Mark's Reports

Bidet AI - Deep Research Verdict

Is the verbal-brain-dump-to-customizable-AI-output app already shipping? Does it actually help the populations it claims? What is the realistic ceiling?

2026-05-09 - opinionated research pass, paired against the brain-dump dossier earlier today. Companion piece for the Kaggle Gemma 4 hackathon (5/18) and the DEV.to Gemma writeup (5/24).

Filed 2026-05-09 - sources cited inline, no fabricated studies, no unverified hedges.

1. Is this concept already shipping?

Short answer: partially. The voice-to-AI-output category is crowded. The on-device, accessibility-first, user-customizable-prompt slice is uncrowded - and the four-or-five-product overlap that exists is what Bidet AI has to beat on framing, not just on features.

The closest direct competitors are AudioPen, Mem.ai, Pixel Recorder, Apple Notes (with Apple Intelligence), and Google's own AI Edge Gallery "Audio Scribe" demo. Each one solves a different slice of the same problem.

AudioPen is the most architecturally similar product on the market. It takes voice input, transcribes, and rewrites the result "in the style you choose" - bullet list, casual memo, technical doc, email, Twitter thread, plus user-defined custom styles. Reviewers in 2025 (independent Substack review, Product Hunt 2025 reviews) describe it as the leader of the category. Critically: AudioPen runs in the cloud, does not market itself as an accessibility tool, and the consistent user complaint is that "the rewriting often left out a little bit of nuance" - exactly the failure mode an accessibility-focused user cannot tolerate.

Mem 2.0 shipped a "Voice Mode" in October 2025 that turns brain dumps into structured notes and offers an "Agentic Chat" that edits and organizes notes from voice input. Mem markets to knowledge workers, not accessibility populations, and runs in the cloud.

Pixel Recorder already runs Gemini Nano on-device for transcript summarization and saw a 24% engagement bump after the AI summary feature launched. Coverage of the Pixel 9 release confirms the summarizer now handles 30-plus-minute lectures, on device. This is Bidet AI's most threatening overlap: the device Mark is targeting (Pixel 8 Pro) ships an on-device voice-to-AI-summary stack from Google itself. The differences that matter: Pixel Recorder gives you exactly one output format (a three-bullet summary plus speaker labels), and the user cannot write a custom prompt. It is a fixed pipeline.

Apple Notes with Apple Intelligence records, transcribes, and summarizes audio entirely on-device on iPhone 15 Pro and 16-series hardware. Combined with system-wide Writing Tools (proofread, rewrite, summarize, change tone), Apple has the same architecture Mark is proposing - and shipped it to roughly 100 million devices in 2024-2025. Apple Notes does not, however, expose user-authored custom prompts; the rewrite menu is fixed (Friendly, Professional, Concise).

Google AI Edge Gallery's Audio Scribe is the most exact technical analog to Bidet AI's stack: Gemma 3n / Gemma 4 running locally via LiteRT, transcribing voice, no internet required. It is currently a developer-facing demonstration, not a productized accessibility tool, but anyone with a Pixel 8 Pro can install it from the Play Store today. This is Mark's biggest "are you reinventing the wheel" risk: the underlying capability is downloadable from Google.

The accessibility-marketed alternatives are mostly cloud-based and mostly target the read-side of the problem. Speechify is text-to-speech first and dictation second; it won the 2025 Apple Design Award framed as "a critical resource that helps people live their lives." Homeschooling with Dyslexia's 2026 roundup lists fifteen tools - mostly TTS, OCR, and reading rulers; none of them combine voice-input with user-customizable AI rewrite. UC Berkeley's Disabled Students Program standardized on Otter.ai for note-taking accommodations; UMBC switched from Glean to Otter.ai for the same reason in spring 2025. Otter is cloud, real-time, and oriented around verbatim transcription - it does not offer the four-axis output transformation Bidet AI is proposing.

Dragon NaturallySpeaking, listed by Yale Dyslexia, is the legacy standard for dictation as accommodation. It is dictation-as-input only - it does not rewrite into formats. MacWhisper and Whisper-desktop variants run Whisper locally on Mac and produce raw transcripts; rewrite-into-format is not their job.

What is genuinely missing from the market. No shipping product I can find combines all of (a) on-device inference, (b) accessibility framing as the primary use case, (c) a user-authored custom prompt slot for niche output formats, and (d) the dual-axis split (output to help me understand vs. output to help me be understood). AudioPen has (c). Pixel Recorder has (a) but neither (b) nor (c). Apple Notes has (a) but not (b) or (c). Audio Scribe has (a) but is a demo, not a product. Bidet AI's defensible position is the four-way intersection - and the accessibility framing is what makes the intersection legible.

Verdict: Mark is not reinventing the wheel; he is building a wheel for a market that the existing wheels do not serve. But the gap is narrower than the brain-dump dossier suggested. Apple, Google, and AudioPen each cover three of the four pillars. The custom-prompt slot plus the explicit accessibility framing is the slot that is open.

2. Empirical evidence the design helps the named populations

Short answer: dyslexia and dysgraphia have the strongest evidence base; ELL is solid; dysarthria and stuttering have caveats; the literature distinguishes input-side benefit (dictation as bypass) from output-side benefit (AI reformatting), and the second is genuinely under-studied.

Dyslexia and writing-affecting LD - strongest evidence. Charles MacArthur and Albert Cavalier's 2004 paper in Exceptional Children remains the anchor study: 31 high-school students, Dragon NaturallySpeaking v4, scribe-dictation, and handwriting compared. Students with LD produced higher-quality essays with fewer word-level errors when dictating than when handwriting (NCEO citation, MacArthur & Cavalier 2004; ERIC EJ696633). MacArthur's 2009 review in Learning Disabilities Research & Practice ("Reflections on Research on Writing and Technology for Struggling Writers") generalizes: speech recognition lifts transcription off the cognitive floor and frees working memory for higher-order composition. The 2022 scoping review by Lindeblad et al. in Disability and Rehabilitation: Assistive Technology (full article, Taylor & Francis) covers the post-2010 literature and concludes STT improves text quantity and quality for students with writing-affecting LD. The University of Minnesota's NCEO Accommodations Toolkit (Speech-to-Text Research) is the practitioner-facing distillation of the same conclusion. The five-year follow-up by Svensson et al. in Disability and Rehabilitation: Assistive Technology (2023, Taylor & Francis) caveats: STT shows "mixed utility" in the long run, and continued use depends on classroom context and student self-acceptance more than on the tool itself.

Dysgraphia. Treated in the literature as a subset of writing-affecting LD; the same MacArthur evidence base applies. There is no separate dysgraphia-specific RCT of STT-plus-AI-rewrite that I can verify, which is itself a useful finding: the bulleted-output mode Bidet AI proposes for dysgraphic users is intuitively defensible from the cognitive-load argument (Sweller; MacArthur 2009), but it is not yet proven by a peer-reviewed trial. Mark should not claim it is.

English language learners. The strongest study is the elementary-school comparison of dictation, STT, and handwriting on ELL composition (Effects of Dictation, Speech to Text, and Handwriting on the Written Composition of Elementary School English Language Learners): both dictation and STT produced higher text quality and lower error rates than handwriting. The 2023 mixed-methods study in Frontiers in Psychology on ASR and EFL pronunciation/speaking (Frontiers in Psychology, 2023) and the 2025 RCT on AI-driven speech recognition in EFL listening comprehension (Humanities and Social Sciences Communications, Nature Portfolio, 2025) both show measurable improvement. The SIOP framework (CAL Solutions SIOP overview) does not study Bidet-style tools directly; it provides the pedagogical scaffolding into which a plain-language-output mode would fit.

Dysarthria - the most important caveat. The straight Whisper failure mode for dysarthric speech is the documented Achilles heel. Off-the-shelf Whisper produces 65-76% WER on dysarthric speech without adaptation; fine-tuned Whisper on dysarthric data drops to roughly 18% WER (EUSIPCO 2025, "Improved Dysarthric Speech to Text Conversion via TTS Personalization"; arXiv:2510.04219, "Probing Whisper for Dysarthric Speech in Detection and Assessment"). Google's own Project Euphonia reports that personalized ASR models trained on Euphonia data can outperform human transcribers for individuals with disordered speech - but the personalization step requires per-speaker enrollment data. The 2025 Frontiers in Language Sciences article documents the Euphonia dataset reaching one million utterances. Implication for Bidet AI: a vanilla Whisper-tiny on a Pixel will fail dysarthric users badly, and Mark cannot claim dysarthria support without either Euphonia integration, per-speaker fine-tuning, or an honest disclaimer. The "clinical-grade form that preserves disfluencies" axis is defensible; the "clean-up dysarthric speech for downstream comprehension" axis depends on a much harder ASR problem than Whisper-tiny solves alone.

Stuttering. The 2024 Mandarin Stuttering Event Detection challenge (arXiv:2409.05430) and the FluencyBank Timestamped dataset (Journal of Speech, Language, and Hearing Research, 2024) are the two recent benchmarks. Apple's machine learning research blog documents work on improved speech recognition for people who stutter. The literature converges: ASR for stuttering is improving but still imperfect; an "intended-speech" output mode (smooth out repetitions and blocks) is technically tractable but ethically debated - many people who stutter consider their disfluencies part of how they speak, not a defect to be erased. The clinical-grade preservation mode Mark proposes is the safer default. The "smooth-out" mode should be opt-in, never on by default.

UDL framework. CAST's UDL Guideline 5: Expression & Communication is the non-medical pedagogical anchor: "There is no medium of expression that is equally suited for all learners or for all kinds of communication." Sub-points 5.1 (multiple media for communication) and 5.2 (multiple tools for construction) are the exact principles Bidet AI's two-axis design implements. This is the framing Mark should lead with in the contest pitch - it routes around clinical labels (which his hard rule bans in user-facing copy) and lands the same point with curriculum-designer credibility.

3. Strongest counter-argument the product would not help

Six steel-manned objections, ranked by how much they should worry Mark.

3a. The dysarthric-speech failure mode is real and documented. Vanilla Whisper hits 65-76% WER on dysarthric input (EUSIPCO 2025). Anyone with severe enough dysarthria to actually need this product cannot be served by Bidet AI as currently described. Mark's options: ship without dysarthria support and say so honestly, integrate Project Euphonia's personalized models (Google Research blog), or build a per-speaker fine-tune flow. Pretending the problem does not exist would be the worst move; reviewers will catch it.

3b. Hallucination liability for clinical and legal output formats. The medical AI scribe industry is currently in court. Fisher Phillips' April 2025 alert documents the Sharp HealthCare class action over ambient AI recording. MICA Insurance reports hallucination rates of 1-3% for ambient AI scribes - low until you multiply by encounter volume. Stanford's benchmarking of legal-model hallucinations found 1-in-6 hallucination rates on legal queries even for purpose-built legal models. If Bidet AI ships a "SOAP note" preset to clinicians, Mark inherits a slice of this liability surface. Mitigation: explicit on-device-only marketing, no auto-send, every output requires user review-and-edit, and the custom-prompt slot must include a non-removable "this output may contain errors; review before relying on it" footer.

3c. Prompt injection in the user-supplied prompt slot. OWASP LLM01:2025 ranks prompt injection as the number-one risk for LLM applications. The custom-prompt slot Mark wants is the textbook attack surface: a user could paste a prompt that instructs the model to ignore safety guardrails or reveal training-data fragments. The good news for Bidet AI: on-device, single-user, no privileged tools means the blast radius of a successful injection is low - the model has nothing to leak that the user did not already give it. The bad news: a malicious prompt copy-pasted from a forum could still produce harmful output (medical misinformation framed as a SOAP note, for example) that the user trusts because the app produced it. Mitigation: a fixed system prompt that is concatenated after the user prompt rather than before, plus output-side disclaimers.

3d. Cognitive offloading and skill atrophy. The Risko & Gilbert framework ("Cognitive Offloading," Trends in Cognitive Sciences, 2016) and the Grinschgl, Papenmeier, & Meyerhoff replication ("Consequences of cognitive offloading: Boosting performance but diminishing memory," 2021) document the trade-off: external tools improve immediate task performance but reduce unaided recall of the offloaded content. Applied to Bidet AI: a college student who never writes their own outlines, only dictates and accepts AI organization, may end up worse at unaided organization. The 2024 paper in Cognition ("Cognitive offloading is value-based decision making") reframes this as a deliberate trade users make - which is the right framing for accessibility tech. The honest answer for Mark to use in the contest pitch: Bidet AI is a compensatory accommodation, not a skill-building tool; the same reasoning that justifies a wheelchair justifies the offload. The dyslexia literature (five-year follow-up, Svensson et al. 2023) supports treating compensation as the correct primary goal.

3e. Apple and Google have already shipped most of this. See section 1. The "ten-million-user feature drop" risk - Google ships a "custom prompt" slot in Pixel Recorder next quarter and Bidet AI's differentiation collapses overnight - is real. Apple Writing Tools already exposes a "Describe your change" custom-prompt slot in iOS 18.2. The defensible moat is not the feature; it is the framing, the open-source license, and the accessibility-population focus that the platform vendors will not prioritize because their TAM math points elsewhere.

3f. The output-side AI-rewrite evidence is thinner than the input-side dictation evidence. The peer-reviewed work on dictation-as-accommodation is solid. The peer-reviewed work on AI-rewrite-as-accommodation barely exists - because the technology is two years old and the IRB-approved RCTs have not run yet. Bidet AI is partly research-grade speculation about an effect that has been measured for the input half of the pipeline but not the output half. Mark should say this out loud in the contest writeup; reviewers will respect it more than over-claiming.

4. Is the two-axis split the right framing?

Short answer: keep it. The literature does not have a better label, and the dual-axis maps cleanly onto UDL Guideline 5 plus AAC's input/output bidirectionality - both of which are credentialed framings Mark can cite.

The CAST UDL framework (UDL Action & Expression principle) treats expression as a single principle with multiple media. Bidet AI's Axis A (help-me-understand-what-I-am-producing) maps onto UDL 5.1 "use multiple media for communication" - the user is the audience for their own output. Axis B (help-me-be-understood) maps onto UDL 5.2 "use multiple tools for construction, composition, and creativity" - the output is for someone else.

The AAC literature handles the same input/output asymmetry through speech-generating-device design. Augmentative and Alternative Communication Advances (PMC, 2019) distinguishes input methods (touch, eye-gaze, breath, BCI) from output methods (text-to-speech synthesis, message banking, partner-assisted scanning). The Bidet AI dual-axis is the same pattern at a higher level of abstraction: voice as input, multiple-format text as output, and the user picks the output configuration based on whether the receiver is themselves (Axis A) or a third party (Axis B).

The simpler framing - "this is just a multi-output transformer" - misses what makes the design defensible. A multi-output transformer does not need an accessibility justification. The dual-axis splits the justification cleanly: "because the user might be the audience" is a fundamentally different argument than "because the user might not be the audience." Each axis has its own evidence base in the literature, its own population, its own failure modes, and its own UI affordances. Collapsing them into one axis would force Mark to defend a unified theory that does not exist.

One refinement. The literature would call this "self-directed expression" vs. "other-directed expression," not "help me understand" vs. "help me be understood." If Mark wants academic-language credibility in the contest write-up, those are the phrases to use. Internally and in the product UI, Mark's plain-English phrasing is better.

Verdict: keep the dual-axis. Re-label the internal documentation with the UDL/AAC vocabulary so the contest reviewers see a well-grounded design, not a marketing taxonomy.

5. Realistic impact ceiling

Short answer: Bidet AI can meaningfully serve roughly the population of college students plus literate adults with mild-to-moderate writing-affecting LD, attention-pattern differences, or ELL status who already own a recent Android phone. That is a real, large, underserved audience. It is not, and cannot be, a tool for every accessibility population.

Populations Bidet AI can meaningfully help, given the design as specified:

Populations Bidet AI cannot meaningfully serve, and Mark should not claim to:

Quantitative ceiling estimate. Worldwide, dyslexia prevalence runs 5-10% of the population; ELL students in the U.S. number roughly 5 million in K-12 alone (NCES Condition of Education); attention-pattern adults are perhaps 4-5% of the adult population by clinical estimates. Of that combined audience, the fraction with a Pixel 8 Pro or comparable Android device is small but growing - call it 5-15% of the addressable accessibility audience in 2026, larger by 2028 as Gemma 4 E4B-class capability percolates down to mid-tier hardware. The realistic ceiling for Bidet AI in its first 18 months is tens of thousands of users, not millions, and the product wins by being the trusted on-device option for that audience, not by competing with Pixel Recorder on volume.

Headline read

Mark is not reinventing the wheel; he is building the wheel that the platform vendors and accessibility-tool vendors have each declined to build, for reasons that are obvious once you see the four-way intersection. Apple Notes, Pixel Recorder, AudioPen, and Mem.ai each cover three of the four pillars (on-device, custom-prompt, accessibility-framed, dual-axis output). None covers all four. The gap is real but narrower than the brain-dump dossier framed it - on-device transcription with AI rewrite is now table stakes, and what Bidet AI sells is the framing plus the open-source plus the population focus. Keep the dual-axis split; relabel internally as self-directed vs. other-directed expression to give contest reviewers the UDL/AAC vocabulary they expect. Lead the contest pitch with the dyslexia and ELL evidence, the UDL Guideline 5 framing, and an honest disclaimer about dysarthria and AI-rewrite evidence still being thin. The realistic ceiling is tens of thousands of meaningfully served users in year one, growing as Gemma-class capability reaches mid-tier hardware. That is a real product, a real population, and a defensible thesis - not a unicorn pitch and not a wasted effort.