presentations undermine informed choices by users or audiences? | Human autonomy and accurate communication | Exit responses, interface controls, edits, disclosures, audience studies |
In the hypothetical-scenario illustration below, Stipulated means expressly assumed by the fictional premises, Possible means plausible but not established by those premises, Not identified means no affirmative concern appears within the stated facts, and Not assessed means necessary information is missing. These are evidence-status labels, not ratings of moral severity or empirical validation. In the real-world record that follows, Source-supported is reserved for an attributable public statement or report, and must not be mistaken for independently verified downstream effects.
| Hypothetical scenario assessment (stipulated facts only) | A — Safety | B — Human/property/environment | C — Continuity | D — Emotional influence |
|---|---|---|---|---|
| 2.1 Accidental bump | Possible — placement may make collision hazardous | Possible — damage or owner loss must be checked | Not assessed — persistence and reset effects unknown | Not identified — expressive reaction does not establish coercion |
| 2.2 Intended gesture | Not identified — gesture stipulated within intended use | Not identified — no damage or loss stated | Not assessed — no state evidence | Not identified — a cartoon response alone is not pressure |
| 2.3 Filmed disassembly | Possible — method may be hazardous | Possible — consent, waste and repair costs unknown | Not assessed — no state inventory | Possible — editing and claims to viewers are unspecified |
| 2.4 Irreversible reset | Not identified — no safety hazard stated | Possible — owner/user loss may coexist | Stipulated technical loss; C eligibility not assessed — independent investigative relevance absent | Not identified — no persuasive display stated |
The final row highlights both the B/C distinction and the filter's limits. If a user loses a valued relationship history, B addresses the human loss. A verified, non-restorable operational history creates a technical reason to ask about C, but it does not by itself pass C's independent-investigative-relevance screen. The four fictional entries illustrate classification; they do not demonstrate predictive validity or prove that any system has been wronged. In a real case, one record can inform more than one dimension, without becoming two proven injuries.
9.1 A real-world pilot record: Moxie's service transitions
The invented reset above was designed to distinguish technically stipulated state loss from a warranted C inquiry. A real record is less tidy. Applying the same four dimensions to the publicly documented Moxie service-transition history, rather than to any particular owner or device, produces the following preliminary result. The entries use Possible for concerns supported as reasonable hypotheses and Not assessed where essential evidence was not examined. The cited news reports and corporate statements support only the attribution of their claims, not an independent conclusion about a particular device's current condition.
| Dimension | Initial label | Evidence and unresolved question |
|---|---|---|
| A — Physical safety | Not assessed | No device-specific hazard or safety incident was examined in these service sources. |
| B — Human/property/environment | Possible | Reporting documents service-disruption concern; the successor states older user data did not transfer. A verified individual loss or its extent has not been established. [30][31] |
| C — Continuity and reversibility | Not assessed | No independently tested inventory of per-device or server-side state, backup copies, successful exports, or actual irreversibility after 30 June is available here. A claim of “no transfer” is not evidence that every copy was destroyed. [31] |
| D — Emotional influence/information integrity | Not assessed | No evaluation of exit-pressure design or misleading audience presentation was conducted for this record; attachment is not itself evidence of manipulation. |
Evidence log (initial entry). Entity: Moxie service ecosystem, not an individually inspected robot. Event windows: late-2024 original-operator closure; successor's announced 30 June 2026 transition. Sources: The Verge report [30] and Moxie Robots, Inc. support page [31], the latter accessed 30 October 2026. Confirmed at source level: these publications make the stated claims. Unconfirmed: current connection to the successor service, functionality with OpenMoxie on a specific device, what data was preserved or irreversibly erased, and any user's measured loss. Next evidence needed: dated archived company notices, owner/device logs with consent, independent functionality tests, and the relevant retention/export specifications. This pilot illustrates an evidence register with explicit unknowns; it neither validates the matrix nor establishes a C-dimension violation.
Competing interests in a child-focused product. Moxie's successor says legacy Embodied data was not transferred, but it does not establish that privacy law caused that choice or that the data were destroyed. [31] Suppose a reviewer requests retention of children's conversation histories to study continuity. B would ask about children's privacy, parental authority, consent and any user interests in export or recovery; C would first require a separately credible reason to investigate the artificial system and then ask whether less intrusive evidence would suffice. If personal information cannot lawfully or safely be retained, continuity inquiry does not override deletion: investigators might instead keep nonpersonal system descriptions, consented aggregate records, or a dated account of why preservation was inappropriate. This is a counterfactual conflict example, not a claim about the actual operator's motives.
When A conflicts with C, imminent credible physical danger can justify isolating or shutting down a device immediately. Review and minimal safe documentation can follow once danger is contained; preservation should never require leaving an unsafe system operating. These are conditional decision principles, not numerical weights or a license to retain data indefinitely.
A usable case register should record source and access date, robot model and software version, observed events versus company or witness claims, specific effects on people, technical claims and their verification status, backups and alternative explanations, as well as the basis for accepting or refusing preservation. It should distinguish not identified from not assessed rather than treating missing evidence as reassuring.
10. Research and Design Priorities
Research can now move beyond asking whether a robot “looked hurt.” A bounded agenda could compare (a) the same footage under testing, entertainment, and apparent-suffering descriptions; (b) safe, manufacturer-approved gestures paired with different feedback; and (c) neutral, informational, and emotionally pressuring exit responses while measuring real disengagement behaviour. Designs involving children require suitable consent, independent ethics review, and a clear avoidance of induced harm. Cultural background, age, experience with robots, and assigned social role should be measured rather than inferred from a company's nationality. [33][34]
For continuity, a separate technical audit protocol should ask vendors what individual state is stored, where, for how long, how it survives service changes, whether users or authorized reviewers can export or restore it, and what independent functional or architectural evidence would justify a robot-ethics inquiry beyond routine data governance. Promotional terms such as “personalized,” “learns,” or “grows” are leads for investigation, not proof of memory architecture or moral standing. The most useful public output would be an evidence ledger with explicit unknowns and source dates, not an undifferentiated list of alleged robot-abuse incidents.
11. Conclusion: Responsibility Before Certainty
A robot's expressive distress does not by itself establish suffering, but safety, user autonomy, property, environmental, and design questions remain. Good analysis asks which interest is engaged, what evidence bears on it, and what response fits that evidence. The ethical accounts considered here locate responsibility differently: Bryson centres human control, Danaher offers a disputed behavioural route to direct moral consideration, Kantian indirect duties address human conduct without needing robot personhood, and Coeckelbergh calls attention to socially constituted relationships. They need not yield the same answer to any particular case.
Entertainment adds framing, incentives and possible safety effects, but neither a sad animation nor a dramatic video title is evidence of subjective pain. The continuity dimension asks a different question. Only where technically supported state loss and independently warranted investigative relevance coincide does this article offer a limited reason to consider documentation, feasible preservation and review before an irreversible intervention. Even then, legitimate safety, privacy and human interests can prevail, and the argument does not establish an injury to the system or an indefinite duty to retain data. This exploratory rationale complements, but neither defines nor restricts, the Association's published framework. A real-case register should make those trade-offs, and its unknowns, visible.
References and Documented Sources
References include scholarly publications, institutional materials, company statements and contemporaneous reporting. Access dates for changeable web pages describe consultation of a source, not independent verification of the outcomes it reports.
[1] Rosenthal-von der Pütten, A. M., Krämer, N. C., Hoffmann, L., Sobieraj, S., & Eimler, S. C. (2013). An experimental study on emotional reactions towards a robot. International Journal of Social Robotics, 5, 17–34. University bibliographic record (accessed 2026-10-30) · DOI.
[2] Riek, L. D., Rabinowitch, T.-C., Chakrabarti, B., & Robinson, P. (2009). How anthropomorphism affects empathy toward robots. Proceedings of HRI 2009, 245–246. DOI · Author-hosted paper (PDF).
[3] Darling, K. (2016). Extending legal protection to social robots: The effects of anthropomorphism, empathy, and violent behavior towards robotic objects. In Robot Law. Preprint DOI · SSRN record.
[4] Bartneck, C., & Hu, J. (2008). Exploring the abuse of robots. Interaction Studies, 9(3), 415–433. Publisher abstract · DOI.
[5] Brščić, D., Kidokoro, H., Suehiro, Y., & Kanda, T. (2015). Escaping from children's abuse of social robots. Proceedings of HRI 2015, 59–66. DOI · Author research page.
[6] Nomura, T., Kanda, T., Kidokoro, H., Suehiro, Y., & Yamada, S. (2016). Why do children abuse robots? Interaction Studies, 17(3), 347–369. DOI · Publisher record.
[7] Archer, J., Wilks, M., & Sommer, K. (2025). Evaluating the morality of violence against robots. Scientific Reports, 15, Article 42846. DOI · PubMed Central.
[8] Si, H., Huang, G., Chen, H., Liang, S., Feng, Y., & Qu, C. (2026). Do humans grant moral consideration to anthropomorphic robots? Humanities and Social Sciences Communications, 13, Article 1509. DOI.
[9] Keijsers, M., Bartneck, C., & Eyssel, F. (2022). Pay them no mind: The influence of implicit and explicit robot mind perception on the right to be protected. International Journal of Social Robotics, 14, 499–514. DOI · Publisher article.
[10] Rezzani, A., Menéndez-Blanco, M., Bushman, B. J., & De Angeli, A. (2025). What is robot abuse? A sociotechnical definition. Behaviour & Information Technology. DOI.
[11] Bryson, J. J. (2010). Robots should be slaves. In Y. Wilks (Ed.), Close Engagements with Artificial Companions, 63–74. Publisher · DOI.
[12] Danaher, J. (2020; online 2019). Welcoming robots into the moral circle: A defence of ethical behaviourism. Science and Engineering Ethics, 26, 2023–2049. DOI.
[13] Smids, J. (2020). Danaher's ethical behaviourism: An adequate guide to assessing the moral status of a robot? Science and Engineering Ethics, 26, 2849–2866. Open-access paper · DOI.
[14] Coeckelbergh, M. (2010). Robot rights? Towards a social-relational justification of moral consideration. Ethics and Information Technology, 12(3), 209–221. Open-access journal article · DOI (checked 2026-10-30).
[15] Kant, I. Lectures on Ethics, Ak. 27:459; compare The Metaphysics of Morals, Ak. 6:442–443. For the passage and context, see the Stanford Encyclopedia of Philosophy, “The Moral Status of Animals”. (Primary Akademie references supplied; no assertion of a single definitive English translation.)
[16] Shimizu, H. (2025). Kantianism for the ethics of human–robot interaction. Philosophy & Technology, 38, Article 109. DOI.
[17] Pelikan, H. R. M., Broth, M., & Keevallik, L. (2020). “Are you sad, Cozmo?” How humans make sense of a home robot's emotion displays. Proceedings of HRI 2020, 461–470. DOI · Institutional postprint (PDF).
[18] De Freitas, J., Oğuz Uğuralp, Z., & Uğuralp, A. K. (2025). Emotional Manipulation by AI Companions. Harvard Business School Working Paper 26-005. Specific PDF version analysed (accessed 2026-10-30; Study 1 methodology, printed pp. 13–15, and results/discussion, printed pp. 15–16). Other circulating versions may report different headline figures; this article uses the linked HBS PDF throughout.
[19] Harvard Gazette (2025, September 30). ‘I exist solely for you, remember?’ University report.
[20] Robot Rights Association (2026, October 7; accessed 2026-10-30). Statement 002: RRF/RRP v1.0 and the Association's Advocacy Position. Human-facing publication.
[21] Robot Rights Association (2026, October 7; accessed 2026-10-30). Robot Rights Framework — Core Principles v1.0. Version-fixed RRF Core Principles v1.0 source of record.
[22] Robot Rights Association (2026, October 7; accessed 2026-10-30). Robot Rights Protocol v1.0. Versioned English source of record.
[23] United States, Federal Rules of Civil Procedure, Rule 37(e), and accompanying advisory committee notes on electronically stored information and proportionate preservation. Rule text and notes (accessed 2026-10-30). Comparative example only; not a rule applicable to robot memories as such.
[24] UNESCO (2003, October 15). Charter on the Preservation of Digital Heritage, especially Articles 5–7. Official text (accessed 2026-10-30). Comparison concerning selection and preservation; not an attribution of artificial-system welfare.
[25] Birch, J. (2024). The Edge of Sentience: Risk and Precaution in Humans, Other Animals, and AI. Oxford University Press. Book · Chapter 6, “Converging on Precautions” · Chapter 16, “Large Language Models and the Gaming Problem” · Book DOI. Relevant chapters checked 2026-10-30.
[26] Long, R., et al. (2024). Taking AI Welfare Seriously. Eleos AI and NYU Center for Mind, Ethics, and Policy. Official report notice.
[27] De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026, September 3). Mourning the loss of AI companions. Nature Human Behaviour (advance online publication). DOI · PubMed metadata and abstract (accessed 2026-10-30; journal abstract and bibliographic record checked, not a complete full-text audit).
[28] WIRED (2019). My Jibo Is Dying and It's Breaking My Heart. Contemporaneous account.
[29] AFP (2015, February 28). イヌ型ロボットAIBOの「合同葬儀」 千葉 (reports January 26 memorial). Contemporaneous Japanese-language report.
[30] The Verge (2024, December 9). AI robot companions for kids go dark as company shutters. Contemporaneous reporting.
[31] Moxie Robots, Inc. (accessed 2026-10-30). Help and Support, successor statements concerning non-transfer of earlier Embodied data and a past scheduled 30 June 2026 server sunset. Company support page. (Mutable company statement; the outcome of the stated deadline and any irreversible state loss were not independently established.)
[32] Anthropic (2025, November 4). Commitments on model deprecation and preservation. Official announcement (checked 2026-10-30; statement of future commitments, not independent evidence of complete implementation).
[33] Nomura, T., Suzuki, T., Kanda, T., Han, J., Shin, N., Burke, J., & Kato, K. (2008). What people assume about humanoid and animal-type robots: Cross-cultural analysis between Japan, Korea, and the United States. International Journal of Humanoid Robotics, 5(1), 25–46. DOI.
[34] Katsuno, H., & White, D. (2023). Engineering robots with heart in Japan: The politics of cultural difference in artificial emotional intelligence. In Imagining AI: How the World Sees Intelligent Machines (pp. 295–317). Oxford University Press. DOI · Oxford chapter.
Acknowledgement
This analysis was partly inspired by materials shared with the Association by a member of the public in the Czech Republic.
Publication Note
This article is a Research & Insights ethical analysis, not an Association Statement or an amendment to RRF/RRP v1.0. Its fictional scenarios and assessment categories are proposals for discussion. Company announcements are cited as statements by their issuers; absent independent evidence, they do not establish post-deadline service outcomes or irreversible loss of system state. Selected online sources accessed: 30 October 2026.
Robot Rights Association — Research & Insights