Industries / Government, education and community / Nonprofits
Volunteers
Charities recruit, train, match and keep unpaid helpers, who do much of the work in many organizations.
7
Milestones achieved, from 1884 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1884
Volunteers who moved in
Samuel and Henrietta Barnett opened Toynbee Hall in London's East End as a place where future leaders would live and work as volunteers, face to face with poverty. Its residents included Clement Attlee and William Beveridge, and Toynbee Hall calls itself the world's first university settlement. Jane Addams visited it before opening Hull-House in Chicago.
Credit
Samuel and Henrietta Barnett
1958
Volunteers sent abroad
Alec and Mora Dickson recruited 16 British volunteers and sent them overseas, answering a letter from the Bishop of Portsmouth asking for people to teach English in Borneo. The effort became Voluntary Service Overseas (VSO), which says it has since worked in more than 90 countries with more than 80,000 volunteers.
Credit
Alec and Mora Dickson
1995
Openings posted online
Idealist began in the summer of 1995 on one old computer, with no full-time staff, to list ways to do good. In early 2025 it merged with VolunteerMatch, which it calls the largest online volunteering platform; Idealist says the combined site now connects 15 million people a year with more than 200,000 organizations in 180 countries.
Credit
Idealist; VolunteerMatch
2015
Sight by video call
Be My Eyes connects blind and low-vision people by live video to sighted volunteers who describe what the phone's camera shows. The idea came from Hans Jørgen Wiberg, a Danish craftsman with low vision. The app launched in January 2015, 10,000 volunteers signed up within 24 hours, and 1 million had signed up by May 2018, by the company's count.
Credit
Hans Jørgen Wiberg; Be My Eyes
2021
Practice with an AI caller
The Trevor Project, a US crisis service for LGBTQ young people, began training volunteer counselors with an AI, built on OpenAI's GPT-2, that plays "Riley," a teenager in distress, a part staff used to play by hand. It was trained on thousands of past role-play transcripts and lets trainees practice at night and on weekends, when nearly 70% of the charity's counselors volunteer. The AI never talks to real young people, and people review the practice transcripts. Google.org gave the charity $2.7 million and nearly 30 staff fellows over two years for this and other AI work.
Credit
The Trevor Project; Google.org
2023
AI takes the first look
In March 2023 Be My Eyes added Be My AI, which describes a user's photo in words using OpenAI's GPT-4, with human volunteers still a call away. In a study of 14 visually impaired users, Jingyi Xie and colleagues found they used it for household, spatial and social tasks, while noting it was not yet built around users' goals.
Credit
Be My Eyes; OpenAI
2024
Fewer regular volunteers
In England, 16% of adults volunteered through a group at least once a month in 2023/24, down from 27% in 2013/14 and the lowest rate the government's Community Life Survey has recorded with its current method. In the US, the share of people aged 16 and over who volunteered through an organization fell to 23.2% in 2021, then rose to 28.3% in the year to September 2023, still 1.7 points below the pre-pandemic level, by Census Bureau and AmeriCorps figures.
LessonCounting volunteers is not enough; in the US their number came back after 2021, but the hours each one gave kept falling.
Credit
UK Community Life Survey; US Census Bureau and AmeriCorps
Watch for
Feedback makes practice count
Researchers built an AI that plays a client and gives feedback, and split 94 novice counselors by lottery into practice alone or practice with feedback. Those who got feedback improved in listening skills such as reflections and questions; those who only practiced did not, and their empathy declined.
Achieved whenAn outside team repeats the result, or a charity uses the approach in routine volunteer training and an outside evaluation finds better performance.
Claimed by
Ryan Louie, Emma Brunskill, Diyi Yang and colleagues, May 2025
Calls that train AI
Be My Eyes says it will give Meta video collected through its platform, which it describes as the lived experience of blind and low-vision users and the volunteers who help them, to train AI models, including those behind Meta's AI glasses. It says personal information is removed and users can opt out.
Achieved whenMeta or an outside group publishes a measured gain on blind users' real tasks from models trained on this data.
Claimed by
Be My Eyes, February 2026
Open
01
Audit stale volunteer listings
A count of how many volunteer openings posted on large online boards are still real. Sample listings in several cities and check each against the organization's own website. For a subset, send an honest email that asks whether the role is open, and record how long a reply takes. Work within each board's terms of use.
The briefCredit · unclaimed02
Test AI role-play training
An AI trainer for a non-crisis volunteer role, such as food-pantry intake or a tenant information line. It plays fictional people drawn from the charity's own training scenarios and gives feedback after each session. New volunteers are split by lottery: half practice with the AI and half with the usual staff role-play. Trainers who do not know which group a volunteer was in score a final practice session.
The briefCredit · unclaimed
Open 01 · the brief
Audit stale volunteer listings
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A count of how many volunteer openings posted on large online boards are still real. Sample listings in several cities and check each against the organization's own website. For a subset, send an honest email that asks whether the role is open, and record how long a reply takes. Work within each board's terms of use.
Why it is still open
Advice to volunteer managers has long been to check online postings monthly and delete filled ones, but no published audit of how many listings are out of date was found.
Built with what exists
A web agent such as Browser Use (see the Note "Deleting 80% to start over"), or careful hand sampling; Python; a spreadsheet; an email account for the inquiries.
First step
Take 100 listings for one city and check each by hand against the organization's own site, to set the categories.
Done looks like
1,000 listings in five cities checked, with the share still open, already filled, removed, or unanswered after two weeks, each with a 95% confidence interval; method and data published, and the check repeated after three months.
Open 02 · the brief
Test AI role-play training
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An AI trainer for a non-crisis volunteer role, such as food-pantry intake or a tenant information line. It plays fictional people drawn from the charity's own training scenarios and gives feedback after each session. New volunteers are split by lottery: half practice with the AI and half with the usual staff role-play. Trainers who do not know which group a volunteer was in score a final practice session.
Why it is still open
The Trevor Project has trained volunteers with an AI role-play since 2021, but the published account reports no controlled comparison with staff role-play. A 2025 randomized study compared AI practice with and without feedback, not AI against people, and only for counselors.
Built with what exists
An AI model such as Claude or Gemini given written persona sheets; the feedback design from Louie and colleagues' study; the charity's existing scoring sheet; a spreadsheet for the lottery.
First step
Write three fictional personas from one charity's training scenarios and run ten practice sessions with current volunteers to check that they feel realistic.
Done looks like
At least 60 new volunteers assigned by lottery; final scores from two trainers who do not know the groups; the difference between groups with a confidence interval, plus staff hours spent per trainee; results published whether or not the AI group does better.