Our Principles

In every organization there are three fundamental questions to answer: How are resources procured? How are resources allocated? Who makes the decisions?

How to procure resources: I propose to structure the legal entity for this auto-grader as for-profit to remain competitive. OpenAI, the maker of ChatGPT, started out as non-profit. To vastly oversimplify, OpenAI learned that this structure was not adequate to obtain and generate the financial resources necessary for continued success. The transition was messy, politically fraught, and waged in public. So to start, I propose we skip that chapter and launch as for-profit in order to sustainably obtain the resources needed for continued success.

How to allocate resources: I elaborate more throughout these principles on how the money should flow, but in a nutshell I intend to give any profit back to impacted stakeholders.

Who makes the decisions: Unlike OpenAI, however, I want to structure the organization through the vehicle of a stewardship trust, a legal structure that mandates the enforceable consideration of other principles besides a fiduciary duty to pure profit in corporate decision-making. While I am still in the process of setting up a webpage explaining the details for how it would work here, suffice it to say that there would be elements of governance shared between the for-profit company and the stakeholders involved. While not a perfect analogy, the clothing company Patagonia has a corporate structure centered around a similar purpose trust which legally binds the company to the considerations of environmental stakeholders in its corporate decision-making.

(Full disclosure: Right now I have CompGrader housed under The Firm, LLC, which is the LLC owned 100% by me that I use to hold all my side projects where there's the prospect of money changing hands. But I suspect this is a temporary placeholder.)

It's one thing to say we're doing good things. Many tech companies start out this way. But it's another thing entirely to continue. As such, I propose this for-profit structure seek accreditation as a certified benefit company, under a well-known accreditor/certifier such as B Lab.

While the specifics may vary by individual accreditors, I propose meeting independently set and verified standards on the topics of “governance, workers' rights, community impact, environmental impact and stewardship of its customers” (as the Wikipedia page for B Lab frames it).

Individuals, organizations, industries, governments, and whole societies need both stability and flexible adaptability to be successful. This tension is often brought to bear in whether or how a particular organization or industry is regulated. I want to work constructively in this area, in a spirit of ongoing conversation, with stakeholders who believe more regulation is appropriate.

With that in mind, I submit that “peer review” rather than government mandate is the most appropriate way for AI graders to be regulated, if they are to be regulated at all. Whether that peer review looks like the existing accreditation system familiar to higher education, the self-enforcing regulatory body of finance known as FINRA, or the ABA for lawyers is still an open question. But given the particular importance of academic freedom involved with AI graders, I'd prefer any regulation in this space to happen by peers rather than by government.

As of now, my software uses open weight models. While I am open to allowing changes to this model if the demand is there to do so, my preference is to use models whose source and weights are open for inspection and, ideally, subsequent tuning.

A benefit of this approach is that student data is not sent to third-party tech titans. The software is fully contained and, with appropriate infrastructure, can remain on a university's campus servers or servers managed by CompGrader, not sent raw to a third party.

Academic freedom is a special concern of the First Amendment. While CompGrader is not the government, ensuring individual instructors have control over how this grader is used is a key focus of mine.

Also key to this conversation is the idea of choice, and allowing faculty to choose. The choices others may make might not be the ones I or you agree with, but that is a cherished point of academic freedom.

A benefit of using open-weight models (See “Openness and Privacy”) is that I've been able to design a product that has hot-swappable models built in. This means that a faculty member, institution, or disciplinary community could train their own model to enforce their own standards of fairness and accountability. I do not want to create the false choice of “grade everything by hand” or “use only the AI model chosen for you.”

There are traditional concepts of prudent supply chain management that apply here (for example, where the code will run, whether the energy that powers it is ecological, who makes the servers, whether the workers are paid fairly, and so forth). Again, I point to existing frameworks for independent assessment and accreditation (such as B Lab certification) for these concerns.

But there are new chains of slavery and dependence on my mind. What happens to student workers, for example? And I address those primarily in the economics section below.

What concerns me most for the near future, while we are not there yet, is that at some point AI will become sentient and, crucially, be able to feel pain. It is this threshold, the ability to feel pain, which I take as the next major litmus test for humanity as we engage in the complicated question of, “What is the right thing to do?” In particular, we must consider anew what it means to bind sentient entities (human or otherwise) or to “[break] the chains of new forms of slavery,” to use the language of The Pope in his recent encyclical letter on AI. I have no definitive answers for these questions, but I can think of no better place to attempt to answer them than in partnership with faculty at a university.

I am no Luddite and I am an AI enthusiast, conditioned on the benefits of AI being widely shared and the displacement of workers mitigated.

In terms of economic mitigation, I plan to hire primarily students to help develop, maintain, provide support, etc. for the grader. While professional staff will be necessary, hiring student workers is a key idea on my mind.

I am no Luddite and I am an AI enthusiast, conditioned on the benefits of AI being widely shared and the displacement of workers mitigated.

As such, I pledge to give as much money as I can to organizations whose missions center undergraduate and graduate employee workers, and for scholarships earmarked for the type of student who would have otherwise found some financial security through grading work yet has lost it due to reliance on a cheaper AI grader.

While I propose to structure this as a private company, financial transparency is crucial. Although I am unsure of the exact specifics, I pledge to keep the accounting books open and will provide as much financial transparency as I can without causing harm. This will likely look like periodic disclosures similar to those made by publicly traded companies (e.g., the 10-K report to the SEC) or a non-profit's tax return (e.g., the 990 tax form).