Harnessing the Power of Conversational and Generative AI: An Introductory Guide for Labor and Community Organizers
At this point, I don’t need to introduce you to AI. This past year alone, we’ve seen content about generative AI popping up everywhere from the annoying social media influencer that’s lying to you about how you can make $10,000 a month with it to the wealthy executives warning that AI will end humanity (interestingly, they always base this apocalyptic warning on how scary it is that this technology is now in the hands of the people). Well, if they’re so worried about this technology being used by us, then maybe we should actually explore what it can do.
To skip to one of the 5 methods now, I have curated a list for you.
First, a note on privacy and security
From the beginning, let me include a very important warning: This guide is only intended for above-ground organizations and legal activists. While this article does not get into what tactics are appropriate for different organizations, it’s important to remember that different levels of struggle require different magnitudes of privacy and security. You should always have in-person conversations with your organization to understand what is, and isn’t, okay to write online.
Even in the union movement, it’s generally always a good idea to understand what can and cannot be written or shared online. While a union organizer isn’t likely to face criminal charges for accidentally violating a contract provision or the National Labor Relations Act’s (NLRA) requirement for strike notices, union leaders should still be careful about what they put in writing online — whether it’s on a Google Docs agenda or in a conversation with chatGPT. And, out of deep respect for our members, we should always be concerned with protecting the data which they entrust to us. Not having good “digital hygiene” as an organizer can result in repercussions against the organizer and their union or organization, ranging from fines and injunctions to possible criminal charges.
The reader should remember that everything they write online is potentially discoverable—should a lawsuit or more serious charge arise from their organizing activity—meaning that it can be used as evidence against them in court. This includes information entered in conversation with AI. As a primer, the EFF has a list of security guides that are concise and informative. That being said, above-ground organizing typically uses legal tactics which carries less risk than forms of organizing that includes extra-legal tactics, so it is generally okay for them to more fully utilize digital tools.
There is always a contradiction between security and organizing. Organizers who want to build majorities in their organization have to engage in structure-based organizing, rather than the self-selection method of just allowing already interested people to join. While I won’t go in-depth on that here, it’s important for the reader to understand that movements and organizations grow most when recruitment is structured around “turf,” or geographical lines where the universe of potential members is finite, such as a workplace, apartment building, or neighborhood. This allows organizers to approach people who are not already on our side rather than only engaging those who already agree with us. In order to do this, organizers first need to form lists that allow them to know their “universe,” make assessments, and systematically work through that list to recruit members, identify leaders, and build majorities.
However, list generation can include sensitive data and can be used against activists in the case of state repression. Therefore, having a good sense of what kinds of lists are appropriate, what data is okay to store in those lists, and whether they should be online or on paper, is essential. In a union, lists are generally always okay to store online (ensuring they are private and the account storing them has the proper security measures in place) as they are already available to the employer and the public (after all, if your workplace is union, your employer and the National Labor Relations Board (NLRB) already know who is and isn’t a member). In an organization engaging in more risky tactics, the organizer should be far more careful and ensure that such questions have been deliberated on thoroughly by the organization before producing a list, especially online.
While digital tools can never replace on-the-ground and in-person organizing, organizers can use them to augment their work when appropriate. The emerging models of generative and conversational AI can enhance digital organizing, mobilizing, propaganda, and educational programming. One final caveat: as a union organizer, the majority of these suggestions are aimed at other worker organizers. However, they can be adapted to other types of organizations.
So now that we have all that out of the way, let’s get into the five ways that AI can be utilized for these purposes.
1. Generating, modifying, and analyzing lists
ChatGPT-4 introduced a new feature called code interpreter. With this feature, the user can upload spreadsheets directly into the platform and direct the AI to modify or edit the list. This is especially helpful as a time saver. Traditionally, you would either have to do these things manually, cell by cell, or by using your own code, such as Visual Basic or Python, to perform complex operations. ChatGPT-4 does this for you by using Python to perform almost any task you need it to on your spreadsheets. It can debug its own code, work through problems, and fix its own errors.
Example 1: Compiling a list of union members
Let’s say that I’m working with several different lists in one of the workplaces that I organize. At workplaces that are under the provisions of a collective bargaining agreement, it’s typical to deal with multiple lists of the bargaining unit (all the workers in a workplace that fall under the protections of the union contract), and for these lists to all come from different sources. Often, lists can be structured differently and contain errors, requiring substantial effort to merge them into one single list. This process can be lengthy if you don’t know how to write your own code to do the manual tasks for you. With ChatGPT-4, however, the task becomes much more simple.
In this example, I was able to upload the different lists and direct ChatGPT to merge the lists. This included small tasks which take a long time, such as merging two datasets where one spreadsheet includes the person’s full name under one single column and the other that splits first and last name into two columns. Take, for example, the below spreadsheets. I used test data with over 100 values for each spreadsheet but simplified here for the reader’s convenience.
I prompted ChatGPT with the following (also note how it understands that I am just using this to provide an example here):
The result was a successfully merged spreadsheet. However, some names on the first had initials for the middle name, or additional first names, that the second list did not have. So, I worked with ChatGPT to develop conditions that would allow it to merge names based on last name and some match in the first name. Below is an example of how ChatGPT is able to work through problems and adjust its own code. Eventually, the names were merged successfully.
Later, I realized that we had some missing information for each member. I prompted ChatGPT to generate a spreadsheet of all contacts who had missing contact information. It did so perfectly and I was able to download the file directly.
Generative AI can be used like this to analyze other kinds of data, too. For example, it can analyze data from crowdsourcing efforts, such as surveys or petitions, or aid in fundraising by analyzing funding trends and identifying donors or sectors to prioritize.
2. Analyzing data to create visualizations
As a language model, ChatGPT-4 can be used to analyze data and create visualizations by generating code that can be used to create charts and graphs. This can be done by providing the model with data in a structured format, such as a CSV file, and then using the generated code to create visualizations using a tool such as Matplotlib or Seaborn. However, it can also create charts and graphs directly if you enable “code interpreter.”
Example 2: Creating charts of average global sea level rise
Let’s say that I’m an environmental organizer and want to create visualizations to assist me in teaching people about rising sea levels and the disparate impact they have on different regions. In order to do this, I uploaded a CSV file that contained data from the National Oceanic and Atmospheric Administration (NOAA) which give estimates of the changes in mean sea levels. At first, I wanted it to generate a map. However, there were some limitations to its geospatial capabilities which it listed below along with suggestions of what I can do should I want to proceed. (Note that it offered to guide me through the process of inputting the code into geospatial software myself, and that it would have written and debugged the code for me)
Because I was just using this as an example, I directed it to simply move forward with generating three graphs. The response is below, where it shows its work and explains what each graph means.
And the result? The following three graphs. Note that I could have prompted it to change the aesthetics.
3. Analyzing texts and creating summaries
Generative AI can be used to analyze text data, such as essays, to generate summaries to identify key themes. This is called “natural language processing” and can be helpful in summarizing important texts or as a starting point for creating social media posts for articles you want to promote.
Example 3: Creating social media posts based on a member’s essay
Oftentimes, organizations want to promote the essays or op-eds that their members write on social media. To do this, they have to extract the main points from the text. As an example, I wrote a rather lengthy essay critiquing certain theoretical trends in the feminist movement. I wanted to condense this essay into a series of posts so that I can share some key points in a more accessible format before publishing the full-length essay.
The process was very simple and did not require ChatGPT-4. I was able to do it on ChatGPT-3.5. I prompted it with “Summarize into 10 compact bullet points” and then copy and pasted the entire essay. The result was the following.
Before posting to Instagram, I edited a few of the points in order to more accurately reflect what I was trying to say. However, if you were to read my essay, you would find that these points accurately list the main themes and also adopt my writing style. The final Instagram post can be viewed here.
As you can see, the applications of this are enormous. It doesn’t eliminate the need for you to edit the output before posting or sharing, but it does create a baseline that allows you to edit and produce content much more quickly.
Example 4: Summarizing important articles for political and popular education programs
It can even aid in learning by inputting entire texts and prompting it to summarize key themes, identify trends, etc. For example, I copy and pasted an article I wrote for Truthout regarding worker surveillance and then prompted it to: “Identify the author and publication information. Summarize key concepts and themes. Condense into 5 main bullet points.” The results are as follows.
Great start! But now I want to know what sources the author cited in the text. So I use the following prompt and get a very accurate response.
Now, I’m interested in learning more about the specific means of surveillance the author references. But, I also want to know if there are other methods by which employers monitor their workers that the author misses. So I use the prompt “What means of surveillance were identified in the article? Are there other means which could have been considered by the author?” The results are as follows.
Note that there is a text limit and that you might have to upload longer texts separately and ensure that it knows they are the same text. Alternatively, you can use a splitter to get around this.
Generative AI can also be used with emails and other communications to identify key themes and sentiments. This can help organizers to better understand the concerns of their membership and tailor their messaging accordingly.
4. Generating information on your legal rights to organize
While generative AI cannot replace the need for an attorney, especially when dealing with cases of retaliation, it can provide organizers with relevant legal information to assist workers in asserting their rights.
Example 5: Generating references to defend organizing rights
Let’s say that I know that workers are protected under federal labor law but I don’t know exactly where or how. I can use the following prompt to get exact information along with a source.
Alternatively, let’s say that I wanted to know about the protections workers have from retaliation but I don’t know what section of the law those protections are contained within. In this case, I used the same prompt but with a more general statement.
While all such examples should be vetted by an attorney or legal team prior to publishing, this provides a pretty accurate snapshot and directs the organizer where to look should they need more information.
5. Creating Workshop Templates
Generative AI can be incredibly helpful in creating workshops. It can provide sources and references, generate templates, create content, identify ways to engage participants, and even develop processes for workshop evaluation.
Example 6: Creating a “Know Your Rights” Workshop
Building off the previous example, let’s say that I want to organize a “Know Your Rights” workshop for a group of workers who want to form a union in California. I need to know the relevant federal and state laws and need to have a structure for the workshop. I can use AI to generate those items with the following prompt.
The result is the following workshop, complete with reference links to the corresponding enforcement agency.
But wait, nine agenda items seem a bit excessive. So I prompted it to condense it down to 4 main items that would be more digestible. The response is below.
Note that it came up with a time limit on its own. You could include a time limit and other restraints in your prompts.
Final thoughts and other considerations in using generative AI for organizing
Although AI can be used to augment organizing work, there are several limitations. The most effective use of these tools comes from effective prompt engineering, which essentially means understanding how to write prompts to get the desired output. This requires some trial and error but can also be augmented by different prompt plug-ins (I won’t list them here as I have not reviewed potential privacy concerns with third-party plug-ins). And, in addition to the security risks noted in the introduction, there are other restraints you should be aware of. For example, ChatGPT, which was the tool used in this article, has limited knowledge up to the year 2021. However, organizers can get around that by using other AI models and—of course—simply doing their own research.
You can look at Google Bard and Bing AI, both of which have access to more current information. Bing AI is a conversational AI, best used in the Microsoft Edge browser, and is geared towards search engine functions. Rather than explain the differences myself, I thought I’d allow Bing AI to.
Whatever tool you decide to use, just remember the considerations for safety and privacy. As organizers, we have an immense responsibility to put the needs of the movement and our members first. Generative and conversational AI are, at the end of the day, simply tools that can be helpful in some situations and harmful in others. They should always be used with consideration for security and privacy and a concern for the overall well-being of the organization and membership.
If you found this article helpful, please feel free to comment below and share how you use generative or conversational AI.
