Public Evaluation of Public Benefit

A mechanism for more robustly aligning an organization with a public benefit purpose.

Luke Thorburn

Published 12 Aug 2026

The challenge: robustly aligning organizations with a public benefit purpose

In general, existing methods for aligning an organization (e.g., a non-profit, trust, benefit corporation) with a public benefit purpose are imperfect and limited. In particular, while some frontier AI companies (e.g., OpenAI, Anthropic) and neolabs (e.g., Inflection, Thinking Machines, Discovery Loop) are structured as fortified Delaware PBCs with an official public benefit purpose, there are plausible scenarios in which the operative purpose of those organizations will drift towards private interests without any party violating a formal duty with sufficient unambiguousness to trigger meaningful corrective feedback.

Given the scale of influence of frontier AI companies — which mediate significant fractions of human attention and decision-making and increasingly have reserve control over a meaningful fraction of economic activity — it is especially important that we develop mechanisms to robustly align them with their public benefit purposes in the long term.

In general — public interest organizations

There are several hard problems in public interest corporate governance, including:

• the procedural and epistemic challenge of defensibly determining what the public interest is in a given context;

• structuring the incentives of all relevant decision-makers (staff, management, directors, stewards) to support the purpose;

• structuring the influence of funders and other external actors in such a way that supports the pursuit of the public benefit purpose alongside, or instead of, private interests;

• creating effective transparency and feedback mechanisms that can detect and correct deviations from the purpose; and

• empowering a steward organization to successfully push back against the actions of the organization it is stewarding (when substantial power lies with its execs and staff, who can choose to leave).

There are existing governance mechanisms for mitigating each of these challenges (e.g., fiduciary duties, veto-share and steward-ownership models), but in both theory and practice they seem insufficient to prevent purpose drift in the long term if there are opportunities to gain significant profit or wield significant power. When purpose drifts, it is usually due to:

• inarticulacy in how the purpose is measured or operationalized;

• self-scoring by those responsible for pursuing the purpose;

• capture by external actors; and/or

• compulsion by the strategic environment (race dynamics and equilibrium pressures that force the hands of even purpose-aligned actors) .

Arguably a prerequisite to addressing most of the above challenges is having a good mechanism for defining and evaluating adherence to the purpose. What constitutes public benefit? How well is it being served? Before anyone can carry out a public benefit purpose (or align incentives with it, or enforce it), someone has to know what counts as adherence or deviation in the first place.

In this respect, current public interest legal structures are lacking. For example, in Delaware Public Benefit Corporations (PBCs), the board decides what it means to pursue the public benefit, and how to “balance” the public benefit purpose with the pecuniary interests of shareholders. When the PBC structure is used to attempt to robustly align organizations with public benefit purposes, this has fundamental limitations.

• The stated public benefit purposes are in most cases too vague to substantively inform or constrain the board’s discretion, or to function as an adjudicative standard against which to evaluate its decisions .

• There is a lack of guidance or mechanisms for the board on how to satisfy its fiduciary duties (e.g., it is required to balance multiple interests without guidance on what distinguishes satisfactory from unsatisfactory balancing) .

• The board is responsible for evaluating its own actions (self-scoring).

For more background on Delaware PBCs, see here.

In particular — frontier AI companies and neolabs

We are interested primarily in the corporate governance of powerful technology companies, and particularly frontier AI companies and neolabs, because they aim to create and wield enormous power, and many of them have already designed mechanisms (theoretically at least) intended to ensure that they maintain a holistic public benefit purpose. Due to the scale of their influence, it is especially important that they are robustly aligned with their public benefit purposes in the long term.

For example, both Anthropic and OpenAI are structured as Delaware PBCs with additional fortifications, including steward entities with special voting shares that determine board selection and other control rights. They both have official public benefit purposes (words to the effect of “benefiting humanity”).

For more background on the legal structures of Anthropic and OpenAI, see here.

While valuable in other ways, these fortifications arguably do not address the core limitations in how the public benefit purposes — both of which are vague, (conceptually) thick, and politically contested — are interpreted and evaluated. In both companies, it seems plausible that significant, societally consequential purpose drift could occur (e.g., via influence over economic activity, human relationships, public discourse, public health, governments, geopolitics, etc.) without any party violating a formal duty with sufficient unambiguousness to trigger meaningful corrective feedback. Plausible scenarios include:

• the company gradually optimizing for revealed preferences (usage, engagement, etc.) which drive revenue, and defining public benefit in those terms, while neglecting significant negative externalities or the many ways revealed preferences can diverge from stated preferences or welfare ;

• more generally, investor pressure/expectations and stock compensation of employees incentivizing the company to define public benefit in a manner most compatible with increasing its share price (e.g., when determining trade-offs between risks and deployment);

• internal decision-makers systematically under-weighing some risks or benefits in their decision-making because they themselves are relatively insulated from those risks, or can already access those benefits;

• government pressure, valuable defense/intelligence contracts, reduced transparency due to security clearances/classification, and a revolving door between policy teams and national security agencies leading to enmeshment with state interests and “benefiting humanity” being operationalized as “furthering the strategic interests of one country”;

• a race to the bottom among 2-4 frontier labs, in which one lab’s decision to compromise on the public interest for a competitive advantage normalizes that conduct and incentivizes other labs to follow; or

• atrophy of the steward entity over time — exercising formal powers (removing a director, etc.) is socially costly, so meaningful pushback is never practiced or normalized and, when a clear violation occurs, the steward is left with intact powers on paper but no mandate, social license or credibility with which to use them.

And with such broad, ambitious purposes that implicate the lives of billions of people, it seems:

• difficult for the PBC directors to be epistemically well-positioned to make decisions about what constitutes public benefit (including being sufficiently informed); and

• difficult for directors to be seen as legitimate decision-makers on inherently political questions about what constitutes public benefit.

Without a stronger mechanism for determining what constitutes public benefit, it seems likely that purpose will drift away from the stated public benefit purpose.

Mechanism: public evaluation of public benefit via representative deliberation

We believe that representative deliberative processes (where a randomly selected group of people deliberate together as part of a structured and facilitated decision-making process to respond to a specified remit) can, if used to evaluate the adherence of an organization to its public benefit purpose, substantially mitigate the above limitations. For brevity, we’ll refer to this approach as Public Evaluation of Public Benefit (PEPB).

For more background on representative deliberative processes and their use in other high-stakes contexts, see here.

Rationale

Substantively, PEPB helps to prevent purpose drift by:

• reducing inarticulacy in how the public benefit is operationalized, by:

  • surfacing deep/substantive information about what it means for the organization to adhere to its public benefit purpose, including in contexts with protracted conflict or difficult normative trade-offs;

  • contributing to a culture of continually refining the organization’s understanding and measures of public benefit in dialogue with high-quality public input;

• avoiding self-scoring, by:

  • outsourcing evaluation to a representative deliberative body;

• reducing the risk of capture by:

  • helping to establish an internal organizational norm of habitually and substantively engaging with deliberative public feedback;

  • attracting people who are motivated by the public benefit purpose and willing to push back against actions that deviate from it;

  • delegating evaluation to a body of people who are randomly selected, anonymous, and do not personally stand to profit or accumulate power, and thus are difficult to capture;

  • creating a mandate with high perceived legitimacy that makes it more likely that steward entities, interested shareholders, and other oversight bodies (incl. attorneys general) will push back against the operating company if it acts in a manner inconsistent with the public interest (in extremis, by removing directors or bringing derivative suits);

  • increasing the chance that legal action brought against directors or stewards in order to uphold the public benefit purpose is successful (e.g., by creating proper metrics and a paper trail of an actor stonewalling credible information about their duties to the public benefit);

• reducing the risk and impact of compulsion by:

  • raising the bar for substantive public interest governance within the industry, and so creating incentives for other organizations to adopt similar measures to compete for talent and minimize relative regulatory scrutiny; and

  • in all the above-mentioned ways, making it more likely that the organization would resist pressures to deviate from the purpose.

Legally, PEPB is a natural operationalization of provisions in the Delaware PBC statute :

• §365(a) requires that directors balance the stockholders’ pecuniary interests, “the best interests of those materially affected by the corporation’s conduct,” and the corporation’s stated public benefit purpose(s); §365(b) provides a safe harbor under which a balancing decision is deemed to satisfy directors’ fiduciary duties if it is, among other requirements, “informed.” Representative deliberative processes are a natural mechanism for eliciting deep/substantive operationalizations of what constitutes the public benefit and of the interests of those materially affected, and thus constitute evidence of informedness — strengthening the board’s position within the safe harbor.

• §365(b)'s safe harbor also requires that the decision be “not such that no person of ordinary, sound judgment would approve.” This is a deferential rationality floor that protects directors, not an affirmative standard a board must satisfy — and a deliberative body’s disagreement with the board would not, by itself, establish that the standard was violated. But a representative panel of randomly selected people is the closest available empirical proxy for the hypothetical person of ordinary, sound judgment.

Mechanism

There are different ways this could be implemented, but for the purposes of having something to critique and iterate on, here’s a concrete proposal for a PBC with a steward entity.

The deliberative process

A high-level specification for the deliberative process is described below. As the design is refined and piloted, the process could gradually be formalized into a “PEPB Standard.”

• The process would be required to be run at least annually.

• Participants in the process would be chosen via stratified random selection from a population that is as close to the relevant public as possible (e.g., if the beneficiary named in the public benefit purpose is “humanity”, the sampling frame may start with US adults for practicality, and progressively expand to the global human population as the infrastructure for doing so improves).

• The remit of the process would be to audit the PBC’s conduct over the previous year and evaluate how satisfactorily it balanced the relevant interests and adhered to its stated public benefit purpose.

• The process design would include, among other stages:

  • triage to select the most significant forms of adherence to or deviation from the public benefit purpose on which to deliberate; and

  • an adversarial design, where the deliberative body can hear from and question both company representatives and external experts / critics.

• The output of the process — the “decision shape” — would depend on how it is integrated into the legal structure (see next section). Options include:

  • recommendations (with minority reports) as to actions the PBC should take to better balance or adhere to its public benefit purpose (these could be similar in format to the outputs of Citizens’ Initiative Reviews of ballot measures that have been used in several US states);

  • findings structured to map directly onto the categories of the required §366(b) stockholder statement (objectives, standards, factual information, assessment) about the PBC’s furthering of its purpose; or

  • the completion of a standard evaluation rubric (incl. reasons) that maps to a standard set of consequences (e.g., executive compensation, actions taken by the steward entity, etc.).

• The deliberations, the outputs, and the actions taken in response would be made public.

Integration into legal structure

We see at least two potential pathways to integrating the PEPB process into the legal structure of the organization (both to require that the deliberative processes are run and to give them meaningful power).

1. As a third-party standard —

   §366(c) of the Delaware PBC statute allows a PBC to, in its Certificate of Incorporation, “require that the corporation … attain a periodic third-party certification addressing the corporation’s promotion of the public benefit or public benefits identified in the certificate of incorporation and/or the best interests of those materially affected by the corporation’s conduct.” A sufficiently mature version of PEPB, conducted with a consistent process design and quality, could be constituted as such a standard, and required in the Certificate of Incorporation.

2. As a decision-making process for the steward entity —

   While limiting board power directly is legally constrained, empowering a steward entity via special stock classes is well-precedented. The steward entity’s constitutional documents could require it to act on PEPB findings, using the process to inform how it exercises its existing oversight powers.

We believe that either of the above pathways can be used to give the process meaningful power, though there are uncertainties (e.g., no case law on what happens when a PBC fails to attain its required third-party certification), and details still to be determined (e.g., exactly what form of influence the process can have over the PBC in the case of a sufficiently positive or negative evaluation).

Viability

Precedents

Representative deliberative processes have a long history (versions were used in ancient Athens and Renaissance Italy), and are frequently run all around the world, most commonly to inform governments — e.g., Ireland, France, Ostbelgien, and the European Commission, among many others — but also increasingly in corporate settings including cooperatives (e.g., MONDRAGON), public utilities (e.g., power, water), and large tech companies (e.g., Meta, Anthropic). Over 1000 processes have been run in recent decades. Lightweight versions of this are also used within corporations for decision-making already.

There is an increasingly mature ecosystem of deliberative practitioners who are qualified to serve as third-party providers of such processes.

Progression

There are several dimensions along which this could be trialed in a low-risk way and, if successful, made progressively more substantive. These include:

• polity — start with US adults, then expand to other countries, and eventually to the global population (to the extent that laws and sampling/recruitment infrastructure support this);
• bindingness — start merely advisory, then require a response, and eventually give real power to shape the incentives the PBC is subject to;
• institutionalization — start as a (mere) experiment at the company’s discretion; later require it in relevant constitutional documents; and
• confidentiality — start with the deliberative body having access only to public information about the company, and later develop mechanisms for granting the body access to relevant confidential information.

Legal design considerations

The process design and legal integration may need to account for:

• board primacy and anti-fettering constraints;
• confidentiality and disclosure requirements for public companies (i.e., if participants in the deliberative process were given access to material non-public information); and
• laws relating to foreign influence and export controls (i.e., if some participants in the process are from other countries).

Background

Delaware PBCs

In Delaware Public Benefit Corporations (‘PBCs’), questions about what constitutes public benefit are answered as follows :

• The public benefit purpose(s) are stated in the Certificate of Incorporation.

• The board of directors:

  • has a duty to manage the corporation in a manner that “balances the pecuniary interests of the stockholders, the best interests of those materially affected by the corporation’s conduct, and the specific public benefit[s] … identified”;

  • is deemed to have satisfied this duty with respect to a balancing decision if the decision is “both informed and disinterested and not such that no person of ordinary, sound judgment would approve”; and

  • at least biennially, must provide shareholders with a report containing: (i) objectives the board has established to promote the public benefit purpose, (ii) standards it has adopted to measure progress toward those objectives, (iii) factual information, based on those standards, about the corporation’s success in meeting the objectives, and (iv) an assessment of the corporation’s success in meeting the objectives.

Essentially, the board decides what it means to pursue the public benefit. When the PBC structure is used to attempt to robustly align organizations with public benefit purposes, this has fundamental limitations.

• The stated public benefit purposes are in most cases too vague to substantively inform or constrain the board’s discretion, or to function as an adjudicative standard against which to evaluate its decisions .

• There is a lack of guidance or mechanisms for the board on how to satisfy its fiduciary duties (e.g., it is required to balance multiple interests without guidance on what distinguishes satisfactory from unsatisfactory balancing) .

• The board is responsible for evaluating its own actions (self-scoring).

  • The law allows for use of third-party standards or independent evaluations, though these have their own limitations. Some other benefit corporation statutes require them (e.g., California), but not Delaware .

Legal structure of AI companies

The PEPB mechanism is applicable to public interest organizations in general, but for reference, we summarize below the legal structures of two frontier AI companies, Anthropic and OpenAI. Both are structured as Delaware PBCs with additional fortifications. Specifically:

• both have public benefit purposes:

  • Anthropic – “responsibly develop and maintain advanced AI for the long-term benefit of humanity”

  • OpenAI – “ensure that artificial general intelligence benefits all of humanity, including by conducting and/or funding artificial intelligence research”

• both have steward entities, intended to protect the purpose, primarily via control rights over who sits on the board of the PBC:

  • Anthropic — 1,000 non-economic Class T shares held by the Long-Term Benefit Trust that let the Trust elect or unilaterally remove a majority of the board (4 of 7 seats), veto any attempt to weaken the Trust’s role or shift authority away from the board, and receive five days’ advance written notice before certain actions are taken — most distinctively, before control of any AI model exceeding board-set (quarterly-updated) capability thresholds is transferred to another party.

  • OpenAI — 1 non-economic Class N share held by the OpenAI Foundation that grants the Foundation a controlling share (66.67%) of the votes when electing new directors, the ability to unilaterally remove directors, and a veto over any amendment to the Public Benefit clause, any sale or liquidation event, and the issuance of any new stock with control rights.

• at OpenAI, there are additional requirements via an MOU with the California Attorney General, including:

  • OpenAI — The PBC board “must solely consider the [public benefit purpose] (and may not consider the pecuniary interests of stockholders or any other interest) in respect of safety and security issues”. These decisions are made by the Safety and Security Committee of the board of the OpenAI Foundation, the chair of which has full observation rights at all PBC board and committee meetings.

• both have declared that they have confidentially submitted draft S-1 forms to the SEC, in preparation for potential IPOs.