{"id":"f1eff3e7-3b19-4e5a-8a77-d7171af54adf","arxiv_id":"2505.23035","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"People adapt their speech to AI assistants with short, explicit, polite-free phrases, a pattern the paper names Machine-Facing English and organizes into five traits.","lead":"This paper proposes Machine-Facing English, a register people use when speaking to AI assistants, with five traits such as over-explicit phrasing and shortened commands. It is a conceptual guide for designers and language teachers, but its supporting observations are anecdotal and not released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's empirical backbone is unverifiable: every quantitative claim (37%, 94%, 22%, 15%, 1.8s vs 0.6s) lacks a protocol, dataset, or analysis description, and Section 2.1 explicitly disclaims a corpus while Section 3.3 cites one.","rationale":"The reader's weakest assumption—that the undocumented qualitative observations and the 3,336-command micro-corpus may not exist or may not have been analyzed as reported—is precisely the load-bearing weakness. The five traits named in the paper are presented as observed regularities, and the quantitative figures are embedded in the trait descriptions as evidence of their reality and effect. Because the paper supplies no protocol, data, or analysis code, a skeptical reader cannot distinguish a well-grounded empirical finding from an illustrative narrative. The contradiction between Section 2.1 ('does not present a standalone corpus') and Section 3.3 (a micro-corpus of 3,336 commands) is a concrete textual marker of this gap. The proposed check directly tests whether the reported metrics are reproducible; if they are not, the empirical supporting structure for the central claim disappears. The conceptual framing of MFE as a possible register remains useful as a hypothesis, but the paper's own conclusion overstates the evidentiary basis. Therefore, the reader's REJECT verdict should stand unless the underlying data and methodology are released and verified.","tokens_in":6935,"tokens_out":2726,"duration_ms":31342,"concrete_test":"Request the raw de-identified command logs, acoustic measurements, and session notes underlying Sections 3.1, 3.3, 3.4, and 4.1, together with exact operational definitions (e.g., how 'replacement' was counted, how pitch range was computed, the baseline for the 37% figure). Independently recompute each reported statistic and have two coders re-run the thematic analysis from the raw transcripts to assess inter-rater agreement. If any statistic cannot be reproduced or the data are not available, the empirical claims should be withdrawn and the paper reframed as a hypothesis-generation proposal rather than an empirical register description.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that sustained human-AI interaction normalizes a stable register called Machine-Facing English rests on empirical support that cannot be checked. Section 2.1 states the study 'does not present a standalone corpus,' yet Section 3.3 reports a 'micro-corpus of 3,336 commands' in which '94 percent of participants replaced email with message within five interactions.' No collection protocol, inclusion criteria, annotation scheme, or inter-rater reliability is given; there is no way to know what was counted, over how many participants, or whether the replacement was per user or per command. Section 3.1 reports a '37 percent' reduction in mis-parsed reminders after a single anecdotal alarm-setting episode, with no baseline or sample size. Section 3.4 reports an acoustic comparison with 'a 22 percent reduction in pitch range and a 15 percent slower articulation rate,' but no recording setup, speaker demographics, utterance counts, or statistical tests are supplied. Section 4.1 cites 'log telemetry reported in the literature' for 1.8s versus 0.6s post-response silence, but no specific study is identified for that measurement. The internal contradiction between Section 2.1 and Section 3.3 is not cosmetic: the reported numbers are the only quantitative support for the five traits, and if they are unreproducible or selectively assembled, the empirical claim collapses into anecdote. The conceptual taxonomy may remain plausible, but the paper's stated empirical conclusion—that MFE is a register 'normalized' by interaction—is not independently verifiable from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes \"Machine-Facing English\" (MFE) as an emergent register shaped by sustained human-AI interaction, and identifies five recurrent traits: redundant clarity, directive syntax, controlled vocabulary, flattened prosody, and single-intent structuring. The account is anchored in register theory, enregisterment, audience design, and Gricean pragmatics, and the authors report qualitative observations from bilingual Korean/English product-testing sessions plus several quantitative metrics (37%, 94%, 22%, 15%, 1.8s vs. 0.6s). The drafting process was assisted by an NLD-P-based assistant, which is transparently disclosed.","tokens_in":7292,"tokens_out":5167,"duration_ms":52850,"significance":"If accepted as a descriptive taxonomy, the paper offers a useful vocabulary for a growing phenomenon and connects it to established sociolinguistic and pragmatic frameworks. The reflexive disclosure of AI-assisted drafting and the explicit call for future empirical validation are strengths. However, the paper's quantitative claims are load-bearing for the empirical assertion that human-AI interaction \"normalizes\" MFE, and those claims lack any described measurement protocol. As written, the contribution is best understood as a conceptual hypothesis with illustrative observations, not as an empirically established register description.","major_comments":[{"comment":"The manuscript states in Section 2.1 that it \"does not present a standalone corpus,\" yet Section 3.3 reports a \"micro-corpus of 3,336 commands\" in which \"94 percent of participants replaced email with message within five interactions.\" No collection protocol, inclusion criteria, annotation scheme, or inter-rater reliability is provided, and Section 5.4 repeats that no dedicated annotated corpus currently exists. The internal contradiction must be resolved, and the 94% statistic must either be traceable to a described dataset or removed.","section":"§2.1, §3.3"},{"comment":"The 37% reduction in mis-parsed reminders reported in Section 3.1 and the 22% pitch-range reduction with 15% slower articulation rate reported in Section 3.4 are presented as measured outcomes, but no baseline, sample size, recording setup, speaker demographics, or statistical tests are given. These figures are load-bearing for the traits of redundant clarity and flattened prosody; as written they should be labeled as illustrative anecdotes rather than reported as quantitative findings.","section":"§3.1, §3.4"},{"comment":"The claim that \"log telemetry reported in the literature\" gives an average post-response silence of 1.8 seconds for AI-mediated turns versus 0.6 seconds for human baselines is unattributed. No specific study is cited for this measurement, so the claimed threefold increase cannot be checked. The number should either be supported by a precise citation or removed.","section":"§4.1"},{"comment":"The abstract's assertion that \"sustained human-AI interaction normalizes\" the five MFE features is stronger than the evidence presented in the body. A revision should either supply the missing empirical support or explicitly reframe the central claim as a working hypothesis whose validation is left to future work, consistent with the limitation statement in Section 5.6.","section":"Abstract"}],"minor_comments":[{"comment":"The citations \"Jun (2024)\" and \"Kim & Su (2024)\" do not appear in the reference list, while several listed references (Cohn & Zellou 2021; Cohn et al. 2024; Van Engen & Bradlow 2007; Ward & Tsukahara 2000) are not cited in the text. The bibliography needs a careful pass.","section":"§5.3 / References"},{"comment":"The excerpt from Evalyn's system prompt is introduced with a pseudo-coded block, but the meaning of fields such as [IDENT], version = 1.8, and [RULE:AUTHORIAL_MASTERY] is not explained; a short sentence clarifying that this is illustrative configuration notation would improve readability.","section":"§6.2"},{"comment":"Table 1 is labeled as \"system-responsiveness patterns,\" but the table itself only lists utterance pairs without any system-response data; the caption should be revised to describe the content accurately as illustrative MFE versus naturalistic phrasing.","section":"§3.6"}],"recommendation":"major_revision","confidential_remarks":"The quantitative claims in Sections 3.1, 3.3, 3.4, and 4.1 are the main obstacle to publication. I would ask the authors either to provide the underlying data and a full measurement protocol or to remove the numbers and reframe the paper as a purely conceptual proposal. The Section 2.1/3.3 contradiction is serious enough that the editor may wish to request a statement of data provenance before sending the paper back for revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the five traits are a plausible synthesis of existing findings, but the paper's quantitative claims cannot be checked. Section 2.1 says no standalone corpus exists; Section 3.3 reports a micro-corpus of 3,336 commands with 94% replacement. No protocol, data, or inter-rater reliability. The same goes for the 37% reduction, the 22% pitch range reduction, the 15% slower articulation, and the 1.8s vs 0.6s silence telemetry, the last of which is attributed to 'literature' without naming a study. This is the entire empirical support for the claim that MFE is a normalizable register, so it matters.\n\nWhat's actually new: little. The five traits restate known work on machine-directed speech, clear speech, and prompt engineering (Cohn & Zellou 2021; Reitter et al. 2006; Apple/Google HIG). The contribution is the packaging—a single label plus a Hallidayan framing. That is fine as a teaching device, less so as a research result. The reflexive NLD-P drafting is gimmicky, though the authors do disclose it.\n\nFair credit: Sections 5.4 and 5.6 explicitly acknowledge the lack of a corpus and the need for validation. That honesty makes the unbacked numbers worse, because the authors clearly know the standard.\n\nThe citation pattern is another soft spot. The Korean studies (Lee 2019; Park & Kim 2021; Kim & Lee 2020; Choi 2022) have thin bibliographic detail; I could not verify them, and they appear too conveniently to confirm each trait. Before anything else I'd ask for a full reference audit.\n\nBottom line: this is a reasonable theoretical proposal, but as an empirical paper it is not reproducible. I'd still send it to review—the register question is timely, and a serious referee could push the authors to either drop the unsubstantiated precision or run a real study—but it needs major revision, and I would not bet on salvage as-is.\n\nThe paper is for HCI and sociolinguistics readers who want a vocabulary for 'prompt-ese,' and for instructors looking for a compact framework. Not for anyone needing replicable evidence.\n\nMy recommendation: engage with it, but demand the methodology report and raw data before believing any of the numbers.","headline":"Plausible taxonomy, but the quantitative claims are unverifiable and internally contradicted by the paper's own disclaimers.","tokens_in":7803,"tokens_out":3776,"would_cite":false,"duration_ms":35236,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that habitual interaction with AI assistants is crystallizing a distinct English register, Machine-Facing English, whose five traits improve machine parseability but compress expressive range.","keywords":["Machine-Facing English","register","human-AI interaction","enregisterment","pragmatics","communication accommodation","voice assistants","language variation"],"falsifier":"A preregistered corpus study of AI-mediated voice and text commands from diverse users would settle the claim: if the five traits do not cluster together as a recognizable register—if redundancy, directive syntax, and lexical convergence are better explained by task type or interface design—MFE-as-register would be weakened. The same study could check whether the reported effects (a 37 percent drop in mis-parsed reminders, a 22 percent pitch-range reduction, a 94 percent lexical convergence to 'message') replicate.","tokens_in":6764,"feed_emoji":"🤖","tokens_out":6976,"duration_ms":63674,"temperature":0.7,"pith_summary":"The paper claims that people who regularly address voice assistants and chatbots are collectively evolving a new variety of English, which it calls Machine-Facing English (MFE). It identifies five recurring traits—redundant clarity, directive syntax, controlled vocabulary, flattened prosody, and single-intent structuring—and argues these are adaptive responses that make commands parse more reliably while narrowing the expressive resources of everyday speech. The argument is grounded in register theory, enregisterment, audience design, and communication accommodation theory, with qualitative observations from bilingual (Korean/English) product-testing sessions and a micro-corpus of 3,336 commands. If MFE is real and stable, it is not a user failure or a temporary workaround but a language change driven by technological affordances, with direct implications for interface design and language teaching.","feed_headline":"Talking to AI is hardening into a new English register","feed_subtitle":"Five recurring traits make commands easier for machines to parse but leave less room for natural speech.","key_machinery":"The analytical machinery is the sociolinguistic concept of register—language shaped by situation—specifically Halliday's Field–Tenor–Mode model (topic, participant relations, channel), Agha's enregisterment (the process by which repeated usage stabilizes a style into a recognized register), Bell's audience design, and Giles and Ogay's communication accommodation theory. The paper uses this machinery to reinterpret scattered user-interface observations—redundant alarm-setting, imperative commands, lexical drift, flattened pitch, and split intents—as surface signs of one coherent register, with feedback loops from the machine accelerating stabilization. It also introduces Natural Language Declarative Prompting (NLD-P) as the reflexive method by which the authors drafted the paper itself with AI assistance under human curation.","core_discovery":"On the paper's terms, Machine-Facing English is an emergent, stable register that arises when sustained human–AI interaction normalizes syntactic rigidity, pragmatic simplification, and hyper-explicit phrasing. The five named traits—redundant clarity, directive syntax, controlled vocabulary, flattened prosody, and single-intent structuring—are each tied to documented user behaviors: announcing every temporal parameter after a failed alarm command, replacing hedged requests with verb-first imperatives, converging on system-preferred vocabulary, dropping pitch range and slowing articulation when the assistant mishears, and splitting a complex reminder into two discrete commands. The paper reads these behaviors through Halliday's Field–Tenor–Mode model and Agha's enregisterment to conclude that MFE is a crystallized register that feeds back into system design, pedagogy, and linguistic theory, and that it embodies a persistent tension between communicative efficiency and linguistic richness.","pith_inferences":["A testable extension the paper does not run: a large corpus of AI-mediated commands from users with different self-reported AI exposure should show MFE tracers (slot-filling syntax, redundancy, flat prosody) increasing with exposure, controlling for task type and interface.","The Korean/English bilingual data imply MFE will take language-specific shapes—honorific and particle omission in Korean, word-order and prepositional phrasing in English—so a multilingual comparison could determine which MFE traits are universal and which are typologically mediated.","Because the manuscript was itself drafted with AI assistance under NLD-P, the paper's own prose could be treated as a naturalistic MFE sample; an independent stylistic analysis of its composition would be an informal check on whether the described register actually shows up in the writing."],"forward_implications":["Interface designers can reduce users' cognitive overhead by making parsers tolerant of paraphrase, regional variation, and mild disfluency, rather than requiring users to produce command syntax.","Voice interfaces should give real-time feedback and incremental confirmation ('Did you mean the Beatles?') to prevent costly full restarts and long post-error silences.","Language teaching can use MFE's explicit, modular structures as scaffolding for temporal expressions and imperative syntax, but must balance them with open-ended practice so learners retain indirectness and rapport management.","If MFE is a stable register, enregisterment loops will keep strengthening it as AI use grows, making the efficiency-versus-richness tension a permanent feature of human–AI communication.","Longitudinal tracking could reveal whether MFE conventions bleed into human–human discourse over time, signaling broader linguistic influence."],"supporting_citations":[{"why":"Supplies the Field–Tenor–Mode register framework that organizes the whole account.","marker":"Halliday 1985, 2006"},{"why":"Provides audience design, the idea that speakers shape style to their addressee, which grounds MFE as audience adaptation.","marker":"Bell 1984"},{"why":"Provides enregisterment, the process by which repeated usage stabilizes a style into a recognized register.","marker":"Agha 2003"},{"why":"Supplies communication accommodation theory, explaining hyper-articulation and dropped politeness toward machines.","marker":"Giles & Ogay 2007"},{"why":"Supports the redundancy trait by showing speakers supply surplus information absent shared mental models.","marker":"Clark 1996"},{"why":"Grounds the directive-syntax trait in speech act theory, where imperative form carries unambiguous illocutionary force.","marker":"Searle 1969"},{"why":"Supplies the lexical-convergence model behind the controlled-vocabulary trait.","marker":"Reitter, Moore, & Keller 2006"},{"why":"Provides the Cooperative Principle whose breakdown explains erosion of implicature in machine-facing speech.","marker":"Grice 1975"},{"why":"Informs the flattened-prosody trait through clear-speech research linking intelligibility pressure to reduced prosodic richness.","marker":"Schwab & Zellou 2020"}],"fun_headline_variants":["AI dialogue crystallizes a new English register","Machine-Facing English: a hybrid register from AI talk","How AI conversations are reshaping English into a leaner register","Five traits mark Machine-Facing English, a new register","English gains a new register shaped by AI interlocutors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The account stands on undocumented qualitative observations and a 3,336-command micro-corpus for which the paper provides no protocol, raw data, or inter-rater reliability; if those data are not reproducible or were selectively assembled, the numerical support for the five traits collapses, though the traits themselves might still be plausible.","fun_headline_variants_meta":{"raw":{"variants":["AI dialogue crystallizes a new English register","Machine-Facing English: a hybrid register from AI talk","How AI conversations are reshaping English into a leaner register","Five traits mark Machine-Facing English, a new register","English gains a new register shaped by AI interlocutors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2604,"prompt_tokens":930,"completion_tokens":1674,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":1611}},"tokens_in":546,"tokens_out":1674,"duration_ms":9944,"temperature":1.0,"reasoning_tokens":1611,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:54:56.030482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A preregistered corpus study of AI-mediated voice and text commands from diverse users would settle the claim: if the five traits do not cluster together as a recognizable register—if redundancy, directive syntax, and lexical convergence are better explained by task type or interface design—MFE-as-register would be weakened. The same study could check whether the reported effects (a 37 percent drop in mis-parsed reminders, a 22 percent pitch-range reduction, a 94 percent lexical convergence to 'message') replicate.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Field–Tenor–Mode register framework that organizes the whole account."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides audience design, the idea that speakers shape style to their addressee, which grounds MFE as audience adaptation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides enregisterment, the process by which repeated usage stabilizes a style into a recognized register."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies communication accommodation theory, explaining hyper-articulation and dropped politeness toward machines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the redundancy trait by showing speakers supply surplus information absent shared mental models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grounds the directive-syntax trait in speech act theory, where imperative form carries unambiguous illocutionary force."},{"cited_title":"D., & Keller, F","cited_arxiv_id":null,"evidence_quote":"Supplies the lexical-convergence model behind the controlled-vocabulary trait."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Cooperative Principle whose breakdown explains erosion of implicature in machine-facing speech."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Informs the flattened-prosody trait through clear-speech research linking intelligibility pressure to reduced prosodic richness."}],"review_version":1}