Bidet AI — Accessibility / Readability UX Options Beyond OpenDyslexic

Research date: 2026-05-10 Audience: adults with ADHD/ADD, dyslexia, dysgraphia, low vision, fatigue-driven reading difficulty, ELL learners, late-deaf/HoH adults, anyone reviewing their own brain-dump output. Framing rule: college-lecture / adult / professional only — no K-12 audience naming.

TL;DR — top 3 options Bidet should ship in v0.3 alongside OpenDyslexic

Atkinson Hyperlegible (Braille Institute, SIL OFL) added to the font picker as the default hyperlegible option. Stronger evidence base than OpenDyslexic, freely licensed, 150+ language support in Hyperlegible Next, and the Braille-Institute provenance is a credible contest talking-point. OpenDyslexic stays in the picker because users prefer it subjectively, but it should not be the default.
Read-aloud (Android TextToSpeech) on the Clean tabs. Highest-evidence accessibility win on the list (Wood et al. 2018 meta-analysis, weighted ES = 0.35; consistent gains for dyslexia, aphasia, intellectual disability). Cheap to ship, big audience surface, and pairs with the brain-dump premise: speak in, listen back, verify you were understood.
Text-spacing controls (line-height, paragraph-spacing, letter-spacing, word-spacing) meeting WCAG 2.1 SC 1.4.12 minimums as user-adjustable presets. Spacing has stronger empirical support than any "dyslexia font" (Zorzi et al., and the consistent BDA 2023 guidance). One slider — Comfort: Compact / Standard / Spacious — covers most of the population without overwhelming the UI.

These three combined map cleanly onto UDL Guideline 5 (multiple means of expression and communication) for the contest pitch and onto WCAG 2.1 AA for the credibility footer.

Verified options

1. Font choice beyond OpenDyslexic

1a. Atkinson Hyperlegible / Atkinson Hyperlegible Next — SHIP

What / why: Sans-serif designed by the Braille Institute for partially-sighted readers; emphasizes letterform distinction (open counters, distinct b/d/p/q, unambiguous 0/O, 1/l/I). Hyperlegible Next (2025) adds variable weight (7 weights) and 150+ language coverage including a monospaced cut.
Population: low vision, fatigue-driven reading difficulty, late-deaf adults reading TTS-equivalents, anyone on a small phone screen.
Evidence: Won Fast Company Innovation by Design 2019; added to Cooper Hewitt permanent collection 2024; adopted by the Braille Institute itself as official typeface. Distinguishability principle is well-grounded in low-vision typography literature, though there is no large peer-reviewed RCT showing a reading-speed advantage over Arial. The provenance + design rationale is the cite, not an RCT.
Implementation: easy — Google Fonts, SIL OFL, drop-in Compose FontFamily.
Default state: ON as default hyperlegible option.
Concerns: Same caveat as every "accessibility font" — no font is a silver bullet. Pair with spacing controls.

1b. OpenDyslexic — KEEP, but not default

What / why: Weighted-bottom letterforms intended to anchor letters and reduce flipping/rotating. Subjectively preferred by some dyslexic readers.
Population: dyslexic readers who self-report it helps.
Evidence (mixed → mostly negative): Marinus et al. 2016 (Annals of Dyslexia, PMID 26993270): no improvement in reading rate or accuracy in alternating-treatment design; no participant preferred the font. Wery & Diliberto 2017: no benefit. Kuster et al. 2018 (Dyslexie sister-font): no benefit. Counter-positive: Franzen et al. (Annals of Eye Science abstract AB004): claimed efficacy in longer standardised text; abstract-level only. The strongest reading of the literature: the wider letter spacing baked into OpenDyslexic is doing the work, not the letter shapes.
Implementation: easy — SIL OFL, already shipped in v0.2.
Default state: OFF, opt-in in the font picker. Don't lead with it.
Concerns: If we lead with OpenDyslexic in marketing we inherit its mixed-evidence baggage. Lead with Atkinson Hyperlegible + spacing; OpenDyslexic is a respect-the-user choice.

1c. Lexie Readable (K-Type) — SHIP

What / why: Designed by Keith Bates per BDA-aligned principles: non-symmetrical b/d, handwritten a/g, generous spacing, large x-height. Captures Comic Sans's much-loved-by-dyslexics qualities without the comic associations.
Population: dyslexic adult readers, ELL learners, anyone who finds Comic Sans helpful but socially awkward.
Evidence: No head-to-head peer-reviewed RCT, but design principles are explicitly drawn from BDA 2023 style guide; many UK schools and dyslexia advocates recommend it. Subjective-preference evidence only.
Implementation: easy — Regular + Bold are free for personal/educational/charitable use; commercial license £10. We can ship Regular + Bold under the free path; add a small license note in About.
Default state: OFF, opt-in.
Concerns: Italic + Heavy weights are paid; we ship without them.

1d. System default (Roboto on Android) — SHIP as the always-available baseline

What / why: No surprises. User's own system accessibility settings (font scale, bold text) carry over unmodified.
Population: everyone who's already configured Android accessibility.
Evidence: Android Accessibility Suite + system-level font scaling is the platform-recommended baseline; respecting it is itself a WCAG conformance behavior.
Implementation: trivial — just don't override.
Default state: ON as factory default until the user picks something else.

1e. Sans Forgetica — SKIP (see "Skipped" section)

2. Text size / line height / paragraph spacing — SHIP

What / why: Sliders or step controls for: font size (12–24 pt range), line height (1.0×–2.0×), paragraph spacing (0.5×–3.0× font size), letter-spacing (0–0.20× font size), word-spacing (0–0.20× font size).
Population: dyslexic readers (spacing is the most evidence-supported intervention for them), low vision, ADHD readers who lose place mid-paragraph, anyone using the app outside in sunlight.
Evidence:
WCAG 2.1 SC 1.4.12 (Level AA, 2018) mandates the app must not break when users set: line height ≥ 1.5×, paragraph spacing ≥ 2× font size, letter-spacing ≥ 0.12× font size, word-spacing ≥ 0.16× font size. Bidet should not just tolerate these settings — it should offer them as presets. Source: W3C WAI Understanding 1.4.12.
British Dyslexia Association Style Guide 2023: font 12–14 pt (1–1.2 em / 16–19 px), letter-spacing ~35% of average letter width, line spacing 1.5×, left-aligned (not justified), bold for emphasis (avoid italic and underline). This is the most-cited practitioner standard.
Zorzi et al. 2012 (PNAS) found increased letter-spacing improved reading speed and accuracy in dyslexic children. Reproduced in adults across multiple follow-ups.
Implementation: easy — Compose LocalDensity, TextStyle(lineHeight=…, letterSpacing=…), Modifier.padding. Ship 3 presets (Compact / Standard / Spacious) backed by the WCAG-1.4.12 minimums in the Spacious preset, plus a custom-slider screen for power users.
Default state: Standard preset ON by default; don't ship Spacious as the default (looks loose to non-target users).
Concerns: None significant. This is the single most evidence-supported intervention on the list.

3. Color schemes — SHIP a curated set

What / why: Three to four themes:
Light (system default)
Dark (battery + photophobia)
Sepia / cream (warm low-contrast — black text on cream/yellow background)
High-contrast (true black on true white, or true white on true black, no greys)
Population: photophobia / migraine / Irlen-syndrome (use sepia or dark), low vision (use high-contrast), low-light reading (dark), warm-bias dyslexic readers (sepia).
Evidence:
Dark mode + photophobia: Mixed. American Migraine Foundation lists light as the most-common migraine trigger; ~63% of self-reporting users say dark mode reduces screen-related headache symptoms (Slack internal data, theraspecs.com summary). Controlled trials show no objective difference in eye strain (TheraSpecs). Honest framing: subjective relief, not proven physiological effect — but the cost of offering it is near-zero and a meaningful population requests it.
Sepia / warm backgrounds: Rello & Bigham CMU 2017 ("Good Background Colors for Readers"): warm colors (peach, yellow, orange) led to faster reading for both dyslexic and control readers; cool blues led to shorter fixation times. Effect sizes are modest. The "Irlen overlays" literature is genuinely controversial — WHO, AAP, and RANZCO point to lack of reliable evidence for Irlen syndrome as a diagnosis. We should offer the option without claiming it cures anything.
High-contrast: WCAG 2.1 SC 1.4.6 (AAA) requires 7:1 contrast for normal text. Critical for low-vision users.
Implementation: easy — Compose MaterialTheme color schemes; one-tap theme switcher.
Default state: Follow system dark/light setting by default. Sepia + high-contrast are opt-in.
Concerns: Don't market sepia as treatment for dyslexia or visual stress — claim only "many readers find this more comfortable." Don't reference Irlen by name in marketing.

4. Read-aloud (TTS) on the Clean tabs — SHIP (recommended core a11y feature)

What / why: Tap a paragraph or sentence to play it via Android's TextToSpeech engine. Whole-tab "Play All" with pause/skip/resume. Speed control 0.5×–2.0×.
Population: Largest evidence base on this list. Dyslexia, dysgraphia, fatigue-driven reading difficulty, low vision, late-deaf adults reviewing transcribed audio, ELL learners (multilingual TTS voices), anyone reviewing a long brain-dump output without staring at the screen.
Evidence:
Wood, Moxley, Tighe & Wagner 2018 Journal of Learning Disabilities meta-analysis: weighted effect size g = 0.35 for TTS on reading comprehension in students with reading disabilities. (PMC5494021)
Harvey et al. 2023, Reading and Writing: TTS significantly improved comprehension vs. no-TTS, with disproportionate benefits for dyslexic readers.
Knight et al. 2018 (aphasia): TTS rate adjustment improved comprehension when matched to user-preferred WPM.
Svensson et al. 2019, Disability and Rehabilitation: Assistive Technology, 5-year follow-up: TTS adoption durably improved school+work outcomes for dyslexic readers.
Counter-evidence: Floyd & Judge 2012 found no difference for middle-schoolers with reading-related LD. Effects vary by task and population. Net direction: positive.
Implementation: medium — Android TextToSpeech is built-in (Donut+); Compose integration is straightforward. The work is in the UX: highlight the currently-spoken sentence, persist playback state, handle audio focus correctly, support Bluetooth headset transport keys (MediaSession). ~2–4 days of build.
Default state: Feature ON, but no auto-play. A persistent "speaker" icon on each Clean tab, plus per-paragraph tap.
Concerns: Default Android TTS voices vary by device; quality on Pixel is good but on budget phones can be poor. Consider documenting the user-installable Google TTS voice. No content concerns — TTS is purely renderer-side.

5. Reading ruler / current-line highlight — SHIP (medium priority)

What / why: A subtle band that follows the current line/paragraph; everything outside is dimmed. Or a thin underline on the actively-read line, advanced by tap or by TTS playback position.
Population: ADHD readers who lose place mid-paragraph (this is the #1 reported reason to use one), dyslexic readers with tracking difficulties, anyone reading dense text on a phone.
Evidence:
Yu et al. 2023 CHI "Digital Reading Rulers": evaluated 6 ruler designs with 91 dyslexic + 86 non-dyslexic readers; statistically significant increase in reading speed for dyslexic readers; "keep my place" + "focus" were the top reasons cited; a non-trivial subgroup specifically used them for ADHD-related focus needs.
This is one of the few accessibility features with a recent peer-reviewed RCT-style evaluation showing measurable benefit on the target population.
Implementation: medium — Compose Modifier.drawBehind or overlay, plus a touch-driven "active line" state. Slightly more complex if we couple it to TTS playback position.
Default state: OFF, opt-in in the Reading panel. Once on, persists.
Concerns: Some users find it distracting; that's why it's opt-in. Make sure the dimming respects dark/sepia themes.

6. Bionic Reading / first-letters-bold rendering — SHIP cautiously, with honest copy

What / why: Bolds the first 30–50% of each word ("Bionic Reading"). Premise: helps the eye anchor; some readers report subjective focus improvement.
Population: ADHD readers (largest self-report constituency), dyslexic readers (mixed reports), motivated learners.
Evidence (LARGELY NEGATIVE — be honest in copy):
Snell 2024 SSRN preprint "No, Bionic Reading Does Not Work" — direct rebuttal.
Možina, Kovačević & Blaznik 2025 (SAGE Open): eye-tracking study, no significant differences in fixation duration/count between Bionic and standard fonts.
Readwise 2022 study, ~2000 readers: participants read 2.6 wpm slower on average with Bionic; comprehension worse by 5–8 percentage points.
Theaker 2023 (~110 students): Bionic did improve self-reported motivation and self-efficacy in students with LDs — even though objective speed/comprehension didn't improve.
Çakmak 2024: ADHD subgroup showed sustained comprehension at higher reading speeds with Bionic — small but suggestive.
Recent eye-tracking, PMC12565662 (2025) "Guiding the Gaze": Bionic does alter gaze patterns; whether that's helpful is task-dependent.
Net: No evidence Bionic Reading improves objective speed or comprehension for the general population. Some evidence of motivational/self-efficacy gains. Some evidence ADHD readers benefit on focus, not speed.
Implementation: easy — pure CSS/AnnotatedString styling; bold first N characters per token where N = ceil(len(token) * 0.4) or similar. ~half-day of build.
Default state: OFF. Opt-in only.
Concerns: Critical that we don't market this as "scientifically proven." Marketing copy should say something like "Some readers find first-letter emphasis helps them keep focus. Try it on, leave it off — your call."

7. Letter / word spacing controls (separate from font) — covered by Option 2 above

Bundled into the spacing presets. WCAG 1.4.12 already requires support; we just expose it.

8. Auto-summary / TLDR toggle — SHIP as a per-tab toggle, not a new tab

What / why: A "Summarize" toggle on each Clean tab that prepends a 1–2-sentence TLDR generated by the same Gemma 4 model that built the Clean output. Per Mark's spec — toggle, not new tab.
Population: ADHD readers (the centrality-deficit literature is exactly the use case), fatigue-driven readers, anyone re-reading their own brain-dump for verification, ELL adults skim-reading before deep-reading.
Evidence:
Cain & Bignell 2014; Lorch et al. 2010: ADHD readers show "centrality deficit" — recall of peripheral details intact, recall of central ideas impaired. Pre-loaded summaries help anchor the schema before deep reading.
Cognitive Load Theory (Sweller): TLDR-first reduces extraneous load and frees working-memory bandwidth for the actual content. Specifically called out in 2025 ALT Texts review of ADHD digital-learning accessibility barriers.
Counter-evidence: Recent studies on AI summarization show people who use ChatGPT-summaries score lower on comprehension than people who read the source — but this measures replacement, not pre-summary scaffolding. Bidet's design (TLDR plus full output) is the safe pattern.
Implementation: medium — second prompt to Gemma 4 with the cleaned output as input; cache the summary; toggle visibility. ~1 day.
Default state: OFF, opt-in. When on, persists per-tab.
Concerns: Hallucination risk. Summary should never replace the cleaned output; it should sit above it visually with a small "AI summary — verify against full text" footer.

9. Color-coded sections / visual chunking — SHIP (light touch)

What / why: Bold section headers in a subtly-different accent color; bulleted lists with discreet icons; clear visual separation between sections of the cleaned output.
Population: ADHD readers (chunking is one of the most-cited evidence-based ADHD reading strategies), dyslexic readers, ELL learners, anyone scanning before reading.
Evidence:
Cognitive Load Theory — chunking and visual scaffolding reduces extraneous load (Sweller, van Merriënboer, & Paas 1998).
CAST UDL Guideline 2 (Representation — Language and Symbols) explicitly recommends "highlight patterns, critical features, big ideas, and relationships."
BDA 2023 style guide: bold for emphasis (avoid italic/underline), clear headings.
Implementation: easy — Compose Text with section styles; the LLM prompt already produces structured Markdown for the Clean output; we just need to render it well.
Default state: ON.
Concerns: Restraint. Two icon types max, one accent color. Avoid the "Confluence-page-from-2014" look.

10. One-tap copy / share with format-preserving — SHIP

What / why: Two share buttons per tab: "Copy as plain text" and "Copy as formatted (Markdown)." Plus standard Android share-sheet integration.
Population: AAC users pasting into communication apps, anyone moving Clean output into Gmail/Docs/Slack, students pasting study notes into Notion/OneNote, the brain-dump-to-anywhere workflow that is the actual product.
Evidence:
WCAG 2.1 SC 1.3.2 (Meaningful Sequence) and 1.3.1 (Info and Relationships) — preserving structure across boundaries.
AAC tools depend on round-trip-able text; HCI literature on AAC is unanimous on this.
Implementation: easy — ClipData.newPlainText + Intent.ACTION_SEND. ~half-day.
Default state: ON.
Concerns: None.

Mark-specific items

Confirm dual-axis ("interpret you to others / interpret others to you") — KEEP

This isn't a UX widget, it's product architecture, but it pairs with every option above. Every accessibility setting should apply to both axes — the "Clean for me" output and the "Clean for others" output should both honor the user's font/spacing/theme/TTS choices. Internal label: self-directed expression vs. other-directed expression (per UDL 5). User-facing copy: Mark's plain language, unchanged.

Teacher reviewing a student's brain-dump-to-text output for understanding-check

Per Mark's video concept ("Hey, does this kid understand what we just studied?") — this is a reviewer use case, not a learner-accommodation use case. Two features support it without crossing the K-12 marketing rule: - Auto-summary toggle (Option 8) — reviewer turns on TLDR to triage 30 student outputs in 5 minutes. - Side-by-side raw/clean view — reviewer sees the original disjointed transcript next to the cleaned version. Already in Mark's video concept. Make sure it's a first-class screen, not an afterthought.

Both stay safely in the "adult/professional accessibility" framing in marketing — they're features, used in any review context (manager reviewing a team brain-dump, professor reviewing a college student's lecture-notes recap, etc.).

Skipped or low-evidence options

Sans Forgetica — SKIP

Why skip: The marketed memory-boost effect has failed to replicate across multiple peer-reviewed studies. Taylor et al. 2022 (Cognitive Research: Principles and Implications, 4 experiments, ~300 participants): no impact on memory accuracy in DRM paradigm. Wetzler, Pyke & Werner 2021 (SAGE): "Disfluent fonts are not always desirable difficulties" — title is the conclusion. Geller et al. 2020: Sans Forgetica recall was worse than Arial. The original RMIT 2018 study reported a 1% difference vs. Arial.
Risk if shipped: Marketing baggage. Anyone who knows the literature will see Sans Forgetica in the picker as a credibility flag.
If users specifically ask for it later: revisit. Not before.

Irlen-style colored overlays as a marketed feature — SKIP the marketing, KEEP the sepia/warm theme

The sepia theme delivers ~80% of the perceived benefit without invoking Irlen syndrome by name. The Irlen-syndrome diagnosis is rejected by WHO, AAP, and RANZCO; we should not align our marketing with it.
Galuschka et al. 2014 systematic review: weak/inconsistent evidence for colored overlays in reading interventions.
Net: offer the warm theme; don't claim it treats anything.

Custom rainbow-per-letter color schemes (sometimes called "color-coded letter dyslexia tools") — SKIP

No peer-reviewed support; visually noisy; will look unprofessional in screenshots.

Animated text / typing animations / scrolling auto-readers — SKIP

Movement is a known WCAG 2.1 SC 2.2.2 (Pause, Stop, Hide) hazard and a vestibular-disorder trigger. Don't ship.

Forced sentence-by-sentence "reveal one at a time" mode — DEFER

RSVP-style reading apps (Spritz et al.) have an evidence base that is worse than reading rulers — comprehension drops at high WPM. Skip for v0.3; revisit if reading-ruler usage data suggests a power-user subset.

Recommended UX

Settings IA: one new top-level Settings entry, "Reading & Accessibility," with three groups:

Display — Font picker (System, Atkinson Hyperlegible (default), Lexie Readable, OpenDyslexic), Theme (Follow system / Light / Dark / Sepia / High-contrast), Spacing preset (Compact / Standard (default) / Spacious / Custom).
Reading aids — Read aloud (master toggle, voice picker, speed slider), Reading ruler (off / line-highlight / paragraph-highlight), Bionic-style first-letter emphasis (off by default, with an honest one-line caption: "Some readers find this helps focus. Mixed evidence — try it."), Auto-summary toggle (default off, persists per-tab).
Output — Side-by-side raw + clean view (toggle), Copy as plain / Copy as formatted, Share-sheet defaults.

One first-run dialog (skippable): "Bidet works for everyone, but a few quick settings can make it work for you. Want a 30-second tour?" Three swipeable cards: pick a font, pick a spacing, try read-aloud on a sample sentence. Done. No dark patterns; the skip button is the same size as the next button. Settings can be reached from the Clean-tab overflow menu at any time so users who rely on accommodations don't have to dig.

This grouping maps directly onto the contest pitch: UDL Guideline 5 (Expression and Communication, multiple modalities) → WCAG 2.1 AA (1.4.12 spacing, 1.4.3 contrast, 1.4.6 enhanced contrast, 1.3.2 sequence) → BDA 2023 style guide (font + spacing + emphasis defaults). All three are name-droppable in the writeup with citations that hold up.