{"id":"c3e4faa2-d677-4f00-a22e-9d632634610f","arxiv_id":"2505.02519","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper argues that current AI sign language technologies, trained on interpreter-mediated data and framed as substitutes for interpreters, threaten deaf people's linguistic rights and require deaf-led design and governance.","lead":"This paper argues that AI language technologies, especially sign language tools, risk eroding the hard-won linguistic and accessibility rights of deaf people by being designed as substitutes for human interpreters. It draws on Deaf Studies and crip technoscience to show how interpreter-centered access models shape AI development in ways that sideline deaf communities and calls for deaf-led approaches.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Erosion claim rests on uncited anecdotes; the paper's urgency depends on current substitution rather than hypothetical risk.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern I find: the claim that erosion of linguistic rights is already underway rests on under-specified anecdotal evidence and is then extrapolated into a broad policy prediction. My stress-test agrees with that assessment and adds precision about why it matters: the anecdote is not just a weak piece of evidence for an auxiliary claim, it is the empirical bridge between the paper's well-supported framing critique and its urgent conclusion that rights are already being eroded. If institutions only supplement, not substitute, the paper becomes a plausible warning about future risk rather than a documented claim of ongoing harm. The central framing critique—that AI sign language technologies are benchmarked against interpreters and trained on interpreted datasets—is independently supported by cited literature (e.g., Desai et al., 2024; Vandeghinste et al., 2024; WFD & WASLI, 2018), so the paper should not be rejected. The reader's CONDITIONAL verdict already captures this: accept the conceptual contribution while requiring stronger empirical grounding for the erosion claim. I therefore recommend no change to the verdict, and I agree with the reader's identification of the weakest premise.","tokens_in":16489,"tokens_out":3149,"duration_ms":44484,"concrete_test":"Conduct a policy audit in three to five European countries the paper discusses (e.g., Netherlands, Sweden, Germany, Finland, Belgium): collect public procurement documents, service guidelines, and deaf consumer complaint records from 2023–2026 for encounters where interpreter hours were exhausted or interpreters were unavailable. Classify each case as (a) speech-to-text AI imposed as a mandatory default, (b) AI offered as an optional supplement alongside continued interpreter provision, or (c) AI not used. If the dominant pattern is (b) or (c), the 'erosion already underway' premise fails; if (a) appears systematically, the warning is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central warning is that 'the erosion of linguistic rights may already be underway' (Introduction), and the specific mechanism offered is that deaf consumers are 'increasingly directed to rely on speech-to-text AI ... when their allotted interpreting hours are exhausted or when human interpreters are unavailable' (Section 'The intersection of AI sign language technology development', paragraph beginning 'Anecdotal evidence'). This premise is supported only by 'anecdotal evidence from some European countries', with no countries, dates, sample sizes, or citations. That single anecdotal premise carries the causal bridge from 'developers frame AI as a substitute for interpreters' to 'institutions actually impose AI on deaf consumers', which is what makes the warning urgent rather than merely hypothetical. If the real policy pattern is optional supplementation (AI offered alongside interpreters when interpreter supply fails), the slippery-slope conclusion weakens: the eroded right would be the right to a human interpreter specifically, not necessarily the right to communication access. The opening vignette itself shows a deaf person voluntarily choosing a speech-to-text app in a shop; voluntary use is not erosion. The VRI example is a historical analogy, not evidence of current erosion. The EU Accessibility Act discussion is partly prospective and reports that the interpreter shortage 'is frequently used here as a justification' for avatars, again without documenting actual substitution. The framing critique in the Discussion can stand without the erosion claim, but the strongest version of the paper's conclusion depends on the under-supported empirical premise that substitution is already occurring.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a conceptual and position article arguing that the development and deployment of AI sign language technologies are being shaped by an interpreter-centred ideology of access, which frames these technologies as substitutes for human interpreters rather than as tools supporting deaf languaging practices. Drawing on Deaf Studies, Sign Language Interpreting Studies, and crip technoscience, the author introduces the notion of 'slippery slope rights' to warn that incremental acceptance of AI tools in low-stakes contexts may expand into high-stakes settings, eroding hard-won linguistic and accessibility rights. The paper also critiques the use of interpreted datasets in training AI, the lack of deaf leadership in sign language AI research, and the performative nature of much co-creation work. It concludes by calling for deaf-led, human-in-command approaches to AI development.","tokens_in":16677,"tokens_out":2988,"duration_ms":43069,"significance":"If its central claim is accepted, the paper makes an important contribution by reframing sign language AI from a purely technical problem to a question of linguistic rights, access hierarchies, and collective autonomy. It brings crip technoscience into dialogue with sign language interpreting studies in a way that is largely missing from the current AI literature, and it gives explicit attention to the perspectives of deaf users rather than treating them as passive beneficiaries. The author is transparent about her own positionality and about the fact that some of her earlier work is cited by AI developers as motivation, which strengthens the credibility of the institutional critique. The paper also offers a concrete, useful warning about the risks of training sign language AI on interpreted data. At the same time, the empirical basis for the claim that erosion is already underway is thin and regionally narrow; the paper is best read as a well-argued interpretive essay calling for vigilance, rather than as an established empirical account of current policy trajectories.","major_comments":[{"comment":"The central urgency of the paper rests on the claim that 'the erosion of linguistic rights may already be underway' (Introduction) and that deaf consumers are 'increasingly directed to rely on speech-to-text AI' when interpreter hours are exhausted. This load-bearing premise is supported only by 'anecdotal evidence from some European countries,' with no countries, dates, organizations, or documentary sources given. If the paper is to make the strong claim that erosion is currently happening rather than merely possible, that claim needs systematic documentation or at least named, checkable cases. Alternatively, the paper should be reframed as a risk analysis: the warning about a plausible trajectory does not require current prevalence, but the present wording overstates the evidentiary base.","section":"The intersection of AI sign language technology development (paragraph beginning 'Anecdotal evidence')"},{"comment":"The concept of 'slippery slope rights' is introduced as a central analytical device but is not defined with sufficient precision, and the argument that low-stakes AI use will expand into high-stakes contexts is asserted rather than supported. The VRI example is a historical analogy showing that a technology initially rejected by deaf people can become normalized, but it does not demonstrate that AI sign language technologies are following the same path, nor does it specify the causal mechanism by which voluntary low-stakes use becomes imposed high-stakes use. The paper would be strengthened by an explicit account of the conditions under which this slippage occurs (e.g., funding pressures, legal obligations, interpreter shortages) and by evidence of at least one case where such a transition is already observable.","section":"Introduction and Discussion (the 'slippery slope rights' concept)"},{"comment":"The paper generalizes from a small set of Western countries—Europe, Australia, New Zealand, Canada, and the United States—to broad claims about 'deaf people' and 'linguistic rights' worldwide, including the statement that the shift 'could have devastating consequences' for 'the reality for most deaf people around the world.' The argument is explicitly built on the existence of institutionalized interpreting services and legal recognition frameworks, which are not present in most of the world. The paper should either restrict its empirical and predictive claims to the jurisdictions on which it draws, or add a substantive discussion of how the argument transfers to contexts without such legal protections. Without this scope clarification, the central warning risks overgeneralization.","section":"Throughout, but especially the Introduction and Discussion and conclusion"}],"minor_comments":[{"comment":"The keyword 'lanugage' is a typo and should read 'language.'","section":"Keywords"},{"comment":"Several references contain spacing artifacts such as 'V andeghinste', 'Y oung', 'V an', and 'Resarch'; these should be corrected for consistency and professional presentation.","section":"References"},{"comment":"The opening vignette describes a deaf person voluntarily choosing a speech-to-text app in a shop and a court hearing where an AI translation is imposed, yet later the paper states that current AI sign language technologies are mainly targeted at low-stakes contexts. The vignette's court example appears to depict current practice, so the text should clarify whether it is a hypothetical future scenario or a documented case.","section":"Introduction / Discussion and conclusion"},{"comment":"The description of the SignGPT project states that all Principal Investigators are hearing academics; this is an empirical claim about individuals and should be supported by a source or softened to avoid overstatement.","section":"Setting the stage"},{"comment":"The reference to the EU Accessibility Act 'set to take effect in June 2025' would benefit from a more precise legal citation or a note on the transposition deadline, as the practical effect in member states depends on national implementation.","section":"Discussion and conclusion"}],"recommendation":"major_revision","confidential_remarks":"This is a conceptual and policy essay rather than an empirical study, and the review should judge it on those terms. The main weakness is the mismatch between the strongly worded empirical claim that erosion is 'already underway' and the anecdotal evidence offered for it. This is fixable either by adding documentation or by reframing the paper as a risk analysis. The conceptual core is sound and the paper is likely to be a valuable contribution to the emerging literature on sign language AI and linguistic rights."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper makes a real contribution: it gives a name and a frame to something people in the field have been circling. The \"slippery slope rights\" concept is useful, and the application of crip technoscience to sign language AI is overdue. The critique of interpreter-centered benchmarks—training models on interpreted data and evaluating AI against human interpreters—is sharp and well-supported. The author knows the literature and the community, and it shows.\n\nThe paper is best when it argues that AI sign language tools should not be framed as substitutes for interpreters but as supports for deaf languaging. That argument stands on its own and is well made. The discussion of access labour, hierarchy, and the double bind of deaf NGOs is nuanced and genuinely useful.\n\nThe soft spot is exactly where the stress-test note lands: the claim that erosion of linguistic rights is \"already underway\" rests on anecdote. The passage about deaf consumers being directed to speech-to-text when interpreter hours run out has no countries, no dates, no citations. That single anecdotal bridge carries the urgency of the whole paper. If the real pattern is optional supplementation rather than imposed substitution, the \"erosion\" claim weakens. The author should either strengthen this with documented examples or reframe the contribution as a preemptive warning rather than a description of ongoing erosion.\n\nAnother, smaller issue: the paper generalizes from Western countries (Europe, Australia, NZ, Canada, US) to \"deaf people around the world\" at some points, even though the author is careful elsewhere to note the differences. This scope mismatch is fixable with a few careful qualifications.\n\nThese are not fatal. The conceptual argument holds up, and the paper is honest about the limits of the evidence in places—but not honest enough about how much weight the erosion claim carries. A serious revision should address that, not by adding hedge language but by either finding real evidence or explicitly lowering the claim.\n\nWho is this for? Researchers in Deaf Studies, HCI, accessibility, and AI ethics. It would be a good reading group piece. I'd send it to review: the field needs this argument, and the author is well placed to make it. I'd also cite the slippery slope rights framing in my own work.","headline":"A needed and largely convincing conceptual contribution that names the 'slippery slope' of AI sign language tech, but its strongest warning rests on uncited anecdotes and needs either more evidence or a more modest claim.","tokens_in":17225,"tokens_out":1930,"would_cite":true,"duration_ms":23049,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Framing AI sign language tools as substitutes for human interpreters risks eroding deaf people's linguistic rights.","keywords":["AI language technologies","sign language interpreting","linguistic rights","deaf communities","technoableism","access hierarchies","crip technoscience","slippery slope rights"],"falsifier":"Compare national interpreter-hour allocation and deaf consumers' actual use of speech-to-text AI in two or three European countries over several years, holding demand for interpreting constant. If deaf people's access to human interpreters does not decline as AI captioning becomes available—or if interpreter funding stays flat or rises—the paper's central warning about erosion of linguistic rights would lose its empirical basis.","tokens_in":16276,"feed_emoji":"🧏","tokens_out":9488,"duration_ms":97635,"temperature":0.7,"pith_summary":"This paper argues that the way AI sign language technologies are being framed—as potential substitutes for human sign language interpreters—threatens deaf people's hard-won linguistic and accessibility rights. The central warning is that voluntary use of AI tools can slide into imposed use: when interpreter hours run out or budgets tighten, deaf consumers are increasingly directed toward speech-to-text AI, and what begins as a convenience becomes a replacement. The paper claims this substitution framing is built into the technology itself, because sign language AI is trained on interpreted data and benchmarked against interpreters rather than against deaf people's own languaging practices—the ways deaf people actually use signed and written languages. This matters because the erosion may already be underway, and it would hit hardest deaf people without the literacy, resources, and networks that make AI tools feel empowering.","feed_headline":"Framing AI sign tools as interpreter substitutes risks deaf rights","feed_subtitle":"If institutions treat AI as a cheaper stand-in for interpreters, deaf people may lose hard-won access choices","key_machinery":"The paper's load-bearing concept is 'slippery slope rights': measures initially rejected by deaf communities because of their limitations become normalised over time and then accepted as adequate, so that each small adaptation of AI in low-stakes contexts widens the path to imposed use in high-stakes ones. Alongside it is the 'substitute framing'—the persistent benchmarking of AI sign language technologies against human interpreters in research, funding, and deaf-service advocacy—which the paper argues distorts both the technologies and the policy choices around them. The mechanism is carried by three linked observations: AI sign language systems are trained on interpreted datasets that reflect interpreter signing rather than deaf community usage; deaf people already perform invisible 'access labour' adapting to interpreters, a subordination that AI can compound; and deaf-led perspectives are largely absent from sign language AI research and development.","core_discovery":"The paper's central claim is that sign language AI technologies are not, and should not be, competitors with human interpreters; they are meant to support and reflect the diverse languaging practices of deaf communities. Framing them as substitutes sets a benchmark that undermines their purpose and creates a 'slippery slope' in which access rights that deaf communities once rejected as inadequate—such as automated captioning or signing avatars—gradually become normalised and then imposed in high-stakes settings like healthcare and courts. The paper contends that the erosion of linguistic rights may already be underway, driven by governments and institutions that see AI as a cheaper, more efficient alternative to interpreter services, and reinforced by a field of sign language AI development that is dominated by hearing, non-signing actors, trained on interpreted datasets, and lacking deaf leadership.","pith_inferences":["A testable extension would be to track national interpreter-hour allocation and procurement data in several European countries for five years, comparing deaf consumers' use of human interpreting before and after speech-to-text AI is offered, to see whether substitution actually occurs at scale.","The slippery-slope logic suggests that once AI access is embedded as a default feature in mainstream platforms such as video conferencing and phone operating systems, it will be treated as a universal baseline, making it harder for deaf communities to demand interpreter-mediated access even when they prefer it; the paper gestures at this but does not develop the platform-defaults angle.","The same substitute-framing critique could apply to spoken-language interpreting for migrants and refugees, where interpreting is already treated as a temporary measure; a comparative study would test whether deaf and migrant language access are being eroded by the same AI-driven cost logic.","If deaf-led research becomes the norm, one would expect new evaluation metrics for sign language AI that measure alignment with deaf languaging practices rather than equivalence to human interpreters; the paper calls for deaf leadership but does not specify what such metrics would look like."],"forward_implications":["If the substitute framing persists, AI sign language tools will be judged mainly by how well they mimic human interpreters, not by whether they serve deaf users' actual communication needs.","Low-stakes uses such as automated captions in public transport and hospitality will normalise AI access, making it harder for deaf people to refuse these tools when they are proposed in healthcare, legal, and emergency settings.","Deaf people who cannot calibrate their signing or written-language skills to what AI systems can recognise will be left with fewer access options than they have today, widening existing hierarchies among deaf users.","Sign language AI trained on interpreted datasets will encode interpreter-generated signing as the standard, further sidelining regional varieties, Black ASL, queer signing, and idiosyncratic signing styles.","Without deaf leadership in research and development, 'co-creation' will remain performative and will not change the power relations that shape which access problems AI is built to solve."],"supporting_citations":[{"why":"Supplies the deaf-authored claim that deaf people would go without interpreters if they could, grounding the paper's account of ambivalence and resilience.","marker":"Hall, Brick, & Millios (2024)"},{"why":"Argues sign language interpreting is a 'quick fix' for inclusion and is widely cited to motivate AI sign tools, anchoring the access-hierarchy critique.","marker":"(De Meulder & Haualand, 2021)"},{"why":"Systematic review of 101 sign language AI papers documenting systemic biases and the dominance of hearing researchers, the paper's main empirical evidence.","marker":"Desai et al. (2024)"},{"why":"Shows deaf and interpreter associations still privilege human interpreting in high-stakes contexts, supporting the claim that AI is benchmarked against interpreters.","marker":"(WFD & WASLI, 2018)"},{"why":"Demonstrates that interpreters act as de facto language models for deaf children, supporting the linguistic-subordination argument.","marker":"(Caselli, Hall, & Henner, 2020)"},{"why":"Provides the critical disability-technology lens through which the paper analyses how AI redistributes access and reshapes hierarchies.","marker":"(Hamraie & Fritsch, 2019)"},{"why":"Introduces the 'good enough' critique and warns against training sign language AI on interpreted data, a premise the paper extends.","marker":"(De Meulder, 2021)"},{"why":"Reports deaf users' own distinctions between one-way and two-way, low-stakes and high-stakes access, evidence that AI and interpreters serve different purposes.","marker":"(Picron, Van Landuyt, & Omardeen, 2024a)"}],"fun_headline_variants":["AI sign tools as interpreter substitutes erode deaf rights","Deaf linguistic rights erode as AI replaces interpreters","Hearing-led AI sign development sidelines deaf voices","Deaf-led AI, not substitute interpreters, protects sign rights","Don't let AI sign tools replace human interpreters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the assumption that governments and institutions will actually substitute AI for human interpreters at scale; the paper's direct evidence for this is limited to anecdotal reports from some European countries, so if substitution does not materialise, the central warning loses much of its urgency.","fun_headline_variants_meta":{"raw":{"variants":["AI sign tools as interpreter substitutes erode deaf rights","Deaf linguistic rights erode as AI replaces interpreters","Hearing-led AI sign development sidelines deaf voices","Deaf-led AI, not substitute interpreters, protects sign rights","Don't let AI sign tools replace human interpreters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3845,"prompt_tokens":857,"completion_tokens":2988,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2912}},"tokens_in":473,"tokens_out":2988,"duration_ms":24808,"temperature":1.0,"reasoning_tokens":2912,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:49:05.524983+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare national interpreter-hour allocation and deaf consumers' actual use of speech-to-text AI in two or three European countries over several years, holding demand for interpreting constant. If deaf people's access to human interpreters does not decline as AI captioning becomes available—or if interpreter funding stays flat or rises—the paper's central warning about erosion of linguistic rights would lose its empirical basis.","supporting_citations":[],"review_version":1}