“Our central ambition is philosophy on demand: philosophical judgment of exceptional depth and quality, available quickly enough to inform consequential decisions.”
That’s Graham Clay, writing about Pensatura, an AI philosophy system he and Caleb Ontiveros are developing. Pensatura “coordinates AI tasks that investigate sources, develop arguments, examine objections and construct an overall answer [to a philosophical question], retaining the resulting work so later inquiries can use, challenge and extend it. Each argument in the record states its premises, taken jointly, and an explicit claim about how they support its conclusion. That support claim can itself have grounds and face objections. Interpretations preserve their sources and rival readings; improvements are distinguished from what an author actually argued. The intended result is an expanding body of reasoning whose conclusions remain connected to the grounds on which they depend.”
They report that they’ve built a “working prototype” of Pensatura that includes the foregoing, makes use of Lean to check formalized arguments, and includes “machinery for probabilistic assessment.” At the moment, its scope is philosophical questions about AI research, but they have plans to expand into other areas.
Clay and Ontiveros ultimately see Pensatura as a way of making philosophical knowledge more readily available to decision-makers whose work could be improved by it. It might also make possible the kind of computer-aided philosophy I speculated about here.
I will note that I am ambivalent about this kind of project. I have doubts about the answer-based conception of philosophy at its core, concerns about the drive it embodies towards what we might call “philosophical efficiency,” and worries about what such technology might one day replace. Nonetheless, it is an interesting experiment, and if it comes close to working, it will be a fascinating, impressive, and philosophically important piece of technology.
In the following guest post (written with assistance from AI), Clay provides more details on the project and its motivation, and solicits help from other philosophers.
by Graham Clay
In 2021, I and my coauthor Caleb Ontiveros argued here at Daily Nous that philosophers should help shape how the AI revolution shapes philosophy itself. Among other things, we proposed that “[d]espite the great promise of AI, we maintain that unless philosophers theorize about and help develop philosophy-specific AI, it is likely that AI will not be as philosophically useful.”
Our 2023 Metaphilosophy article, which we found to be a lot easier to publish after ChatGPT was released, developed the case further. In it, we offered a more detailed argument for the conclusion that AI could transform philosophy on a scale comparable to writing, including in how research becomes cumulative and systematic—but philosophers needed to get involved.
However, in both pieces, we deliberately set aside the possibility of systems capable of carrying out substantial core philosophical work themselves; we were focused on AI that acted in a support capacity to human philosophers.
To be clear, we have always thought that this possibility would deserve, at some point in time, a serious research program. Yet we thought that point in time was farther into the future, despite being (very) bullish on timelines.
Now I think we were not bullish enough.
Part of why I think this is that, after shifting our careers to industry (AI and software), we have built the prototype of this sort of system, Pensatura. It is more impressive than I ever imagined in 2019 when I first got research access to GPT-2. I will explain broadly how it works farther below.
Given the scale, complexity, and expense of Pensatura, its development and operation require significant brainpower and compute. Our usage is already in the vicinity of 2 billion tokens a day, and I expect this to ramp to 100s of billions by the end of Q1 of 2027, even as we improve its efficiency rapidly. To this end, we are looking to collaborate with the brightest minds in philosophy and the AI industry, and we are receiving and making inquiries about funding, both academic and commercial.
I’m here today to make an appeal: we want philosophers to help formulate its aims, from the perspective of quality, and investigate the conditions under which it and other AI systems would be deemed successes with respect to the philosophy they produce.
But first let me explain why I think Pensatura, or a system like it, is needed. This requires explaining what distinguishes such a system and what would make it possible in its ideal form.
Our central ambition is philosophy on demand: philosophical judgment of exceptional depth and quality, available quickly enough to inform consequential decisions. Consider an AI laboratory deciding how much authority to give a research system that helped design the evaluations used to assess it. How much a passing result counts depends on how those evaluations were constructed. What that evidence permits depends, in turn, on further arguments about risk, responsibility and justified delegation, with substantive ethical and epistemological premises.
There is already demand for research on and reliable answers to questions of this kind. Anthropic has established a research program examining whether and how AI systems might deserve moral consideration. The UK’s Advanced Research and Invention Agency (ARIA) has solicited philosophical and strategic insights to inform decisions about Safeguarded AI systems. These initiatives are a few of many that give philosophical inquiry a direct role in consequential decisions, including about AI itself. Pensatura’s ambition is to make inquiry of the required depth and quality available across many more such decisions, including eventually in policymaking, medicine, and other fields.
In all cases, the target is an extremely well-justified account: one that commands the strongest relevant scholarship, develops further arguments where needed, confronts serious alternatives, and brings their combined force to bear on the decision. It should establish an “applied” answer at the strength and scope its grounds, including highly theoretical ones, warrant. We want a standard demanding enough to challenge the best philosophers working on the question, and ultimately a depth, breadth, and speed of application far beyond what any human or team of them could sustain.
As implied above, ethical, epistemological and other philosophical questions about AI research autonomy are our first domain of application, as they are at the intersection of importance and funding. But the engine’s ambition is general. When consequential decisions depend on difficult questions about evidence, reasons, obligations or value, we want excellent philosophical inquiry to be available quickly enough, and widely enough, to inform them.
The Philosophical Ideal
Why treat the capacities above as one target? The following argument makes the commitments explicit:
- The ultimate epistemic aim of inquiry into a philosophical question is to provide the most well-justified answer to it (i.e. a proposition that answers it, plus its justification).
- To achieve this aim, one must engage in faithful interpretation of relevant scholarship, further reasoning upon its basis and other evidence, continuations of inquiries in which relevant considerations are added (i.e. dynamism in response to new evidence), and engagement with new concrete cases (i.e. applications of one’s positions, including theoretical ones).
- For each of these activities, one must correctly assess and act on one’s assessment of how its outputs/results bear on, or are affected by, the outputs/results of all others that are relevant to it.
Therefore, an ideal realization of the ideal of philosophical inquiry requires interpretation, reasoning, application and reassessment, successfully coordinated across a range of results in a cumulative manner.
Put in a historical key, ideal philosophical inquiry examines “received” accounts, works through the difficulties they raise, pursues independent reflections and lines of argumentation, and seeks the best-grounded account the subject permits, progressively coordinating and reassessing its principles, consequences, and applications in light of relevant evidence and further inquiry.
This argument is conditional on an account of inquiry and its idealization, both offered for scrutiny (more on that below). Its value is to make the high bar intelligible: an answer can fail through a missed argument, a misreading, an unsupported connection, a mistaken application, or an inadequate response to criticism.
Useful and excellent philosophy can fall short of the ideal, but such contributions are useful and excellent in part because of their contribution to achieving it. The aspiration is to bring those contributions together more deeply and broadly than any individual could. The achievement can reside in an expanding body of philosophical work shared across people and systems.
Building Philosophy on Demand
In his recent discussion of how philosophical achievement might be recognized, Justin Weinberg suggested that AI’s distinctive contributions may appear in dynamic inference-mapping applications. We agree. In fact, Caleb and I have been working towards this ideal for quite a few years, and Pensatura aims to conduct the inquiry such a representation would record. It coordinates AI tasks that investigate sources, develop arguments, examine objections and construct an overall answer, retaining the resulting work so later inquiries can use, challenge and extend it.
Each argument in the record states its premises, taken jointly, and an explicit claim about how they support its conclusion. That support claim can itself have grounds and face objections. Interpretations preserve their sources and rival readings; improvements are distinguished from what an author actually argued. The intended result is an expanding body of reasoning whose conclusions remain connected to the grounds on which they depend. Our working prototype includes this record and orchestration, Lean checking of consequences under explicit formalizations, and machinery for probabilistic assessment.
Suppose further research reveals a weakness in the aforementioned AI laboratory’s evaluations. The system should identify which proposed permissions relied on those results, investigate possible repairs, and preserve judgments that retain independent support. Later inquiries should be able to build on that work while concentrating effort on new facts, objections and unresolved questions. This is also the economic promise of cumulative inquiry: substantial research could serve many subsequent applications.
As noted before, the first version is now built, and our current work is review, evaluation, and improvement. Its philosophical success remains to be demonstrated (more on evaluation and benchmarking below).
Evaluating Philosophical Inquiry
How should we tell whether Pensatura, or another system, achieves the ideal we seek? Serious efforts to evaluate AI’s philosophical abilities are already underway. PhilosophyBench seeks sophisticated original philosophy, including work approaching the quality of the best historical achievements, and investigates how AI affects human philosophical work. Other evaluations examine important components, including conceptual reasoning in the Conceptual Reasoning Index and substantive moral reasoning in MoReBench. The AI Philosophy Competition invites primarily AI-written essays, permitting research, criticism, revision and disclosed human assistance.
As implied already, excellent work in these settings can provide substantial evidence of philosophical ability, just as a strong essay can demonstrate interpretation, integration, originality and application together. (I hope my dissertation did!)
However, we differ from these approaches in four ways (note: not all of these divergences apply to all other approaches):
(i) we are not indexed on traditional output formats like essays because we doubt their utility for the applications we think are most impactful;
(ii) relatedly, we are focused centrally on what a philosophical account makes possible over continued inquiry (i.e., can the system respond appropriately when its grounds are challenged, preserve what remains supported, and use the resulting understanding to answer unfamiliar questions);
(iii) also relatedly, we are pursuing commercial applications; and
(iv) we want our benchmark to assess the full achievement described above across philosophical subfields.
A fear (or hope) of everyone building anything specialized right now is the possibility that a general-purpose LLM, agent, or harness will “eat its lunch” within a short timeline. We don’t want to run from this fear. It’s crucial, in our view, that any benchmark we create should enable us to compare any AI system, whether Claude Opus 5.5, Pensatura, or GPT-2. Indeed, Simon Goldstein’s account of producing a largely AI-written philosophy article illustrates why organized workflows are serious alternatives.
If general-purpose systems achieve the philosophical standard we are aiming for, that would itself be a major success, regardless of which system wins (hence ‘hope’). Either way, Pensatura would be one contender in a benchmark valuable across competing approaches, and our aim is to have the highest score possible while bringing outputs produced by the system to as many applications as possible.
How Our Proposed Benchmark Will Work, More Specifically
So, how will the benchmark work? As the above implies, we are developing a benchmark around complete philosophical inquiries and what their results enable. Each case would begin with a question, relevant context and explicit requirements for the answer. The entrant would construct an account, encounter a previously concealed objection or development, reassess its position, and apply the resulting understanding to a new case. Expert reference packages would contain outstanding accounts and responses across that sequence. Each stage’s actual submission would be retained, allowing us to distinguish grounds already present from valuable reasoning developed later.
Reference work and submissions use a common format that makes conclusions, jointly supporting premises, further grounds, objections and source interpretations explicit, while preserving the natural-language explanation and context their assessment requires. Entrants would remain free to choose their internal architectures. Formal and Bayesian rigor would be central: does the representation capture the intended claims, does the result follow under its assumptions (whether entailment or probabilistic), are those assumptions well grounded, and what does the result establish for the question actually asked?
Building the references is itself substantial philosophical work. Translating outstanding scholarship requires distinguishing the author’s argument from an interpretation or improvement; developing a sufficiently complete account may require new arguments and further research. The benchmark should reward original contributions, important distinctions, discoveries about how arguments fit together, and the quality of the applications they enable. It should admit different defensible answers and warranted suspension. A better answer must be able to beat the reference.
We want an overall score of philosophical excellence, accompanied by component results, coverage, cost and time. However, we are beginning with philosophical questions about AI research autonomy and will expand into other fields as suitable experts join.
Because Pensatura would compete relative to the benchmark, its standards must be defensible independently of Pensatura. We want philosophers—including critics of our approach—to help determine what the common format must preserve and whether the assessment recognizes the strongest work. Reference construction could draw on adversarial collaboration, bringing philosophers who disagree into the work of developing and challenging accounts. The ambition is a test on which even an extraordinarily accomplished philosopher would have to work maximally hard for a long time to receive a satisfactory score. We are seeking academic partners to help lead the benchmark’s development, and dedicated funding for that research program.
Why Go Commercial?
David Strohmaier recently argued here that AI could make philosophy capital-intensive, partly through its ability to investigate connections across debates at a scale unavailable to individual researchers. We want to build the organization that makes this possibility productive: namely, a company with substantial philosophical research capacity, rapid engineering and evaluation, and the ability to deliver the resulting work wherever consequential questions arise.
That requires substantial, sustained funding. Capital supports a core philosophical team, specialist research, engineering, compute, demanding comparisons and reliable interfaces for users. It would let us investigate more competing arguments, pursue more inquiries in parallel, build references across more fields, and shorten the cycle between identifying a weakness and testing an improvement. The benchmark should help establish when that additional effort produces better philosophy.
The commercial routes connect directly to this work. For model developers’ evaluation and alignment teams, we aim to provide demanding evaluation datasets, training resources and assessment services. Philosophical judgments already enter AI development: Anthropic describes Claude’s constitution, which specifies intended values and behavior, as central to its training process. Better philosophical research and evaluation could improve both the judgments systems are trained to make and the standards used to assess them.
Organizations facing ethical and institutional decisions could commission analyses and obtain continuing access to the reasoning behind them: grounds they can inspect, objections they can raise, and applications they can reconsider when facts or assumptions change. A substantial investment in understanding a domain could support many later inquiries while each new case receives the further work it requires. We want a business that funds continuing philosophical research and brings its achievements into far more decisions.
We still have to show excellent philosophical performance, dependable service and value to customers. If we do, that would support an enterprise much larger than any individual research project.
An Invitation
There are substantial risks to this endeavor: persuasive bad arguments, accelerated error, conformity, unwarranted deference, and assistance to harmful projects. Visible reasoning may itself invite excessive trust. Critical involvement must include studying these failures and identifying where reliance is unwarranted. We also need to assess the environmental costs of scaling and its effects on access, research priorities and human philosophical expertise.
The work we need from philosophers is substantial philosophy that contributes to our evaluation and benchmark development. Develop a stronger argument, a better interpretation, a distinction that changes what follows in an important case, or an account of how several arguments contribute jointly to a conclusion. Help determine what warrants a probability assigned to a philosophical premise, or how an inquiry should respond when an interpretation is corrected. These contributions could shape the engine’s methods, becoming demanding reference work, and inform the inquiries delivered to users. We welcome sustained collaborators, potential cofounders, research partnerships and, as funding allows, paid specialist contributions. The work should advance contributors’ own research as well as the system.
We also welcome arguments against this conception of philosophical inquiry and against developing or deploying systems intended to realize it.
Write to me at [email protected] with a question, argument, difficult application or proposal for collaboration. I will also engage with comments and questions here when I can.