alfaPlus lab

Independent · Nonprofit · Published in the open

The benefits of AI arrive unevenly. We work at the far end.

alfaPlus lab is an independent nonprofit lab. We study artificial intelligence, and we look after the people it reaches last.

This chart is the whole argument

Technology never arrives everywhere at once. The width of the blocks below is not a design decision. It is the data.

New technology reaches the places that already have the conditions for it — a connection, a device, someone who knows how — and seeps outward from there. «Seeps» sounds harmless. Historically it has meant a generation.

Diffusion carries both a temporal and a resource gradient. Marginal gains accrue first to groups already holding the infrastructure, the corpora and the human capital; everyone else is displaced to the tail of the curve, a delay historically measured in decades.

LanguageWeb textSpeakers
English49.2%18.8%
Chinese5.1%14.8%
Spanish4.5%7.0%
Arabic1.1%4.6%
Hindi0.6%7.5%
Swahili0.1%2.0%
All others39.4%45.3%
Fig. 1 · Block width is the data. The two narrowest bands — Hindi and Swahili — are under one pixel wide at true scale and are drawn at 2px; that distortion is itself the finding. Switch to «by speakers» and those same languages take up a third of the screen. 〔source to be added〕

Four commitments

Written down so they can be held against us. The time we broke one is in §06.

Access before frontier

We don't compete on whose model is stronger. We watch what happens to the same model when the connection is bad, the phone is old, the language is a dialect and nobody is there to teach it — then we go and get back what was lost.

Our objective is not advancing the capability frontier but the accessibility decay of existing capability under constraint: low bandwidth, low-spec hardware, non-dominant languages, low prior knowledge. Reducing that decay is the objective function.

Reproducible, and readable by non-researchers

Everything we publish ships with its data, its scripts and its failures — and with a version written without jargon. A finding only a peer can read does not exist for the people we serve. This page included; the switch is top right.

All releases carry data, scripts and failure samples, plus a parallel text free of terminology. A conclusion verifiable only by peers conveys no information to the served population. This page implements the same rule; both registers ship with it.

Only promises that can be checked

We don't say «solving educational inequality with AI». It cannot be verified, so it cannot be held against us. We say: save one teacher four hours of marking a week — and then record honestly whether we did.

We do not make unfalsifiable claims. Project goals are stated as observable measures, with falsification conditions declared at the outset. Unmet goals go to §06 and are not restated.

The people served can stop the work

Every project has a co-designer from the field, and they can call a halt. If they do, we stop — and the reason goes in §06 and stays there.

Each project appoints a field co-designer holding unconditional termination rights. Termination decisions and their reasons are published in §06 and are not removed.

Three lines, feeding each other

Research produces judgement, engineering turns judgement into something usable, and the field sends the problems back.

A

Access & Equity

Make mainstream AI work where the conditions are poor: offline, on an old phone, in a dialect, without sight. Adaptation and training for schools and nonprofits are free of charge.

Adaptation against four classes of constraint — bandwidth, compute, language and sensory channel: on-device inference, quantisation for low-spec hardware, dialect and minority-language support, screen-reader and alternative input paths.

Access & Equity

B

Public-interest Engineering

Nonprofits can't keep engineers on staff, so the same work gets rebuilt over and over: sorting records, matching cases, checking claims, writing reports. We build it once and hand over the code, the docs and how to run it.

We carry the engineering nonprofits cannot build in-house. Deliverables include source, documentation and an operations manual, follow data minimisation, and create no runtime dependency on us.

Public-interest Engineering

C

Open Research

Research on who can't use it: how much worse it gets in smaller languages, how far a model that fits on a phone can go, how it fails when someone is trying to claim a benefit, and who is harmed when it does. The results that went badly get published too.

Empirical work on accessibility: capability decay in low-resource languages, the ceiling of on-device models, failure modes and harm assessment in public-service settings. Methods, data and negative results are published together.

Open Research

Published record

Released under CC BY 4.0 or Apache-2.0. Retracted items stay on the list.

A-07 On-device models in low-bandwidth classrooms: fourteen weeks in the field field report 2026-06 · six schools · includes two rollbacks and why · data and scripts attached A-06 Dialect question-answering benchmark v0.3 dataset 2026-04 · seven Chinese dialects · open to community submissions A-05 When the model can't read the policy: failure modes in frontline benefits advice research 2026-02 · includes a list of what we advise against automating A-04 An AI handbook for non-technical colleagues (2nd edition) handbook 2025-11 · no jargon · free to reprint and rewrite A-03 Survey of AI use in county schools retracted 2025-09 · retracted 2025-10-18 · sampling bias, conclusions do not hold; the original stays downloadable

Numbers you can check

As of 2026-06-30, on the same basis as the annual report. Find a discrepancy and write to us; the correction goes in §06.

Partner organisations
Organisations with actual delivery that year; 31 of them at county level or below.
People trained
Repeat attendance counted per session; includes 12 sessions for blind and low-vision users.
Open repositories
Four are archived; archived work stays downloadable.
Admin costs
Denominator is total spend that year; itemised in the annual financial report.
Directed funding
Funding conditioned on conclusions is returned; every source is published.

Corrections and retractions

Corrections don't overwrite the original. Retractions don't remove the entry. The point of this section is that it isn't flattering.

2026-08-26 · rev. 4
Added «directed funding» to §05 and completed the basis notes.
2026-07-02 · halted
«Frontline policy assistant» halted at the request of our field partners. Reason: errors concentrated in welfare and medical-insurance clauses; the harm was not acceptable. The project will not restart.
2026-04-11 · rev. 3
Corrected the claim on this page. It read «we will eliminatenarrow inequality in educational resources»; it now reads «shorten the time that inequality takes to close». Reason: the evidence does not support «eliminate». Changed after field partners challenged it.
2025-10-18 · rev. 2
Retracted A-03, «Survey of AI use in county schools». Reason: sampling bias; the conclusions do not hold. The retracted item stays in §04 and remains downloadable.
2023-06-01 · rev. 1
First published.

If you think someone should be doing this

We are small and we don't plan to get big. We need people who stay more than we need to grow.

Partner with us Schools, nonprofits and public bodies: adaptation and training are free. Write and describe the setting and its constraints. Volunteer Engineering, research, translation, accessibility testing, time in the field. A few hours a week is enough; staying is what counts. Fund the work Undirected donations only, published line by line. Funding with conditions attached to conclusions is returned. Just take it Every tool and report is free to use. You don't need to ask us, and you don't need to credit us.

hello@alfaplus.org