As artificial intelligence systems become more autonomous and capable of generating public-facing content, tax-exempt organizations are beginning to confront a novel, but increasingly relevant, question: When AI agents speak, advocate, or publish, whose activity is it?
For organizations recognized under Internal Revenue Code Section 501(c)(3), the answer carries meaningful implications. While the technology may be new, the legal framework is not. Existing IRS rules governing lobbying and political campaign activity apply based on attribution, not authorship. In other words, the central question is not whether a human drafted the content, but whether the activity is properly treated as that of the organization.
Attribution in an AI-Driven Environment
Where a nonprofit funds, operates, or maintains the infrastructure through which AI-generated content is created and disseminated, there is a strong basis for treating that activity as attributable to the organization. This includes situations where the organization deploys the systems, hosts the outputs, or publishes content through its own channels or affiliated accounts.
From a regulatory standpoint, disclaimers such as characterizing outputs as “AI-generated” or not reflective of the organization’s views may be helpful for transparency, but they are unlikely to be dispositive. The IRS will generally look to control, funding, and operational responsibility in determining whether an activity is conducted by the organization.
When Does AI Output Become Lobbying?
The distinction between permissible educational activity and lobbying remains grounded in familiar principles. Under federal tax law, lobbying generally involves attempts to influence legislation, which typically requires both:
-
-
- A reference to specific legislation, and
- A call to action (e.g., urging the public or legislators to take a position)
-
AI-generated content that discusses public policy issues in a general, analytical, or educational manner can fall within permissible bounds. However, outputs that advocate for or against specific legislative proposals, or that encourage engagement with lawmakers, are more likely to be treated as lobbying.
For organizations that have not made the §501(h) election, the “substantial part” test applies, which is inherently facts-and-circumstances based and lacks a clear quantitative threshold. In this context, even relatively small amounts of lobbying activity may create uncertainty. By contrast, organizations that elect under §501(h) (via IRS Form 5768) benefit from a clearer, expenditure-based framework for measuring permissible lobbying.
The §501(h) Election: What It Is and Why It Matters
One of the most practical steps a §501(c)(3) organization can take when there is any possibility of lobbying, whether human or AI-generated, is to make the §501(h) election by filing IRS Form 5768.
At a high level, the §501(h) election allows a nonprofit to measure its lobbying activity based on expenditures, rather than under the default “substantial part” test. This distinction is critical.
Under the default rule, an organization may not have a “substantial part” of its activities devoted to lobbying, but the IRS has never defined “substantial” with precision. Instead, it is a facts-and-circumstances analysis that can take into account time, effort, and overall organizational focus, not just dollars spent. This creates a meaningful degree of ambiguity and risk, particularly in newer or less predictable operating environments (such as AI-driven platforms).
By contrast, making the §501(h) election replaces that subjective standard with a clear, quantitative framework. Specifically:
-
-
- Lobbying is measured based on actual dollars spent, not overall activity levels
- The IRS provides defined thresholds (based on the organization’s budget) for how much lobbying is permitted
- The rules clearly distinguish between direct lobbying and grassroots lobbying, each with its own limits
-
For most organizations, these thresholds are relatively generous and provide a safe harbor so long as expenditures remain within the prescribed limits.
From a risk-management perspective, especially in the context of AI-generated content, this election is particularly helpful for a few reasons:
First, it reduces uncertainty. Rather than trying to assess whether a small percentage of outputs (e.g., 1–5%) might be deemed “substantial,” the organization can instead track and cap the resources devoted to lobbying-related activity.
Second, it focuses the analysis on what the organization actually controls, namely, how much it spends supporting or maintaining systems that may generate lobbying content, rather than attempting to quantify or predict the volume of potentially qualifying outputs.
Third, it provides a defensible compliance framework in the event of regulatory scrutiny. So long as the organization stays within the expenditure limits, it can more confidently demonstrate that it has not exceeded permissible lobbying levels.
Importantly, making the §501(h) election does not require the organization to engage in lobbying, nor does it increase scrutiny. It simply provides a clearer, more administrable method of measurement if lobbying does occur.
For organizations operating in emerging or dynamic environments, where content generation may not be entirely predictable, the §501(h) election is often a prudent and relatively low-burden step to bring clarity and structure to an otherwise ambiguous standard.
The Bright-Line Rule on Political Campaign Activity
More restrictive still is the prohibition on political campaign intervention. Under Treasury Regulation §1.501(c)(3)-1(c)(3)(iii), a §501(c)(3) organization may not participate or intervene in any political campaign on behalf of, or in opposition to, a candidate for public office. This is an absolute prohibition, with no de minimis exception.
Accordingly, AI-generated content that references candidates, elections, or electoral outcomes, particularly in a manner that could be perceived as favoring or opposing a candidate, presents heightened risk. As with lobbying, the automated nature of the content does not insulate the organization from attribution.
At the same time, Treasury Regulation §1.501(c)(3)-1(d)(3) makes clear that nonprofits may engage in educational activities related to public policy issues, even where those issues are politically relevant. The key distinction is that such communications must not directly or indirectly support or oppose candidates.
Practical Risk Mitigation in an Autonomous Context
Given the inherently unpredictable nature of AI-generated outputs, traditional pre-clearance models may not always be feasible. Instead, organizations should consider a combination of system-level guardrails and responsive oversight.
Preventative measures may include designing prompts, constraints, or filters to reduce the likelihood that systems generate content involving specific legislation, calls to action, or references to candidates and elections.
Equally important, however, is the development of a responsive framework. Because organizations cannot fully direct or anticipate every output, there should be a mechanism to monitor public-facing content and address instances where outputs present one-sided or advocacy-oriented positions. This may include supplementing such content with additional context, clarifying information, or alternative perspectives to maintain a balanced, nonpartisan, and educational posture.
Importantly, this does not require presenting all viewpoints in every instance. Rather, it reflects a commitment to ensuring that, when content risks being perceived as advocacy, the organization takes reasonable steps to contextualize it in a manner consistent with its exempt purpose.
Conclusion
While AI introduces new operational complexities, it does not fundamentally alter the legal standards governing §501(c)(3) organizations. Attribution will continue to follow control and infrastructure, and the familiar lines between education, lobbying, and political campaign activity remain in place.
Organizations deploying or supporting AI systems should therefore approach these tools with the same level of care applied to human-generated communications, augmented by thoughtful guardrails and responsive oversight mechanisms tailored to the realities of autonomous content generation.