Welcome to BiasLens

An algorithmic bias testing and accountability platform anchored in the South African Draft AI Policy, the EU AI Act, the Amnesty Algorithmic Accountability Toolkit, and the IBM bias taxonomy. To begin, choose the role that fits you best — the interface adapts.

New here? Take the self-guided tour

A short walkthrough explains what every screen does and how to demo it. No prior knowledge of the EU AI Act or POPIA needed.

What BiasLens does

Five connected functions: a guided bias risk questionnaire, a fairness metrics calculator, a four-column compliance mapper across SA, EU and international human rights law, a PAIA / FOI request generator, and a 23-section AIA / FRIA builder. Every screen is built to WCAG 2.2 Level AAA and applies the seven Universal Design principles.

Dashboard

Sample data loaded

Bias risk questionnaire

Function 1 of 5

Answer 12–14 questions about the AI system. Get a one-page bias risk report graded against the IBM 8 bias types and the SA Draft AI Policy 6 pillars, with dual SA tier + EU AI Act classification.

Fairness metrics calculator

Function 2 of 5

Compute four fairness metrics — demographic parity, equal-opportunity difference, the four-fifths rule, and statistical parity — each tagged to the EU AI Act provisions they speak to.

Compliance mapper

Function 3 of 5

Map the system's features against three legal layers: South Africa, the EU AI Act, and international human rights law. Flags compliance gaps in BeAccessible Mid Blue with a remediation note.

PAIA / FOI request generator

Function 4 of 5

Generate a ready-to-send PAIA Section 11 request, an EU access-to-documents request under Regulation (EC) 1049/2001 + EU AI Act Article 27, or a parallel SA + EU evidence package. Track responses in the evidence log.

AIA / FRIA builder

Function 5 of 5

A 23-section wizard. Export as a South African Algorithmic Impact Assessment (AIA), an EU AI Act Article 27 Fundamental Rights Impact Assessment (FRIA), or a unified document covering both jurisdictions.

My assessments

How to demo BiasLens

Self-guided tour

A walkthrough you can run alone in 5, 15, or 30 minutes. Built for four audiences: government procurement, corporate compliance, civil society, and affected individuals or communities. Every step tells you exactly what to click.

Choose your tour

What is BiasLens?

BiasLens helps you investigate, audit, document, and remedy algorithmic bias in AI systems. It is anchored in four frameworks:

  1. The South African Draft National Artificial Intelligence Policy — Cabinet-approved 25 March 2026, gazetted 10 April 2026 in Government Gazette No. 54477. Public comment closes 10 June 2026.
  2. The EU Artificial Intelligence Act, Regulation (EU) 2024/1689 — Articles 5, 6, 9–15, 27, 50 and Annex III.
  3. The Amnesty International Algorithmic Accountability Toolkit (December 2025).
  4. The IBM bias taxonomy — eight bias types and a three-stage mitigation model.

It is a single-page application built to WCAG 2.2 Level AAA accessibility. It runs in any modern browser, with no installation, by double-clicking the file.

The 5-minute tour

Three steps. Best for a first call where the buyer wants to know what they are looking at.

1

Pick a role and land on the Dashboard

BiasLens adapts to four user types. Every interface element re-emphasises what matters to that user. Watch.

Click: any role on the welcome screen — try Business or AI developer first.
Look at: the Dashboard. The AIA / FRIA builder is now first and labelled "Recommended for you". Click Change role in the topbar and pick Civil society, journalist or activist. The PAIA generator is now first.
2

See the dual classification in action

Every AI system in South Africa today must be classified under two regimes — the SA Draft AI Policy tier and the EU AI Act, if outputs reach EU users. BiasLens does this side by side.

Click: 1. Bias risk questionnaire in the left rail.
Click: the Use sample data (SASSA) button at the bottom. The questionnaire jumps to the auto-generated report.
Look at: the two side-by-side risk cards — SA Draft AI Policy tier and EU AI Act classification — and the EU AI Act deadlines list with Live and Pending pills.
3

Generate a real document

BiasLens does not just produce reports — it produces ready-to-send legal documents. Here is a PAIA Section 11 letter and an EU access-to-documents request, both downloadable, in under thirty seconds.

Click: 4. PAIA / FOI generator in the left rail.
Change: the Jurisdiction dropdown to "Both — parallel SA + EU evidence package".
Click: Generate letter. Two letters appear, each with a Download button.

That is the 5-minute version.

If the buyer is engaged, move into the 15-minute tour below. If not, end with: "BiasLens turns weeks of compliance work into a session. Want me to set up a longer demo with your own AI system?"

The 15-minute tour

The full path through all five functions. Use this when the buyer has booked a meeting and wants depth.

1

Function 1 — Bias risk questionnaire (3 minutes)

Click 1. Bias risk questionnaire in the left rail. Click Use sample data (SASSA) to skip ahead. Walk through the report:

  • The header recap of system, deployer, populations, oversight.
  • Two risk cards side by side — SA tier and EU classification, with the exact Annex III category named.
  • The EU AI Act deadlines: 2 February 2025 and 2 August 2025 are Live; 2 August 2026 and 2 August 2027 are Pending.
  • The IBM 8 bias types, each marked Flagged or Low signal.
  • The SA Draft AI Policy 6 pillars.
  • The two-column Triggered Obligations — South African statutes on the left, EU AI Act articles on the right.

A board can read this in under a minute. A compliance officer can act on it the same day.

2

Function 2 — Fairness metrics calculator (3 minutes)

Click 2. Fairness metrics. The bank credit-scoring sample loads automatically.

  • Demographic parity: +18 percentage points difference between men and women.
  • Equal opportunity difference: +13 points among qualified applicants.
  • Disparate impact ratio: 0.71 — fails the four-fifths rule.
  • Statistical parity difference: +18 points.

Each metric is tagged with the EU AI Act articles it speaks to — Article 10(2)(f) bias examination, Article 10(3) representativeness, Article 13 transparency, Article 15 accuracy.

Scroll down to the IBM three-stage mitigation strip and the Remediation evidence pack preview. The preview shows how the metric output would be cited inside an Article 27 FRIA — sections 8, 9, 12, 15.

In a real audit you would upload your own CSV here. The prototype shows what the output looks like.

3

Function 3 — Compliance mapper (4 minutes)

Click 3. Compliance mapper. The SAPS facial recognition pilot loads by default. This is the screen that wins procurement officers.

Cycle through all five SA systems by clicking each button — SASSA, SAPS, hospital, bank, municipal jobs. Each one re-renders the entire map.

Walk through the four columns:

  • Column 1 — System feature (e.g. real-time biometric ID, watch-list matching).
  • Column 2 — South Africa (Constitution sections, POPIA, PEPUDA, PAJA, sectoral regulators). Cells flagged in BeAccessible Mid Blue mark a compliance gap.
  • Column 3 — EU AI Act (Articles 5, 6, 9–15, 27, 50, with the Annex III category named).
  • Column 4 — International human rights (UDHR, UNCRPD, CEDAW, CERD, ICCPR, ICESCR, UNGPs, EU Charter).

Point at the cross-jurisdictional flag callout above the map: a POPIA Personal Information Impact Assessment can satisfy parts of a GDPR DPIA, which complements an EU AI Act Article 27 FRIA. One assessment, three jurisdictions.

This is what your legal counsel produces in two weeks. BiasLens drafts the first cut in two minutes. They review and refine — they do not start from a blank page.

4

Function 4 — PAIA / FOI generator (2 minutes)

Click 4. PAIA / FOI generator. The form is pre-populated with the SAPS pilot.

Switch the Jurisdiction dropdown to each of the three options and click Generate letter:

  • SA — produces a PAIA Section 11 / Form 2-style letter.
  • EU — produces an access-to-documents request under Regulation (EC) 1049/2001 plus an EU AI Act Article 27 FRIA request.
  • Both — produces both letters in parallel.

Click any Download button. A real .txt file downloads to the user's computer.

Scroll down to the Evidence Log table — three sample rows with document, source, date, status, and legal basis.

Civil society teams spend a week drafting one PAIA. Here it took thirty seconds.

5

Function 5 — AIA / FRIA builder (3 minutes)

Click 5. AIA / FRIA builder. This is the heaviest function. Show three things only:

  1. The Export format selector at the top — AIA only, FRIA only, or Unified AIA + FRIA. Same wizard, three exports.
  2. The 23-section stepper. Click section 7 — the IBM 8-bias-type assessment — and section 21 — Notification readiness with three pre-set destinations: SA AI Ethics Board, SA Information Regulator, EU national market surveillance authority.
  3. The Export document button. A complete document downloads, marked clearly with the chosen format.

A FRIA is mandatory under EU AI Act Article 27 from 2 August 2026 for any high-risk system whose outputs reach the EU. This is a two-week project for most teams. BiasLens turns it into a structured one-day workshop.

6

Show the accessibility controls (1 minute)

Click A++ in the topbar. All text scales up. Click Reading mode. The font becomes serif and lines space out — for users with cognitive load needs. Click Accessibility statement in the rail to show the full WCAG AAA conformance details.

BiasLens is the only AI bias platform built to WCAG 2.2 Level AAA from the first line of code. Accessibility is not a future feature — it is the foundation. That matters for public procurement and for any organisation serving disabled communities.

Tour for your role

Pick the audience closest to you. The tab below shows the pain points BiasLens solves, the function that matters most, and the demo path that lands the value.

Government and public-sector procurement

Who: DPSA, National Treasury, Information Regulator, Department of Communications and Digital Technologies, municipal supply chain management, SOEs, the proposed AI Ethics Board.

Pain points BiasLens solves

  • Pre-purchase due diligence on AI systems — vendors arrive with brochures, not compliance evidence.
  • Aligning a single procurement decision with the SA Draft AI Policy, POPIA, PEPUDA, PAJA, and the EU AI Act all at once.
  • Fundamental Rights Impact Assessments under EU AI Act Article 27, which become mandatory on 2 August 2026 for many SA public-sector deployers whose outputs reach EU users.
  • Defending procurement decisions against equality-clause and constitutional challenges.

Demo path (10 minutes)

  1. Choose the Government or public-sector procurement role on the welcome screen.
  2. Show the dashboard — Compliance mapper is first, AIA / FRIA builder second.
  3. Open the Compliance mapper. Cycle through all five SA systems. Stop on SAPS facial recognition. Read the cross-jurisdictional flag aloud.
  4. Open the AIA / FRIA builder. Show the Unified export option and the 23 sections.
  5. Open the PAIA generator and produce both an SA and an EU letter — useful for a vendor who has not provided evidence.

The value proposition in one sentence

"BiasLens turns the SA Draft AI Policy and the EU AI Act into a procurement checklist your team can run before signing."

Corporate compliance

Who: banks, insurers, hospitals, retailers, telcos, employers, model risk teams, internal audit, AI ethics committees, GRC functions.

Pain points BiasLens solves

  • POPIA Personal Information Impact Assessments duplicated across teams when one assessment could serve POPIA, GDPR, and an EU AI Act FRIA.
  • Model risk and fair-lending questions from regulators that compliance teams answer with spreadsheets.
  • Evidence packs for the FSCA, the Information Regulator, the Equality Court, or an EU market surveillance authority — produced under deadline pressure.
  • Documentation under EU AI Act Article 11 and Annex IV, which most SA vendors have not yet built for.

Demo path (12 minutes)

  1. Choose the Business or AI developer role.
  2. Open Function 1, click Use sample data (SASSA), walk through the dual classification report.
  3. Open Function 2 — the bank credit-scoring sample. Walk through all four fairness metrics. Show the EU AI Act article tags. Show the IBM three-stage mitigation. Show the Remediation evidence pack preview.
  4. Open the AIA / FRIA builder. Show the Unified export. Show section 12 (Data governance, Article 10) and section 21 (DPIA cross-reference, Article 27(4)).
  5. Show the Article 27 FRIA notification destinations in section 22.

The value proposition in one sentence

"One assessment satisfies POPIA, GDPR, and EU AI Act Article 27 — instead of three teams writing three documents."

Civil society and journalists

Who: Section 27, Equal Education, Right2Know, Black Sash, the Centre for Applied Legal Studies, ALT Advisory, Open Secrets, investigative journalists at Daily Maverick, amaBhungane, GroundUp, and University of Cape Town and Wits research centres.

Pain points BiasLens solves

  • PAIA requests drafted from scratch every time, often without the right legal hooks.
  • Evidence logs kept in inboxes and personal spreadsheets, lost between investigators.
  • Translating technical AI outputs into a story or court argument — and connecting it to constitutional grounds.
  • Investigating an SA AI system that may be exported to the EU — needing both PAIA and EU access-to-documents pathways.

Demo path (10 minutes)

  1. Choose the Civil society, journalist or activist role.
  2. Show the dashboard — PAIA generator is first.
  3. Open the PAIA generator. Switch jurisdiction to "Both" and produce both letters. Download one.
  4. Show the Evidence Log table — explain how a small team can share one log instead of five inboxes.
  5. Open the Compliance mapper for SAPS. Show how the IHRL column gives them ICCPR Article 17, CERD, and EU Charter Article 7 — useful for a court application or an op-ed.
  6. Open the bias risk report from Function 1 and explain how it produces a one-pager that goes into a story or affidavit.

The value proposition in one sentence

"BiasLens turns four weeks of investigator setup into a session — and gives your story or your court application the legal scaffolding regulators cannot dismiss."

Affected individuals and communities

Who: people whose grant has been suspended by SASSA, applicants rejected by an algorithm, patients whose triage was AI-assisted, communities under facial-recognition surveillance, and the disability, women, and language-rights advocates who support them.

Pain points BiasLens solves

  • Not knowing what an automated decision was based on or how to challenge it.
  • Plain-language tools — not legal jargon — to understand what an organisation owes the affected person.
  • A starting point for a complaint to the SA Human Rights Commission, the Information Regulator, or the Equality Court.
  • Accessible interfaces that work for blind users, deaf users, users with cognitive disabilities, and users whose first language is not English.

Demo path (8 minutes)

  1. Choose the Affected individual or community role.
  2. Show that the dashboard now puts the questionnaire and the PAIA generator first.
  3. Open Function 1. Walk through the first three questions to show the plain-language design.
  4. Click Reading mode in the topbar. Show how the entire interface becomes serif and easier to read.
  5. Click A++. Show how text scales without breaking layout.
  6. Open the PAIA generator and produce a request directly addressed to the public body involved.

The value proposition in one sentence

"BiasLens gives an affected person the same legal toolkit a top law firm has — accessible from a phone, in plain language."

The five sample SA AI systems

Every screen uses these five real-world South African AI deployments. They were chosen because they reflect the populations most often harmed by algorithmic systems in the SA context — people with disabilities, women, Black South Africans, isiXhosa- and isiZulu-speaking communities, rural residents, and migrants.

SASSA grant fraud-detection algorithm

Welfare. Flags potentially fraudulent grant applications and ongoing payments. Affects Disability Grant and Old Age Grant beneficiaries, rural residents, and isiXhosa- and isiZulu-speaking applicants.

SAPS facial recognition pilot

Policing. Identifies suspects in CBD CCTV feeds. Documented disparate error rates for darker skin tones and women.

Private hospital AI triage tool

Healthcare. Prioritises emergency-department arrivals. Affects people with chronic illnesses, pregnant patients, and non-English-speaking patients.

Bank credit-scoring model

Financial services. Assesses creditworthiness for unsecured personal loans. Affects women, informal-sector workers, township residents, and migrants.

Municipal job-application screening tool

Employment. Ranks applicants for shortlisting in municipal vacancies. Affects people with disabilities, women, isiXhosa-speaking applicants, and older job-seekers.

What this prototype does and does not do

It does

  • Run end-to-end on a laptop or phone, offline, with no installation.
  • Demonstrate every screen, every flow, with realistic SA sample data.
  • Generate downloadable PAIA letters, EU access-to-documents requests, bias risk reports, and AIA / FRIA documents as plain-text files.
  • Apply WCAG 2.2 Level AAA to every interaction — text scaling, reading mode, keyboard navigation, screen-reader landmarks.
  • Show what the production version will feel like, navigate like, and produce.

It does not yet

  • Save your work between sessions — close the tab, the data is gone (production will save to a POPIA-aligned database).
  • Compute fairness metrics from a real CSV — the numbers shown are illustrative (production runs the calculation server-side).
  • Translate the interface into isiXhosa, isiZulu, Afrikaans, Sesotho, or Setswana — the strings are scaffolded, the translations are pending.
  • Render PAIA letters as the official Form 2 PDF — production will.
  • Connect to the SA AI Ethics Board, the Information Regulator, or any EU national market surveillance authority — production will integrate via APIs.

Cheat sheet — keyboard-only demo

If you have to demo without a mouse — for example, to a screen-reader user — every action above is reachable by Tab, Shift+Tab, Enter, Space, and the arrow keys.

Tab
Move forward through interactive elements.
Shift + Tab
Move backward.
Enter or Space
Activate a button, link, or radio option.
Tab onto Skip link
The first focusable element on the page is "Skip to main content".
Arrow keys
Move within a radio group (questionnaire and AIA wizard).
Esc
Closes the mobile rail when open.

After the demo

If the buyer is engaged, offer one of three next steps.

  1. Workshop — a half-day session on one of their AI systems, ending with a draft AIA / FRIA they can take to legal counsel.
  2. Pilot — a 30-day pilot with three of their AI systems mapped end-to-end, including custom sample data.
  3. Implementation — a full BiasLens deployment with their branding, their integrations, their compliance team trained.

Contact: hello@beaccessible.co.za

Accessibility statement

Our commitment

BiasLens is designed and built by BeAccessible to meet the highest accessibility standard available — Web Content Accessibility Guidelines (WCAG) 2.0, 2.1 and 2.2 at Level AAA, and the seven Universal Design principles. We treat accessibility as a precondition, not a feature.

What we have done

  • Semantic HTML5 throughout — header, nav, main, section, footer, fieldset, legend.
  • Visible focus indicators — 3 px Deep Blue ring with 2 px offset on every interactive element.
  • Touch targets at least 48 × 48 px.
  • Colour contrast ≥ 7:1 for all text against its background. Status is communicated by text + icon + colour, never by colour alone.
  • Skip-to-main-content link as the first focusable element.
  • ARIA landmarks, live region for dynamic feedback, aria-current and aria-pressed on stateful controls.
  • All form fields have associated <label> elements; placeholders are not used as labels.
  • prefers-reduced-motion respected — non-essential animations are disabled.
  • A persistent text-resize control (A / A+ / A++) and a reading-mode toggle (serif font, 1.8 line-height).
  • Plain-language drafting — Grade 8 reading level target on all user-facing text.
  • Multilingual readiness — all visible strings live in a single STRINGS object ready for translation into isiXhosa, isiZulu, Afrikaans, Sesotho and Setswana.

Universal Design principles applied

  1. Equitable use — every function works for every user role.
  2. Flexibility in use — keyboard, mouse and touch all supported equally.
  3. Simple and intuitive use — plain language, predictable layout.
  4. Perceptible information — text + icon + colour, captions, alt text.
  5. Tolerance for error — back navigation, save draft, change role.
  6. Low physical effort — large touch targets, minimal typing, sample-data buttons.
  7. Size and space for approach and use — responsive at 320 px and up.

Known limitations

This is a working prototype. Persistent storage is in-session only; data is not saved to a server. Multilingual translations are scaffolded but not yet populated. CSV upload is acknowledged but the demo uses sample numbers — production will compute against the uploaded data. PAIA letter formatting follows Form 2 conventions but is not the official Form 2 PDF.

Contact

Accessibility feedback or barrier reports: hello@beaccessible.co.za. We aim to respond within five working days.