Election-Year Lines Nonprofits Cannot Cross: AI Content and 501(c)(3) Nonpartisanship

How the IRS Decides: Facts, Circumstances, and Seven Factors

Because the prohibition is categorical, the hard work is definitional. What counts as intervention? The IRS answer is that there is no single formula, and that it examines all the facts and circumstances of each case. The central document is Revenue Ruling 2007-41, which works through twenty-one factual situations across voter education, candidate appearances, issue advocacy, business dealings, and websites. It is the most useful compliance reading a communications team can do, and it is readable without a tax background. The IRS also maintains a landing page on political campaign intervention by 501(c)(3) organizations that collects the published guidance in one place.

A facts-and-circumstances standard has a specific practical consequence that most organizations underweight. It means the same sentence can be permissible in March and a problem in October. It means a statement that is fine on its own becomes a problem alongside the three posts published around it. It means context you did not author, such as what a linked page says this week, becomes part of what the IRS looks at. No reviewer can apply that standard by instinct, which is why the factors need to be written down somewhere a communications coordinator will actually see them.

The sharpest tool in the ruling is its treatment of issue advocacy, which is where charities legitimately spend most of their public voice. A 501(c)(3) organization may take positions on public policy issues, including issues that divide candidates, and it does not lose that right because an election is approaching. What converts issue advocacy into intervention is a combination of signals, and the ruling identifies the factors the IRS weighs. They are worth reproducing in full, because they double as a review checklist.

Where Organizations Actually Cross the Line

Almost nobody arrives at a violation through a decision to endorse. The realistic failure modes are mundane, and they cluster in five places. The first is timing. A statement that has been in your advocacy library for two years gets reposted in late October with a sharper headline, and the content has not changed while the context around it has. Timing is one of the seven factors precisely because the same words carry a different signal when ballots are in hand, and the fix is calendar awareness rather than silence.

The second is voice bleed. Leaders of a charity retain their rights as private citizens, and Publication 1828 is explicit that the prohibition is not intended to restrict free expression on political matters by leaders speaking for themselves as individuals. The constraint is that they cannot make partisan comments in official organization publications or at official functions. In practice, the boundary between a personal account and an organizational one has nearly dissolved. An executive director whose personal profile lists their title, uses the organization's logo as a header, and posts from the same phone that runs the organizational account is conducting an ongoing experiment in how clearly that separation reads to a stranger. Resharing a leader's personal political post from the organizational account collapses the distinction entirely, and so does using staff time or organizational systems to produce it.

The third is linking. Revenue Ruling 2007-41 treats a website as a form of communication by the organization, and the factors for links include the context of the link on your site, whether all candidates are represented, and the directness of the path between your page and material favoring or opposing a candidate. Two consequences follow. You are responsible for where you send people, and a link to a resource page that itself links to partisan material is not automatically far enough away. You are also exposed to drift, because a page that was a neutral civic resource when you linked to it in August may read very differently in October, and nobody on your team is checking.

The fourth is the uninvited or unevenly invited candidate. A candidate shows up at your community event, is handed a microphone during the program, and speaks warmly about your work. Nothing was planned and the exposure is real, because other candidates had no equal opportunity. Organizations that host public events during an election season need a written protocol for this, including who in the room has authority to decline a speaking request and a neutral script for declining it. The fifth is coded issue language, where a statement names no candidate but uses a phrase that functions locally as shorthand for one. The seven-factor analysis is designed to see through that, and the factor about whether the issue distinguishes candidates for a given office is the one that catches it.

Four AI Workflows That Create New Exposure

Drift in a draft is manageable, because a draft gets read. The genuinely new exposure comes from workflows where generated text reaches the public without passing a human who was looking for this specific problem. There are four worth auditing before any election season, and each has a different control.

The first is high-volume short-form content. Social captions, newsletter blurbs, subject lines, and post variants are the output most often generated in batches and reviewed least carefully, because each individual item looks trivial. Short form is also where nonpartisanship is hardest to maintain, since there is no room for the qualifications and context that keep a longer piece balanced. A caption has to land in nine words, and the nine words that land hardest are usually the ones with an opponent in them. Compounding the problem, batch generation means a reviewer skims thirty items for tone and typos rather than reading each against a compliance standard, which is exactly the failure described in our article on AI for nonprofit social media.

The second is autonomous and scheduled publishing. Content calendars populated weeks ahead, automated repurposing that turns a blog post into a thread, and agent-style tools that draft and post on a schedule all share a property that is fine in ordinary months and dangerous near an election: the publication moment is decoupled from the review moment. Copy approved in September for a mid-October slot was approved against a different context than the one it lands in. Anything scheduled across an election date needs to be re-read on the date, not on the date it was written, and systems that publish without a human in the loop should be paused for the window entirely.

The third is any public-facing chatbot or AI search feature on your website. If visitors can ask it questions, they will ask it about the election, and in an election season they will ask it things like which candidate supports your cause, what your organization thinks of a particular officeholder, or how they should vote on a ballot measure. A general-purpose assistant with a thin system prompt will answer, fluently, in your organization's visual identity, and that answer is a statement by the organization. Nobody approved it, nobody will see it, and the visitor has no way of knowing it was not reviewed. This is the highest-exposure item on the list because it generates novel organizational speech at a volume no one is reading.

The fourth is generated imagery and video, which is a distinct category rather than a harder version of text. Synthetic images or audio of a candidate carry exposure beyond the tax question, because most states now regulate synthetic media in election communications. As of 2026, roughly thirty states have enforceable laws addressing AI-generated deepfakes in political advertising and election communications, most of them requiring disclosure within a defined window before an election, and a few prohibiting certain content outright regardless of disclosure. State approaches and timing windows vary considerably, which makes this a jurisdiction-specific question. The simple institutional answer is a flat internal rule: your organization does not generate images, audio, or video depicting any candidate or officeholder, for any purpose, including internal mockups that could escape. Provenance practices help you prove what you did and did not make, which is the argument in our coverage of content provenance and C2PA for nonprofits.

The Nonpartisanship Pass, and Guardrails Before the Draft

A policy states the standard. A review step applies it. The practical mechanism is a nonpartisanship pass: a distinct, required review stage that happens after the usual edit for accuracy and tone, performed by someone looking only for this class of problem, with a written checklist derived from the seven factors. The reason it has to be distinct is that a reviewer checking everything at once checks nothing reliably, and compliance signals are specifically the ones that read as fine to a reader focused on quality. Good copy and compliant copy are different judgments.

The checklist should be concrete enough to answer yes or no. Does this name, picture, or otherwise identify a candidate or an officeholder who is also a candidate? Does it express approval or disapproval of anyone's positions or record? Does it mention voting, the election, or a date in the window? Is the issue one currently distinguishing candidates for an office on our list? Have we published on this issue on a steady cadence independent of the election, or is this a surge? Is there a non-electoral event, such as a scheduled legislative vote, that explains the timing and the reference? Does any link in this piece lead, in one hop or two, to material favoring or opposing a candidate? Would a stranger with no context read this as a preference? That last question is the one that catches coded language, and it is best answered by someone outside the team that wrote the piece.

Guardrails before the draft reduce how much the review has to catch. Put the constraints in the system instructions of whatever assistant or custom workspace your communications team uses, rather than relying on each staff member to remember them in each prompt. Useful instructions include a standing prohibition on naming candidates, parties, or officeholders in generated public copy; a requirement to describe policy positions without attributing motive; a ban on election references, voting language, and deadline urgency; and an explicit instruction that requests to make copy more compelling must be satisfied without introducing an opponent. Add a refusal behavior for requests that would compare candidates. These guardrails are reliable enough to be worth having and nowhere near reliable enough to replace review, which is the standing lesson of our article on adversarial prompts against a nonprofit chatbot.

Change how your team uses revision prompts, too, because this is the cheapest intervention available. Replace "make this stronger" with instructions that specify the kind of strength you want: more concrete, more specific about our program, better evidenced, shorter. Each of those improves copy without reaching for an antagonist. And keep the prompt alongside the output in whatever system holds your content calendar, so a reviewer can see what was asked for. A piece of copy whose prompt history includes "make it more urgent, election is coming" is a different artifact from the same copy produced by a neutral request, and if anyone ever asks, the record matters.

Documentation That Shows Even-Handedness, and When to Call Counsel

A facts-and-circumstances standard rewards organizations that can produce facts. If your nonpartisanship is ever questioned, by the IRS, by a journalist, by a funder, or by a community member who read one post uncharitably, the useful response is a record rather than an assurance. Most of that record is cheap to create at the time and impossible to reconstruct later, which is the entire argument for doing it as you go.

For candidate-facing activity, keep the invitation itself: who was invited, when, through which channel, on what terms, and what each one said. Keep the questionnaire exactly as sent, the full set of responses, and the record of how non-responses were handled. Keep the forum format and the composition of the question panel. For published content, keep a log of what went out, when, who approved it, and which review steps were completed. For AI-assisted work, keep the prompts and system instructions alongside the outputs, plus a note of which tool and which configuration produced them. For the chatbot, keep the scope configuration, the test results, and the log samples that were read. Set a retention period that outlasts the cycle comfortably, because questions arrive after the season ends, and record the whole thing somewhere the organization keeps records rather than in one staff member's drive. Registering election-season publishing as a standing risk with its own controls and review date fits naturally into the practice described in our article on AI risk registers for nonprofit boards.

Board involvement deserves a sentence of its own. The board should approve the election-season policy, know who the named approver is, and receive a brief report after the window closes covering what was published, what was flagged, what was changed, and whether anything required counsel. That is a ten-minute agenda item that converts an operational control into a governance one, and it also means the board is not learning about an incident from a reporter.

Finally, the referral rule. Nothing in this article is legal advice, and the facts-and-circumstances standard means genuinely close calls cannot be resolved by a checklist. Send it to counsel when a communication identifies a candidate or an officeholder who is running and you want to publish it anyway; when you are considering a voter guide, questionnaire, or forum and want the instrument reviewed before it goes out; when a candidate has asked to appear at your event or to use your facility; when a staff member's personal political activity is becoming hard to separate from their role; when you are working alongside a 501(c)(4), a political organization, or a coalition whose partners have different permissions; when an AI system has already published something that worries you; and whenever anyone on the team is arguing that something is probably fine. The cost of an hour of counsel is trivially small against the exposures described at the top of this article, and the willingness to use it is itself a sign of the care the standard looks for.