AI as an Interlocutor in Philosophical Work

Philosophers often develop ideas through dialogue. A conversation partner can ask for a reason, question an assumption, offer a counterexample, or show that a conclusion reaches beyond its support. Language models can now participate in parts of that exchange. They can generate objections, compare positions, summarize a dispute, suggest sources, and respond to repeated questions. Some professional philosophers have begun testing what those abilities can contribute to their own work.

How This Chapter Fits the Course

This chapter brings the first three chapters together. Chapter 1 showed how a visible judgment makes reasons and revisions available for scrutiny. Chapter 2 supplied a method for building and pressure-testing arguments. Chapter 3 explained why fluent AI output still requires evaluation. Here you will apply those foundations to the three roles AI can play in PHIL 123: philosophical practice partner, research mapper, and project and build assistant. ABC-R provides a common method for evaluating AI contributions while keeping the final intellectual and ethical decisions in your hands.

Philosophy Is Already Working With AI

David Chalmers, for example, describes extended conversations with Claude and ChatGPT while exploring philosophical questions and possible new ideas. A research team led by Ye and colleagues interviewed 21 professional philosophers about language models as critical-thinking tools. The study organized their possible uses into three interaction functions. A model might act as an interlocutor that enters a dialogue, a monitor that watches for gaps or errors, or a respondent that reacts to a philosopher’s work. PHIL 123 uses a different three-role structure based on the stages of the course. There are limits, of course. Current models can lose track of subtle positions, supply shallow criticism, or produce fluent answers that the philosopher still has to assess.

Philosophers also work inside companies building these systems. Amanda Askell is a philosopher and research scientist at Anthropic, while Iason Gabriel describes his work as philosopher and research scientist at Google DeepMind. Their roles show that philosophical judgment enters AI development as well as outside criticism of it.

Language models display enough local skill with moral language to make the encounter intellectually useful. Seth Lazar reports that frontier models can often recognize morally relevant details and connect them with reasons. They can also construct plausible arguments in text-based cases. But again, their performance is uneven. A model may reason well about one vignette and poorly about a nearby case, or it may produce an objection without recognizing how serious it is, or defend incompatible positions in separate conversations.

These tools have entered philosophical practice because they can contribute material that deserves examination, but their usefulness depends on a person who understands the relevant concepts and knows what the task requires. That person must also decide whether a contribution deserves acceptance. PHIL 123 asks students to develop that ability. AI will participate in some of the course’s philosophical practice, research, and project development.

Three Roles for AI in PHIL 123

Philosophical Practice Partner

In the first half of the course, students use AI as a philosophical practice partner. Students use these interactions to practice argumentation, Divine Command Theory, natural law, virtue ethics, Kantian deontology, and utilitarianism. A course chatbot may ask which premise carries the moral principle, offer a counterexample, question an inference, or present an objection from within a framework. The exchange creates another opportunity to practice before the student completes the major comparison and judgment work.

The student remains responsible for the interaction. If the AI tutor offers a serious objection, the student identifies the premise under pressure and decides whether to defend it, revise it, narrow the conclusion, or concede part of the argument. The chatbot can supply pressure, but the student will be forced to judge and then steer the conversation accordingly.

Research Mapper

In the second half of the course, AI can serve as a research mapper. Students choose a field, profession, technology, or social domain connected with their final project. An AI research tool can help them identify recurring claims, disputed questions, specialized vocabulary, and source leads. It can also reveal possible stakeholders and several routes into a topic before the student knows enough to search with precision.

A map is an orientation device. It may contain useful routes and dead ends. Missing regions or distorted labels can also change how a student understands the topic. The tool might fabricate a source title or attach a claim to a real article that does not support it. Its report may emphasize companies and engineers while overlooking people affected by the technology. A broad national answer can also miss the conditions of a particular profession in Idaho.

Students therefore open and check at least one important source, compare the map with the field’s actual context, and inspect the framing. They use philosophical frameworks to convert information into questions about welfare, rights, duties, and character. Other questions may concern responsibility, consent, justice, or human capacities. The AI can help show where a debate might be. The student decides which parts of that map can support further inquiry.

Project and Build Assistant

During final-project development, AI may become a project and build assistant. Depending on the assignment and the project format, it can compare approaches, organize material, or test language. It may also generate design options, draft or revise permitted components, troubleshoot code, or provide feedback on a prototype. Some assignments permit AI to contribute to major project components.

The student’s responsibility appears in the design trail. When AI compares a podcast, website, and policy guide, the student should explain why one format fits the intended audience and problem. Generated code needs to be understood and tested before it enters the project. Language revisions require another judgment about whether the new wording preserves the project’s philosophical position. The final product should show choices that the student can reconstruct and defend.

Exact permissions vary by assignment. Canvas identifies when AI use is required, permitted, limited, or prohibited, along with the process evidence students must preserve. Across all three roles, a stable obligation remains. Students inspect the contribution, decide what to use or change, preserve the required evidence, and explain the judgment behind the decision.

Each role changes the emphasis of the evaluation. Framework practice centers on concepts and arguments. Research mapping turns toward sources, field knowledge, and omitted perspectives. Project development requires students to defend audience and design choices while checking the production itself. ABC-R travels across all three tasks without treating them as interchangeable.

ABC-R: Check the Contribution, Then Check the Use

The previous chapter explained why a language model can produce a fluent answer that still contains false claims, distorted sources, shallow reasoning, or a response shaped by the user’s framing. PHIL 123 uses ABC-R to evaluate those contributions. ABC-R is a course method assembled from AI risk guidance, academic-integrity principles, and the practices students will use throughout the semester.

The first three letters evaluate the contribution itself.

• A: Accuracy and Evidence asks whether claims, sources, quotations, framework interpretations, and argument moves are supportable.
• B: Bias and Perspective asks which assumptions and viewpoints shape the response, whose interests appear, and what the interaction may have pushed aside.
• C: Context and Relevance asks whether the contribution fits the course concept, actual dilemma, field, audience, source requirement, and task.

The final letter evaluates the student’s use.

• R: Responsible Use asks whether obtaining and using the contribution this way follows the assignment, preserves the ability the student must develop, respects integrity and privacy, and remains ethically explainable.

A contribution can pass one letter and fail another. It can also pass A, B, and C while failing R. ABC-R ends in a decision about the contribution: use it, revise it, reject it, verify it, ignore it, or investigate it further.

A: Accuracy and Evidence

Accuracy in philosophy includes more than checking names and dates. Students also need to check whether a source supports the sentence attached to it, whether a framework is represented fairly, and whether an objection reaches the argument it claims to challenge. A response can use correct vocabulary while connecting the concepts badly.

PHIL 123’s custom chatbots receive the assigned textbook chapters and course instructions. That shared reference point lets students compare the chatbot’s explanation directly with the course source.

Students who use ChatGPT, Copilot, Claude, Gemini, or another general-purpose platform for additional practice should check its explanations especially carefully against the assigned chapter. Philosophical traditions contain internal disagreements, and a defensible interpretation may still be the wrong one for the concept or version taught in this course. A chatbot discussing utilitarianism, for example, might shift among act, rule, and two-level approaches without explaining the change. For PHIL 123 assignments, the assigned textbook chapter is the source of truth for the framework.

Later in the course, source checks follow the same pattern. An internet link establishes only that a page exists. When using AI to help research, you’ll still need to open the page, locate the relevant passage, and compare it with the generated claim. When the task depends on research, ask whether the source is appropriate for the field and current enough for the claim. The NIST Generative AI Profile treats confabulation and information integrity as risks requiring validation and attention to provenance. For students, that means tracing a claim back to evidence they can inspect.

B: Bias and Perspective

Bias includes stereotypes and unfair patterns, but philosophical work requires a wider perspective check. A response may treat one stakeholder as central, accept the user’s description of the problem, supply an easy objection, or apply one ethical lens as if it settled the case. It may mirror the conclusion the user appears to prefer. Because a chatbot responds inside an ongoing conversation, the interaction itself can strengthen the frame.

Consider a student researching AI documentation tools in nursing. The initial research map emphasizes faster charting, reduced administrative burden, and hospital efficiency. Those are relevant concerns. The map says little about nurses who must correct generated notes, patients whose conversations become data, clinicians who may rely on inaccurate summaries, or hospitals that purchase and govern the system. It also treats time saved as the main measure of success.

The student should not discard the map simply because it has a frame. Every inquiry begins from some selection of questions, and the B check makes that selection visible. Several search terms and source leads can remain. The student then adds nurses and patients to the stakeholder map and investigates how documentation errors are detected and assigned. The revised map can support questions from utilitarianism, Kantian ethics, and virtue ethics while treating hospital efficiency as one consideration among the others.

The UNESCO Recommendation on the Ethics of Artificial Intelligence directs attention to affected groups, representation, human oversight, and the distribution of benefits and harms. Those concerns fit research mapping, but B also applies to a short tutor exchange. Consider strategies like asking whether the objection is serious, whether the chatbot has followed your preferred conclusion, and whether another framework would direct attention elsewhere.

C: Context and Relevance

A statement can be accurate in general and poorly fitted to the work in front of you. Context includes the assigned framework, the details of the dilemma, the field, the audience, and the stage of the project. Relevance asks whether the contribution helps answer the actual question.

Suppose a student plans a final project for small-business owners who are considering AI hiring tools. A chatbot recommends an interactive website that lets users upload applicant data and receive an automated risk score. The idea is technically connected with the topic. It also reproduces the kind of automated judgment the project is supposed to examine, requires access to sensitive applicant information, and demands technical work the student cannot adequately test within the course. A simpler decision guide could help the same audience compare vendor claims, identify affected stakeholders, and ask about appeal procedures without processing anyone’s data.

The student rejects the scoring tool and revises the website idea into a guide. This verdict applies to the current proposal because it fails the project’s audience, privacy, testing, and philosophical purposes. C keeps the contribution tied to those particulars without judging every possible scoring prototype.

The same check applies when a chatbot gives a textbook definition that doesn’t match the assigned chapter, a national statistic that doesn’t answer an Idaho question, or a polished paragraph that exceeds the student’s own understanding. The contribution may contain useful information, but it needs to fit the task.

R: Responsible Use

A, B, and C evaluate whether the contribution is supportable, appropriately framed, and fitted to the task. R turns to the defensibility of the interaction. Questions about academic integrity fall under R.

The first question comes from the assignment. In this course, we’ll adopt the AI Assessment Scale. It gives instructors a way to state different levels of AI permission for different tasks. PHIL 123 uses those boundaries at the level of an assignment or checkpoint. AI may be required for philosophical practice, permitted for research mapping, allowed for selected project work, and prohibited for a judgment trace or peer reply. A contribution that would be useful in one stage may be inappropriate in another.

R also asks what ability the student is supposed to develop or demonstrate. A chatbot may produce a sound Kantian objection. If the assignment asks the student to formulate that objection independently, submitting the generated version bypasses the target skill. If the assignment asks the student to evaluate an objection supplied by a chatbot, the same output becomes appropriate material for analysis. The words have not changed. The educational use has.

Privacy, source rights, and intellectual responsibility also enter here. A student should not place confidential workplace records, identifying information, or private messages into an outside system without permission. A research lead should be followed to the original source so the model’s summary does not erase authorship or distort the claim. If AI influences a submitted argument or design, the student should preserve the evidence required by the assignment and explain the influence accurately.

UNESCO’s guidance for generative AI in education and research emphasizes human agency, privacy, pedagogical fit, and accountable use. The International Center for Academic Integrity grounds academic work in honesty, trust, fairness, respect, responsibility, and courage. R brings those considerations into one course question:

Even if the contribution passes A, B, and C, was obtaining and using it this way permitted, educationally defensible, and ethically accountable?

Suppose an AI system produces an accurate, well-sourced, balanced explanation of a student’s dilemma. The student submits that explanation as a no-AI judgment trace. Its quality does not make that use appropriate. The stage was designed to show what the student could explain and decide after assistance. The student has also concealed the source of the reasoning, so the submission cannot serve as evidence of independent judgment.

If an AI contribution is accurate, balanced, and relevant, why should the path used to obtain it affect the philosophical quality of the final work? A sound argument remains sound regardless of who first formulated it. Requiring process records can look like policing behavior after the intellectual problem has already been solved.

The objection correctly separates an argument’s logical quality from its origin. A generated argument does not become invalid because a model produced it. Yet a course assignment evaluates more than the existence of a good argument. It evaluates whether a student can recognize the governing principle, connect it with the case, respond to an objection, and defend a judgment. An answer key can contain every correct answer while supplying no evidence that the learner can solve the problems. A polished AI contribution creates the same gap when it performs the ability under assessment.

Permission and accountability create further reasons to inspect the path. A student may use an accurate contribution outside the stated boundary, conceal a major source of influence, expose protected information, or accept reasoning they cannot explain. A, B, and C evaluate the contribution’s quality. They cannot tell us whether the student used it honestly, protected the relevant people and sources, or retained the skill the task was meant to develop. R handles those questions.

Responsible use does not require refusing every substantial AI contribution. It requires a defensible relationship between the assistance and the task. A practice partner can supply pressure for the student to evaluate and answer. A checked research map can orient further inquiry. Project assistance remains defensible when the student directs the work and tests the result, then documents consequential choices. The boundary changes with the role, while every role still requires an account of the student’s judgment.

Use AI to Learn, Not Merely to Finish

An AI-assisted product can look better than the learner’s independent work. That difference can be useful. Feedback, examples, and explanations often help people perform tasks they could not yet complete alone. The educational question concerns what the learner can understand, retain, and do after the assistance.

Recent studies have tested parts of that question in different settings. In a Turkish high-school mathematics experiment involving nearly 1,000 students, Hamsa Bastani and colleagues found that unrestricted GPT access improved performance during practice. On a later unaided exam, that group performed 17 percent below the no-AI control. A more structured tutoring interface avoided the penalty by limiting direct answers and preserving more student effort.

A randomized study of 117 university students by Yizhou Fan and colleagues found that ChatGPT-supported students produced stronger essays during the assisted task without a corresponding significant advantage in knowledge or transfer. In another setting, Greg Kestin and colleagues tested a purpose-built AI tutor with 194 Harvard physics students. The tutor used instructor materials, a structured sequence, feedback, and self-pacing. Students showed strong immediate learning gains, although the study did not measure retention or transfer and cannot establish the effects of ordinary chatbots in philosophy.

The studies involve different populations, subjects, tools, and measures. Together they support a bounded conclusion. Product quality during assistance does not by itself establish independent learning. The design of the interaction influences what the learner still has to do.

Imagine a student comparing Kantian ethics with utilitarianism. The chatbot produces a polished explanation, identifies the strongest argument from each framework, and recommends a qualified conclusion. The response may earn high marks if judged only as a finished product. Close the chat, however, and ask the student to explain why Kant’s Formula of Humanity differs from an impartial calculation of welfare. If the student cannot reconstruct that difference, identify the assumption behind the chatbot’s recommendation, or explain why the conclusion was qualified, the product has outrun the student’s present ability.

The gap identifies what the student still needs to learn. They can return to the framework chapters, rebuild the comparison in simpler language, and test one premise at a time. Treating the chatbot’s polished response as an object of analysis lets the student mark its claims and inferences directly. Independent explanation then shows whether the student can perform the relevant philosophical moves without leaning on the generated paragraph.

The same problem appears in moral deliberation. Elizabeth O’Neill, Michal Klincewicz, and Michiel Kemmer argue that an artificial ethics assistant could support self-control, case-level judgment, or clarification of values. It could also collect sensitive moral information, import outside influence, and encourage a person to adopt conclusions without adequate reasons. The philosophical issue concerns the formation of judgment as well as the correctness of one recommendation.

Brendan McCord develops that concern through Adam Smith’s image of the impartial spectator. A person develops practical judgment through repeated attempts to see a situation from perspectives beyond immediate self-interest. An artificial spectator might supply information or objections that help that effort. If it supplies the conclusion and the person simply adopts it, the interaction can bypass the practice through which practical wisdom develops. McCord provides a philosophical argument about moral formation. Experimental studies would be needed to establish how a particular classroom interaction affects learning.

PHIL 123 responds by separating assistance from the abilities students still need to show. Framework practice requires students to check the chatbot against the chapter and rebuild important reasoning themselves. Research includes opening a source and judging whether the AI map represents the field. Project documentation explains how philosophical frameworks and later movements affected the design. Some stages require a no-AI judgment trace so the student can state what they decided after the assistance ended.

These records provide inspectable evidence of current performance. They cannot establish permanent mastery. A student who can explain why an objection reaches P2, locate the passage behind a source claim, or defend a design decision has shown more than a completed interaction. When the student cannot reconstruct the reasoning, the missing explanation identifies the next skill to practice.

Metacognition helps students use that evidence. The term refers here to monitoring one’s own thinking and making decisions about what to do next. During an AI interaction, a student might notice that a counterexample changed their confidence in one premise but left the conclusion intact. Another student might discover that they accepted a source summary because it agreed with an existing belief. Recording the change makes the reasoning available for review. The student can then decide whether to verify a source, revise a premise, seek another perspective, or practice the framework again without assistance.

The record also guards against a common illusion created by fluent output. Reading a clear explanation can produce a feeling of recognition. Recognition is weaker than being able to retrieve, explain, and apply the idea in a new case. A short no-AI reconstruction, source check, or judgment trace tests whether the student can cross that gap. The result may confirm learning or reveal uncertainty; either result gives the student a more accurate picture of what they can currently do.

A useful decision trail is short and specific. AI helped me brainstorm and I edited the result hides the judgment. A clearer record identifies the missing stakeholder that caused a research map to be revised, the inaccurate framework claim that was rejected, or the audience constraint that determined a project design. The record should let another person see the point where the student’s judgment entered the work.

In PHIL 123, students may use AI to encounter objections, compare frameworks, explore unfamiliar fields, and develop final projects. These activities support learning when students know enough to test the contribution and remain responsible for its use. Across the course’s three AI roles, ABC-R gives students a repeatable way to decide why a contribution belongs in their philosophical work, which parts need revision, and which parts should be left out.

References

• Askell, Amanda. Curriculum Vitae. Used to verify Askell’s philosophical training and current Anthropic role.
• Bastani, Hamsa, Osbert Bastani, Alp Sungu, Haosen Ge, Ozge Kabakci, and Rei Mariman. “Generative AI without guardrails can harm learning: Evidence from high school mathematics”. Proceedings of the National Academy of Sciences 122.26 (2025). Used for the contrast between assisted practice performance and later unaided performance in the study’s mathematics setting.
• Chalmers, David J. “What We Talk to When We Talk to Language Models”. Preprint (2026). Used as a first-person example of a philosopher exploring philosophical questions with Claude and ChatGPT.
• Fan, Yizhou, Luzhen Tang, Huixiao Le, Kejie Shen, Shufang Tan, Yueying Zhao, Yuan Shen, Xinyu Li, and Dragan Gasevic. “Beware of Metacognitive Laziness: Effects of Generative Artificial Intelligence on Learning Motivation, Processes, and Performance”. British Journal of Educational Technology 56.2 (2025): 489-530. Used for the distinction between assisted writing quality and measured knowledge or transfer.
• Gabriel, Iason. Dr Iason Gabriel. Used to verify Gabriel’s work as a philosopher and research scientist at Google DeepMind.
• International Center for Academic Integrity. The Fundamental Values of Academic Integrity. Used for the integrity values included in Responsible Use.
• Kestin, Greg, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti. “AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting”. Scientific Reports 15 (2025). Used for the bounded example of a purpose-built tutor and immediate learning outcomes.
• Lazar, Seth. “Are Frontier Models Good at Ethics?”. Cosmos Institute (2026). Used for the limited claim that frontier models can show local textual facility with moral reasons and arguments.
• McCord, Brendan. “The Artificial Spectator”. Cosmos Institute (2025). Used for the Smith-inspired argument about assistance and the formation of practical judgment.
• National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1 (2024). Used for confabulation, information integrity, provenance, and context-sensitive risk evaluation.
• O’Neill, Elizabeth, Michal Klincewicz, and Michiel Kemmer. “Ethical Issues with Artificial Ethics Assistants”. In The Oxford Handbook of Digital Ethics, edited by Carissa Veliz. Oxford University Press (2022). Used for the benefits and risks of AI-supported moral deliberation.
• Perkins, Mike, Jasper Roe, and Leon Furze. “Reimagining the Artificial Intelligence Assessment Scale: A Refined Framework for Educational Assessment”. Journal of University Teaching and Learning Practice 22.7 (2025). Used for task-level AI permissions and boundaries.
• UNESCO. Guidance for Generative AI in Education and Research (2023). Used for human agency, privacy, pedagogical fit, and accountable educational use.
• UNESCO. Recommendation on the Ethics of Artificial Intelligence (2021). Used for attention to affected groups, representation, human oversight, and the distribution of benefits and harms.
• Ye, et al. “Language Models as Critical Thinking Tools: A Case Study of Philosophers”. COLM 2024. Used for the interlocutor, monitor, and respondent roles derived from interviews with 21 professional philosophers.