Research · Accessible AI

Making AI work for more people.

Accessible AI research asks what keeps people from using artificial intelligence, from cost and connectivity to language, literacy and disability, and how to design systems that remove those barriers.

Reaching people where they are.

Focus
Removing barriers to AI
Conditions
Everyday devices, unreliable networks
Products
Temi AI
Why it matters

Most people are not left out of AI by choice. They are left out by design.

Many AI systems assume fast internet, the latest phone, fluent English and the confidence to type the perfect prompt. Across Africa, those assumptions leave many people out.

We study those barriers and design around them, then test what we learn in Temi AI, where real use shows whether a system is truly accessible.

Research areas

Where the work happens.

Accessible AI is made up of 3 research areas. Each has its own questions, methods and links to our products.

ACC-01

Localized Reasoning & Bandwidth Constraints

Running useful AI on local devices and keeping it working over intermittent mobile networks.

Most AI systems assume a fast, constant connection. We explore on-device inference, local caching of state and knowledge, compressed indexes and careful synchronisation, so that a system keeps helping when the signal drops and catches up when it returns.

  • On-device inference
  • Offline-first design
  • Resilient synchronisation
ACC-02

Efficient & Affordable Models

Making models small and efficient enough that cost, data charges and older phones stop deciding who can use AI.

Every request costs someone money, in data, in device storage or in compute. We study quantisation, distillation and smaller task-focused models, and measure the real cost of each interaction, so that useful AI fits everyday budgets and devices.

  • Quantisation and distillation
  • Cost per interaction
  • Older and low-memory devices
ACC-03

Inclusive Design

Designing with people with disabilities, low literacy or little experience of technology, rather than for them.

An AI product is only accessible if people can actually use it. We work on plain-language responses, support for screen readers and other assistive technology, clear ways to correct the system, and testing with the people most likely to be left out.

  • Plain language
  • Assistive technology support
  • Testing with underserved users
Our approach

Start with the barrier, not the model.

Accessibility is easy to claim and hard to prove. We begin with the people who are left out and only call something accessible when they can use it.

  1. 01

    Find the barrier

    Identify who is left out and why: cost, connection, language, ability or confidence.

  2. 02

    Design around it

    Shape the system to the conditions people actually live with.

  3. 03

    Test with people

    Try it with the people facing that barrier, on their own devices and networks.

  4. 04

    Measure and refine

    Track whether they can complete real tasks, and keep improving where they cannot.

The questions we are asking

What we are exploring.

01

What does it take to work offline?

Designing systems that keep working when connectivity drops, and catch up when it returns.

02

What does it take for AI to be affordable?

Lowering what it costs to run and use AI, from data charges to devices, so price does not decide who benefits.

03

How small can useful AI be?

Making models efficient enough for affordable phones and limited data plans.

04

Can AI work for everyone?

Designing with people with disabilities, low literacy or little experience of technology, and checking honestly where our products fall short.

Where it connects

Related work.