The Organizing Center Information on New Technologies and Ethics

We’ve thought a lot about new technologies and ethics

Instead of a summary we’d like to share our full policy below. This is a result of rigorous thought about who we work with, how we work and how to use this technology:

Impact Mitigation, Ethical Design, Guardrails and Collective Power

For organizations working with the Organizing Center. Forward it, question it, check it.

Contents

1. Why we build AI tools

2. Our Gen AI Values

3. How the values show up in the design

4. The guardrails policy

5. What your organization controls

6. The technology underneath

1. Why we build with AI tools

Generative AI is a real technological shift. It is also landing inside systems that were already failing the people our clients organize, and it is led by corporations built for extraction.

Most of the harm now attributed to AI traces less to the technology than to the logic of the companies driving it. Maximize growth, valuation and profit. Treat the impact on communities, workers and the planet as an externality for later. Those defaults are not growing pains. They are what the companies were built to do. Point any powerful tool at that logic and it does what the surrounding system already does, faster. Replace people. Replace relationships. Make the transaction cheaper.

We start from the other end. Values first, then technology. A tool built by people accountable to movements rather than shareholders behaves differently.

Skepticism about this technology is warranted. The costs are real, across democracy, labor, privacy, the environment, and the slower question of what predictive systems do to political imagination over time. None of those costs are resolved by good intentions.

Abstinence is not on offer. AI already runs inside the CRM, the ad platform, the translation tool, the transcript of the last meeting, the autocomplete in the email. Vendors use it and do not always say so. In most organizations, someone behind on a deadline is pasting internal strategy into a consumer chatbot right now, because workloads expect them to scale outputs with AI. Refusing to touch AI as a sector does not reduce its systemic harms at even noticeable levels. It cedes the tools to the people already using them against us, and buys a purity closer to marketing than to commitment. The choice was never whether. It was whether anyone made it deliberately.

→ Like it or not, you are already using AI

2. Our Gen AI Values

Counter to the assumptions of Generative AI in operation today,

We believe the promise of Gen AI is fulfilled when used to:

Elevation > Automation: We aim to elevate, not just to automate.

Make Democracy better, not just politics cheaper: We see the power of gen AI to enable creation, bring people together, empower those fighting injustice - not just increase efficiency.

To promote participation: To build people power, rather than a politics of transaction

Augment through Specialists: Derive power from the wisdom and talents of domain specialists, not replace them for AI slop.

Amplify the voices, visions and solutions for change of diverse communities, rather than dull them with the hegemonic influence of massive LLM data set

Based in Human Intention and Leadership: Gen AI is only as valuable as the intention, direction and strategy we use to collaborate with it. Human leadership rooted in lived experience and community values cannot be replaced by predictive patterning and mimicry.

Additionally, we pledge to be aware of and make all efforts to mitigate the potential harms - individual and systemic - in the use of generative AI tools and to be transparent in our use.

3. How the values show up in the design

Values that do not change what gets built are decoration. Four design commitments carry ours.

Elevation over automation. Automation takes what people do and has a machine do it instead. Elevation takes what people know, and amplifies it, so more people can create at a higher level, more autonomously, and more aligned with collective strategy. Every design decision resolves in that direction.

Friction by design. Commercial AI maximizes adoption through the frictionless experience: say the word, the machine obliges. The research points the other way. A joint MIT Sloan and Accenture study found that moderate, targeted friction, in the form of prompts nudging users to scrutinize what the AI produced, significantly improved accuracy without meaningfully slowing the work. MIT's Renée Richardson Gosline, who led that research, has documented that people handed AI-generated content anchor on it, with roughly 60 to 80 percent of their final work mirroring the machine's first suggestion. Once you have seen it, you cannot un-ring the bell. Our tools push back, ask the clarifying question, and say are you sure that is what you actually want here. That is a feature. Building it requires understanding the work well enough to know where the blindspots are.

The expertise and experience boundary. A good designer knows the limits of their own expertise and when to consult the client. Our tools do the same. Where something is a matter of craft or best practice, the system carries it. Where something is a matter of lived experience, community knowledge, or political judgment, the system asks rather than filling the blank with a statistical average. Locating that line is most of the work.

Two layers, not one. A Movement Intelligence Layer holds organizing knowledge built from fourteen years and 600+ campaigns across 40 countries. A Custom Layer holds your strategy, brand, messaging framework and guardrails. Generic AI has neither. What separates the two is whose knowledge and whose voice the tool works from.

→ Friction by Design: Rethinking the AI Revolution for Movements and Democracy → What the machine is made of

4. The guardrails policy

The Organizing Center operates under a written ethical guardrails policy governing how the tools handle representation, bias, identity and political voice. Three features of it are unusual.

It is not a content moderation policy. Most AI use policies list prohibited outputs and enforce them reactively, to protect a platform from embarrassment. That framing misses most of what matters here.

It treats bias as three problems rather than one.

• System-generated bias is the model producing something biased from a neutral request. Ask for a woman in her forties and the model supplies a stereotype. Specify a dark skin tone and the model lightens it. That is the model's voice overriding the user's, and most of our guardrail work is aimed there.

• User-requested bias is a request carrying an unexamined assumption. The response is usually to surface it and let the person decide, not to refuse.

• Corporate and marketing bias is the register underneath both: the professional voice that positions institutions as heroes and communities as beneficiaries. Progressive organizations are not exempt from any of the three.

It is enforced in layers, not written down and hoped for. Rules live at four levels: hard constraints in the system prompt, programmatic correction of prompts before anything is generated, non-overridable parameters at the image model itself, and softer coaching in the conversation. The level a rule lives at determines how absolute it is.

The policy is also explicit about power and about the limits of our own authority. It states that the Organizing Center serves organizations and communities fighting unjust power differentials, that the guardrails are calibrated for that, and that where the system does intervene the thresholds are not symmetrical between punching up and punching down. It states equally plainly that we have no standing to set representation standards for communities we do not belong to. Few lines, held firmly. Deliberate restraint about the rest.

A duty-of-care register covers the harms that come from language models rather than image generation: crisis disclosure, advice a machine has no business giving, safeguarding, and the slower risks of dependence and deskilling. It names what is handled, what is not yet, and what we do not yet know how to solve.

5. What your organization controls

The guardrails are tiered, because not every decision is ours to make.

Tier 1 is ours and does not move. A small set of things the platform will not produce for anyone, each stated with its reason. Alongside them sit platform commitments: generated images carry a non-removable AI disclosure mark, and creations are attributable to an authenticated user, with only your organization able to resolve a creation to a person.

Tier 2 is on by default and yours to adjust. Most representation guardrails sit here, set for the common case. Your organization can change them deliberately and on the record where your community's context calls for it. A guardrail against sexualized imagery is right almost always, and might be wrong for a sex workers' rights organization representing its own members. An override is not a loophole. It is your organization taking documented responsibility for a choice that belongs to you.

Tier 3 is not ours at all. Community terminology, contested language, how your community is represented, your aesthetic and register.

The mechanism is a setup process we run with you: a workbook covering risk tolerance, brand and voice, strategic priorities, community terminology, and what gets escalated to staff versus handled at the group level. Every setting is recorded with its reason, because a line without a stated reason drifts.

One assumption underneath this is made deliberately and stated openly. The person who sets these controls speaks for your organization, and knows your collective agreement about how you represent your community better than we do. we does not adjudicate that from outside, and is not positioned to.

6. The technology underneath

The language layer: Change Agent

Our custom systems run on Change Agent, a generative AI platform built specifically for causes and campaigns rather than adapted from a corporate product. Three points matter for internal review.

Your data. On Change Agent, your data is never sold, shared, or trained on. You can delete it at will, or use temporary chats that store nothing. Data is held to a high security standard, with geographic options.

Your content is allowed. The terms of service of every major commercial AI platform restrict political content. Change Agent's do not. Political and advocacy work is what the platform exists for, which matters for organizations whose core work gets flagged as a violation elsewhere.

Who is behind it, and who else is on it. Change Agent's first backer was a civil rights organization rather than a venture fund, and its largest funder still is one. Organizations already using the platform include State Voices, Community Change, Movement Voter Project, Center for Common Ground, VOCAL New York, SEIU, and The Movement Cooperative.

The image layer: Black Forest Labs

Image generation runs on FLUX models from Black Forest Labs, a German company headquartered in Freiburg. Three grounds for that choice survive checking. It is subject to EU law and is a signatory to the EU AI Act's General Purpose AI Code of Practice, obligations US-only providers are not structurally subject to. Its models are open-weight and self-hostable, which means real exit-ability rather than vendor lock-in. And unlike OpenAI, Google and xAI, it holds no US defense or intelligence contracts.

That picture is not fully clean. We cannot stand behind all of Black Forest Labs' backers and dependencies. The same is true of the major social media platforms, whose investors and executives we would not defend either, and which we still recommend organizations use, because building power means working where people already are. Image generation offers a narrow field, and every option in it is compromised somewhere. Among them, this is the one we judge most defensible on the grounds above.

On energy and environmental impact

The environmental cost of AI is real, and organizations fighting data center expansion are right to question, “Is this worth it?”. we has not solved the major environmental concerns. The choices available have been made in a direction we are proud of: Change Agent reports that 250 of its users consume roughly the energy of a domestic refrigerator, and the image layer runs on a German and EU grid substantially lower-carbon than the US hyperscale average, under binding EU efficiency law with no US federal equivalent. That reduces a footprint. We acknowledge it does not eliminate one, and it does not resolve the problem at sector scale.