Should AI Help Run a Mutual Aid Network? Coordination Without Bureaucracy
Mutual aid networks run on group chats, shared spreadsheets, intake forms, and payment apps, held together by volunteers who are often exhausted. AI tools promise to lighten that load by sorting requests, translating messages, cleaning up duplicate rows, and drafting the weekly update. But mutual aid is built on values that most technology was not designed around: solidarity instead of charity, no eligibility tests, no hierarchy, and deep care for the safety of neighbors who may be undocumented, unhoused, or targeted. This piece weighs where AI genuinely helps a horizontal network coordinate, where it quietly reintroduces the bureaucracy and surveillance mutual aid exists to avoid, and how a network can make that decision together rather than letting one tech-savvy volunteer make it for everyone.

Picture the coordination side of a neighborhood mutual aid network on a Tuesday night. Forty new rows have appeared in the request spreadsheet since Sunday. Some are in Spanish, a few in Haitian Creole, and one is a voicemail someone transcribed by hand. Three people asked for the same grocery delivery twice because they were not sure the first form went through. The Signal group has 200 unread messages, half of them offers to help and half of them logistics for a rent fund drive. One volunteer, who also works full time, is trying to match drivers to deliveries before anyone goes to bed. This is the moment when someone in the group suggests, reasonably, that "maybe AI could help with this."
That suggestion deserves a serious answer, not a reflexive yes or no. Mutual aid surged across the United States during the COVID-19 pandemic, when hundreds of local mutual aid groups sprang up to deliver groceries, medicine, and cash to neighbors. Many of those networks still exist, and new ones form after every disaster, layoff wave, or policy shock. Most are unincorporated and volunteer-run, and many are deliberately structured without executive directors, boards, or formal eligibility rules. The coordination burden is real, and burnout among the people who carry it is one of the most common reasons networks fade.
At the same time, mutual aid is not just a lightweight version of a social service agency. As legal scholar and organizer Dean Spade has argued, charity tends to sort people into the deserving and undeserving and attach strings to help, while mutual aid tries to meet needs directly and build solidarity among the people involved. In an interview with The Nation, Spade describes mutual aid as redistribution that does not depend on eligibility rules that can be withdrawn. Many AI tools, by contrast, are built to score, rank, flag, and predict. Bringing them into a mutual aid network without care can quietly rebuild the very gatekeeping the network was created to reject.
This article is written for the organizers, coordinators, and volunteers who keep mutual aid networks running, and for nonprofits and fiscal sponsors that support them. It covers the actual coordination tools most networks use, the specific tasks where AI can reduce load without changing who decides, the uses that cut against mutual aid values, how to protect members' data (especially for undocumented neighbors), and how to make a tool decision collectively. If your organization is a more conventional volunteer-run group, our guide to AI for all-volunteer nonprofits may be a better fit. This piece stays focused on what makes mutual aid different: its horizontal structure and its explicit values.
The Real Coordination Stack of a Mutual Aid Network
Before asking where AI fits, it helps to be honest about what mutual aid coordination actually looks like. Very few networks run on purpose-built software. Most assemble a stack of free or cheap consumer tools that volunteers already know how to use. That stack is flexible and fast to set up, which is exactly why it spread so quickly in 2020. It is also fragile, scattered across personal accounts, and often holds far more sensitive information than anyone intended when the first form went up.
Guides like the American Friends Service Committee's guide to creating a mutual aid network describe a familiar pattern: a way for people to ask for help, a way for people to offer it, a way to match the two, and a way to communicate. Each of those functions usually maps to a specific tool, and each tool carries its own privacy and power implications.
Request and offer intake
Google Forms, Airtable forms, hotlines, paper sheets
Neighbors fill out a request form (groceries, rides, rent help, childcare) or an offer form (I have a car, I speak Vietnamese, I can cook on Thursdays). Phone hotlines and in-person tables catch people who are not online. Forms tend to accumulate questions over time, and every new question is new data held about a neighbor.
Matching and tracking
Shared spreadsheets, Airtable bases, Trello boards
Form responses land in a spreadsheet where coordinators assign a volunteer, track status, and close out requests. These sheets are the network's memory, and they are frequently shared with dozens of people through a single link. Duplicates, stale rows, and inconsistent formatting pile up quickly.
Group communication
Signal, WhatsApp, Slack, Discord, text threads
Group chats are where the real coordination happens: "Can anyone drive to Elm Street at 4?" Networks concerned about safety often prefer Signal for its end-to-end encryption. Slack and Discord are easier to organize into channels but store message history on company servers, which matters for anyone who could be targeted.
Moving money
Venmo, Cash App, Zelle, GoFundMe, fiscal sponsors
Many networks redistribute cash directly through payment apps, often through a volunteer's personal account. Larger networks sometimes move funds through a fiscal sponsor so donors can get tax deductions and the money is held by an organization rather than an individual.
Any AI tool you add will plug into this stack, usually by reading the spreadsheet, the form responses, or the chat. That is the crucial point. AI does not arrive as a neutral helper floating above your data. It arrives as another place where your neighbors' names, addresses, immigration concerns, health needs, and financial situations could be copied, processed, and stored. The question "should we use AI?" is really the question "should we send this information to another company, and what will we let that company's software decide?"
Why Mutual Aid Values Make This Harder Than It Looks
Most nonprofit AI advice assumes an organization with a mission, a hierarchy, and a defined client population. A director decides which tools to adopt, staff follow a policy, and clients consent through an intake process. Mutual aid networks often reject each of those assumptions on purpose. There may be no director to decide. The line between "helper" and "helped" is intentionally blurry, because the person who received groceries last month may be driving deliveries this month. And the network may avoid formal intake precisely because formal intake feels like a welfare office.
Those choices are not inefficiencies to be optimized away. They are the point. When a network decides not to ask for proof of income, it is making a political statement that need is enough. When it avoids a single coordinator with final authority, it is protecting itself from capture and burnout. Any technology that subtly reverses those decisions, for example by ranking requests by "urgency score" or by concentrating information in one admin account, changes the character of the network even if nobody votes on it.
No deserving test
AI that scores, flags, or prioritizes requests can recreate eligibility screening by another name. A model trained on general text has its own assumptions about who is "really" in need.
Horizontal power
The person who sets up and controls the AI workflow can become a de facto manager. Tools that only one volunteer understands concentrate power quietly.
Safety first
Members may be undocumented, unhoused, surviving abuse, or organizing against powerful interests. Data that is harmless for a book club can be dangerous here.
None of this means AI is off-limits. It means the test for any AI use in mutual aid is different from the test in a typical nonprofit. The question is not only "does this save time?" but also "does this change who decides, who knows, and who is visible?" A tool that saves two hours a week but makes the network more legible to outside surveillance, or more hierarchical inside, may not be worth it. A tool that saves two hours a week while keeping data local and decisions with people may be a genuine gift to tired coordinators. The broader principles in our article on community-centered AI apply here with extra force.
Where AI Can Genuinely Lighten the Coordination Load
The best uses of AI in mutual aid share three traits. They handle tedious, mechanical work rather than judgment calls. They can be done with little or no identifying information about neighbors. And their output is reviewed by a person before anything happens. When a use meets all three, it can free up real hours for the relational work that actually holds a network together: phone calls, check-ins, showing up.
Summarizing and sorting requests by type, not by worth
Long free-text requests can be summarized into a short line ("groceries for family of four, needs delivery, prefers weekends") and tagged by category so the right sub-team sees them. This is triage in the logistical sense: routing a ride request to drivers and a rent request to the fund team. It is not triage in the sense of deciding who goes first or who qualifies. Keep that distinction explicit in your instructions to the tool and in your group's agreements.
To reduce exposure, strip names, phone numbers, and addresses before sending text to a cloud AI tool. A coordinator can paste only the "what do you need" column, get back categories and summaries, and match them to row numbers locally.
Translation for multilingual networks
Language access is one of the strongest cases for AI in mutual aid. A volunteer who only speaks English can understand a request written in Spanish or Tagalog, and the network can publish its offer form and updates in several languages without waiting for a volunteer translator. For short logistical messages, machine translation is usually good enough to act on.
For anything that matters more, such as know-your-rights information, explanations of how the fund works, or messages about safety, have a fluent community member review the translation. Our guide to translation quality review covers a lightweight review process that works for volunteer groups.
Deduplicating and cleaning spreadsheets
Duplicate submissions, inconsistent phone formats, and stale "in progress" rows are a constant drain. AI features built into spreadsheet tools, or a chatbot asked to write a formula, can help find likely duplicates, standardize columns, and flag rows that have not been touched in two weeks. Asking AI to write the formula, rather than uploading the data, keeps neighbors' information out of the chat entirely.
The same approach applies to archiving: a formula or script that moves closed requests to a separate sheet, and then deletes them after an agreed period, is a practical privacy win. Our article on AI in spreadsheets walks through these techniques.
Drafting volunteer updates and how-to docs
Weekly volunteer updates, reminders about the delivery protocol, "how to use the request sheet" guides, and onboarding explainers are perfect candidates for AI drafting. They contain no private data, they get written over and over, and a quick human edit catches anything off. A clear one-page guide written in plain language also spreads knowledge, which is a horizontal value in its own right: when everyone can follow the process, no single coordinator becomes indispensable.
AI can also help turn a messy chat thread into meeting notes or a decision log, as long as the thread does not contain sensitive personal details and the group has agreed that notes may be summarized this way.
Notice what these uses have in common. None of them asks AI to decide who gets help, how much, or in what order. None requires sending a full list of neighbors to a tech company. And each produces something a person reads before acting. That is the pattern to look for in any new idea someone brings to the group.
Where AI Cuts Against Mutual Aid Values
Some uses look efficient on paper but undermine what makes mutual aid different. These are worth naming plainly, because they are often pitched by well-meaning volunteers who have seen similar tools work in nonprofit or corporate settings. The problem is rarely the technology itself. It is that the technology carries assumptions from those settings into a space that was built to reject them.
Uses to avoid or approach with extreme caution
- Scoring or ranking need. Asking AI to prioritize requests by "urgency" or "legitimacy" rebuilds the deserving and undeserving distinction. If the fund cannot meet every request, the group should decide openly how to share scarce resources, for example by lottery, first come first served, or equal splits.
- Fraud detection on neighbors. Automated flagging of "suspicious" requests treats the people you serve as suspects. Many networks accept some duplication as the cost of not policing people. If abuse becomes a real problem, deal with it through human conversation and group agreements.
- AI bots in the group chat. A bot that reads every message in your Signal or WhatsApp group sends that entire conversation to a third party, undoing the protection that drew you to an encrypted app in the first place.
- Automated replies to people asking for help. A chatbot answering requests can feel like a call center. Mutual aid is relational, and the human reply is often the help.
- Profiling and "insights" about members. Analyzing who requests most often, from which blocks, or with which needs creates exactly the kind of map that could hurt people if it leaked or was demanded by authorities.
The scarcity case deserves particular attention, because it is where the pressure to rank is strongest. When a rent fund has $3,000 and $15,000 in requests, someone has to decide. It is tempting to let a tool "objectively" sort requests so no volunteer has to carry that weight. But an AI ranking is not objective. It reflects whatever patterns the model absorbed and whatever words the requester happened to use, which tends to favor people who write fluently in English and know how to describe hardship in institutional language. A transparent rule the group agrees on, even a random lottery, is more consistent with mutual aid values and easier to explain to the people affected.
Surveillance concerns are not abstract either. Every major cloud provider receives legal demands for user data from US law enforcement, and companies like Google publish counts of those demands in their transparency reports. AI chat logs are subject to the same legal processes. In 2025, as part of copyright litigation brought by The New York Times, a federal magistrate judge ordered OpenAI to preserve consumer ChatGPT conversations, including chats users had deleted. That case had nothing to do with mutual aid, but it shows that "I deleted the chat" is not a reliable privacy guarantee. If you paste a list of neighbors into a consumer chatbot, assume it may persist somewhere you do not control. Our article on AI surveillance concerns for nonprofits explores this risk in more depth.
Data Minimization: Collect Less, Keep Less, Share Less
The single most protective step a mutual aid network can take has nothing to do with AI. It is collecting less information in the first place. Data you never gathered cannot be leaked, subpoenaed, or fed into a tool by an enthusiastic volunteer. Before discussing any AI workflow, it is worth auditing your forms and spreadsheets with a simple question for every field: do we need this to deliver help?
A grocery delivery requires an address and a way to reach someone. It does not require a full legal name, a date of birth, household income, or immigration status. A rent payment requires a way to send money. It does not require a reason, a sob story, or documentation. Many networks discover that their forms grew questions over time ("let's ask about household size for our report") that serve curiosity or funders rather than neighbors. The principles in our guide to the minimum viable dataset translate directly to mutual aid.
Practical minimization steps
- Accept first names or nicknames instead of legal names
- Never ask about immigration status, criminal records, or benefits
- Use a cross street or drop-off point when a full address is not needed
- Delete closed requests on a set schedule, such as 30 or 60 days
- Limit spreadsheet access to the volunteers who actively match requests
Rules for any AI use
- Remove names, phone numbers, and addresses before pasting text into an AI tool
- Ask AI to write formulas or templates instead of handing it the data
- Turn off training on your data and chat history where the tool allows
- Never connect AI tools directly to the request sheet or group chat
- Consider local AI models for anything involving neighbor information
Local AI deserves a mention for networks with a technically inclined volunteer. Smaller open models can now run on an ordinary laptop, which means summarizing or translating text without sending it to any company. They are less capable than the big commercial models, but for categorizing a request or translating a short message they are often good enough. Our guide to local AI for privacy explains the options. The tradeoff is that a local setup can concentrate know-how in one person, so document it well and teach others to use it.
Protecting Undocumented Neighbors and Other At-Risk Members
Many mutual aid networks serve, and are made up of, people for whom data exposure carries serious consequences. Undocumented immigrants and mixed-status families are the most obvious example, but the same care applies to people fleeing abusive partners, people with past convictions, unhoused neighbors, sex workers, and organizers involved in protest or tenant fights. For these members, the question is not whether a data leak would be embarrassing. It is whether a spreadsheet, a chat log, or a payment record could be used to find them.
The practical response is to design as if your records could someday be requested by an authority or seen by someone hostile. That means keeping sensitive conversations in end-to-end encrypted channels, using disappearing messages, and keeping AI tools away from anything that identifies these members. The Electronic Frontier Foundation's guide to Signal groups explains how encryption, admin controls, and disappearing message timers work. Remember that disappearing message settings apply going forward and that copies may already exist on other members' devices, so set them up early rather than after a scare.
Payment apps need a look too. Venmo has historically made transaction details visible by default unless users change their privacy settings, and memos like "rent help for the Garcias" can reveal far more than intended. Volunteers handling funds should set accounts to private, keep memos generic, and never paste payment histories into an AI tool to "reconcile the books." For networks that serve immigrants specifically, our article on AI in refugee and immigrant services covers additional precautions relevant to formal organizations that often partner with mutual aid groups.
A simple "red zone" rule
Easy to remember, easy to teach new volunteers
Some networks find it helpful to define a "red zone" of information that never goes into any AI tool, any unencrypted chat, or any shared document beyond the people who need it. Everyone learns the list during onboarding.
- Anything about immigration status, even implied
- Addresses or locations of people avoiding someone dangerous
- Health conditions, medications, and disability details
- Legal situations, court dates, or contact with police
- Full lists of members, requesters, or donors in any form
Consent and Keeping People as the Decision Makers
Consent in mutual aid is trickier than a checkbox on a form. Someone asking for groceries when their pantry is empty is not in a strong position to refuse terms, and a long privacy notice does not create real choice. Meaningful consent in this context means being honest and specific, in plain language and in the languages your neighbors speak, about what happens to their information. If you use AI to translate or summarize requests, say so on the form: "Volunteers may use translation tools to read your request. We remove your name and contact information first." Then actually do that.
It also means offering a real alternative. A neighbor who does not want their request touched by any AI tool should be able to say so and still get help, perhaps by calling a hotline answered by a person. Opting out should never mean waiting longer or being deprioritized. Our article on informed consent for AI discusses how to write disclosures people can understand.
Keeping humans as decision makers is the other half. In practice, that means AI may suggest and people decide. AI can propose a category, draft a reply, or translate a message, but a volunteer reads it before anything is sent, assigned, or paid. No money moves because a tool said so. No request is closed, declined, or deprioritized by software. When a request is ambiguous, the answer is to ask the person, not to ask the model. These rules are simple to state and easy to erode under pressure, so write them down and revisit them.
Decision rights in an AI-assisted network
- AI may: translate, summarize, categorize by request type, draft messages and documents, write spreadsheet formulas, suggest likely duplicates for a person to confirm.
- People must: decide who receives help and in what order, approve every payment, contact requesters, resolve conflicts, and decide whether to adopt or drop any tool.
- Nobody may: use AI to score need, flag neighbors as suspicious, profile members, or read encrypted group chats.
Making Tool Decisions Together
In most organizations, technology decisions are made by whoever has the authority and the know-how. In a mutual aid network, that pattern is itself a problem. When one volunteer quietly connects an AI tool to the request sheet because it seemed helpful, they have made a decision about everyone's data and everyone's values without asking. Even with the best intentions, that is a hierarchy forming. The process for deciding about AI matters as much as the decision.
A collective process does not have to be slow. Many networks already use consensus, modified consensus, or simple voting for decisions about money and safety. Tool decisions can follow the same path. The key ingredients are a clear proposal, a chance for people most affected to weigh in, a trial with limits, and a scheduled check-in. Importantly, the voices of people who request help, not only those who coordinate it, should shape the decision. They are the ones whose data is at stake.
A five-step collective tool decision
Adapt to your network's existing decision process
- 1. Name the problem, not the tool. Start with "we are drowning in duplicate requests" rather than "let's use this AI app." Sometimes a better form fixes the problem with no AI at all.
- 2. Write a one-page proposal. What the tool does, what data it touches, which company holds that data, what it costs, who will maintain it, and how to stop using it.
- 3. Hear from the people affected. Share the proposal in the languages your members speak, and make space for requesters and members with safety concerns to raise objections, including anonymously.
- 4. Run a limited trial. Try it for a month, on non-sensitive tasks, with at least two people who know how it works so the knowledge is shared.
- 5. Check in and be willing to drop it. Did it save time? Did anyone feel less safe or less heard? Did it shift power? Record the decision and set a date to review it again.
It also helps to write the results into a short shared agreement, a page or two that every volunteer reads. It need not look like a corporate policy. A list of "we use AI for these things, never for these things, and here is how we protect people" is enough. For networks that want a starting point, our guide to AI for grassroots organizations includes approaches to lightweight agreements that can be adapted to a horizontal structure.
Money, Fiscal Sponsorship, and Where AI Fits
Moving money is where mutual aid networks feel the pull toward formalization most strongly. Small networks often redistribute cash directly through a volunteer's Venmo or Cash App. The IRS has said that money received from friends and family as a gift or reimbursement for a personal expense should not be reported on Form 1099-K, and that payment apps treat transfers as personal by default unless marked as payment for goods or services, according to its Form 1099-K FAQs. Even so, running large sums through one person's account can create tax confusion for that volunteer, concentrates trust in a single person, and makes the money vulnerable if the account is frozen. Reporting thresholds have changed several times in recent years and states can set their own, so a volunteer handling significant funds should check current rules with a tax professional.
Fiscal sponsorship is the common next step. A sponsoring nonprofit holds the funds, offers donors a tax deduction, and handles compliance. The tradeoff is that the sponsor will usually need records of how money was spent, and sometimes information about recipients, which can pull a network toward exactly the intake and documentation it was trying to avoid. Before signing on, ask a prospective sponsor exactly what data they require, how they store it, and whether direct cash grants to individuals can be documented in aggregate. Our article on AI and fiscal sponsor compliance covers what sponsors themselves are expected to track.
AI can help with the paperwork side of this relationship without touching recipient data. A volunteer can use AI to draft a fiscal sponsorship inquiry, summarize a sponsor's agreement into plain language for a group discussion, prepare a budget narrative, or write a donor thank-you template. What AI should not do is become a pipeline that sends detailed recipient records to a sponsor or anyone else. Aggregate reporting ("we distributed $12,400 in 83 grocery and rent grants this quarter") meets most real accountability needs while keeping individuals out of it.
So, Should AI Help Run Your Network?
The honest answer is: AI can help some of the people who run a mutual aid network, with some tasks, under rules the network agrees to together. It should not help run the network in the sense of deciding, ranking, or watching. The distinction is between AI as a quiet assistant to tired volunteers and AI as a new layer of management. The first can strengthen mutual aid by giving coordinators back time for relationships. The second recreates charity's bureaucracy with less transparency than a human case worker.
It is also completely legitimate for a network to decide not to use AI at all. Some groups will conclude that the risks to members outweigh the time saved, or that the political meaning of relying on large tech companies conflicts with their values. That is a reasoned choice, not a failure to keep up. The point of this article is not to push adoption. It is to make sure that whatever your network decides, it decides on purpose and together.
A practical starting checklist
- Audit your forms and spreadsheets and delete fields you do not need
- Set a retention schedule and delete closed requests on it
- Agree on a red zone of information that never touches AI tools
- Start with zero-data uses: drafting updates, how-to docs, and formulas
- Add translation and summarization only with identifying details removed
- Tell requesters plainly how tools are used and offer a human-only option
- Make every tool decision through your collective process and revisit it
Conclusion
Mutual aid networks face a real coordination problem. Requests pile up, languages multiply, spreadsheets sprawl, and a small number of volunteers end up carrying far too much. AI can ease some of that burden, especially through translation, summarizing requests by type, cleaning up spreadsheets with formulas, and drafting the endless updates and guides that keep everyone on the same page. Those are meaningful gains for people who are giving their evenings to their neighbors.
But mutual aid's strength comes from its values: need is enough, power is shared, and neighbors' safety comes before efficiency. AI that scores need, flags neighbors, reads encrypted chats, or concentrates data in a tech company's servers works against those values, even when it saves time. The safest path combines aggressive data minimization, a clear red zone of information that never touches AI, honest consent with a human-only option, and a firm rule that people, not software, decide who gets help.
Most of all, treat the question of AI the way you treat every other important question in your network: openly and together. Start with the problem, write a short proposal, listen to the people whose information is at stake, try it carefully, and be willing to walk away. Coordination without bureaucracy is possible with or without AI. What matters is that the tools serve the solidarity, and never the other way around.
Support Your Network Without Compromising Its Values
We help community groups, mutual aid networks, and the nonprofits that sponsor them think through privacy-first coordination workflows, data minimization, and collective technology decisions.
