Where AI can widen the circle: tutors, translators, and a second pair of eyes

A useful question for AI is what help becomes easier to reach. A student might get another opportunity to practise. A farmer might ask a question in a familiar language. A blind person might get a description of an image without waiting for someone nearby.

These possibilities deserve attention. They also require different kinds of evidence. Better lesson results, a growing user base, and a higher translation benchmark score tell us different things about whether a tool improves someone's life.

A six-week tutoring pilot in Nigeria

During June and July 2024, a randomized pilot in nine schools in Benin City offered secondary students teacher-guided, after-school sessions with Microsoft Copilot. The reported gain was 0.31 standard deviations on a combined assessment of English, AI knowledge, and digital skills. English alone improved by about 0.24 standard deviations.

That is encouraging evidence for the program as a whole. It does not isolate the chatbot's contribution from extra study time, teacher guidance, and structured activities. The short pilot also had substantial assessment attrition, which the authors examined through sensitivity analyses. Longer-term learning and results elsewhere remain questions.

Research references: [1]

Help for the tutor

Tutor CoPilot suggests responses during human-led online maths tutoring. The trial's analyzed sample included 783 tutors and 1,013 students in grades 3–6. Access increased lesson exit-ticket passing by four percentage points, from about 62% to 66%. The gains were nine points for the lower-rated tutor group and seven for the less-experienced group. Rating and experience were separate measures.

The estimated $20 per tutor per year covered annualized API usage, not the whole tutoring service. The tool was used in about 29% of treatment sessions. Tutors reported grade-level mismatches. End-of-year maths gains were not statistically significant; mixed student exposure limited that analysis.

Research references: [2]

Agricultural advice people can use

Farmer.Chat accepts questions through text, voice, or images and supports farmers and extension workers. Its developers' 2024 preprint reported more than 15,000 users across four countries. That population included several kinds of agricultural workers.

In a July–August 2025 survey conducted by 60 Decibels with Digital Green, 70% of 450 Kenyan respondents said they had applied its advice in the previous 30 days. The survey had a 19% response rate. This is evidence of self-reported use among respondents, not a demonstrated causal increase in harvests or income.

Research references: [3] [4]

Translation beyond the best-served languages

Meta's No Language Left Behind work, published in Nature in June 2024, describes translation across 200 languages. It reported an average 44% improvement over prior systems on the specified BLEU benchmark comparison. This is a relative benchmark improvement, not a claim that translations are 44% more accurate in every setting. The paper made research resources available for non-commercial use.

The model card is explicit about limits: NLLB-200 is a research model, and medical, legal, and certified translations are outside its intended use. Language and domain coverage require checking before anyone relies on the output.

Research references: [5] [6]

A second pair of eyes

Be My Eyes combines volunteer video assistance with Be My AI image descriptions. People can use the AI without a volunteer and can choose human assistance when needed.

In a January 22, 2025 announcement, the company reported a community of 750,000 users and 97% satisfaction with AI descriptions. These are dated company reports, not independently established measures of accuracy or improved independence.

Its safety terms warn against relying on Be My AI for medication dosages, emergencies, or as a replacement for mobility aids. A convenient description still needs an appropriate level of caution.

Research references: [7] [8] [9]

The practical test

The strongest lesson here is to be specific about the job a tool is doing and how success is measured. The education trials tested particular programs. Agricultural surveys describe respondents' experiences. Translation benchmarks test language performance. Accessibility services show useful functions and demand, while leaving important outcome questions open.

Human involvement also takes different forms. Tutor CoPilot supports a person making instructional choices. Be My Eyes offers optional human help. A translation model does not guarantee a human reviewer. These examples do not establish that putting a person somewhere in a system automatically makes it effective or safe.

For builders, the questions are practical: Who can use the service? What happens when it gives a wrong answer? Who can help? What evidence would show a lasting benefit?

Devices, affordability, connectivity, accessible interfaces, and relevant language support can still limit access. Treat those as design requirements. Measure who is being missed as carefully as who is using the tool.

Evidence over hype, every issue. Next for Humanity explores the technologies reshaping human life and what thoughtful builders can do next.