PhilSci-L03b: GenAI, Writing and Philosophical Learning

In short

How may you use generative AI (GenAI: ChatGPT, Claude and the like) in this course, and why does Sebastian De Haro (the course coordinator) draw the line where he does?

His answer: AI is fine when it strengthens your own learning (explaining a term after you have tried, quizzing you, commenting on text you wrote) and not when it replaces it (doing the reading, writing what you submit). Writing with AI is only responsible for someone who is already an expert and can already write philosophy, because only they can catch its mistakes.

Not on the exam. No mock exam question touches this session and the syllabus does not list it as course content. Read it for the rules that govern your paper and presentation. The rules and the paper requirements come first; the reasoning behind the rules follows.

The line: does AI strengthen or replace your learning?

De Haro takes neither side of the usual debate. He does not cheer AI on and he does not ban it. His test is a single question.

The central question

“The central question is whether AI strengthens or replaces your own learning.”

Uses that support learning: asking for clarification of a difficult definition, requesting alternative examples, testing whether you can explain an argument, and using feedback to revise text you have written. Uses that replace it: outsourcing the reading, submitting AI-generated answers as your own, using generated text before you can judge its quality and accuracy, and not acknowledging AI use.

What the supporting uses have in common (my reading; De Haro does not spell it out) is that you have already done the work. You read the text and got stuck on a definition. You understand a concept well enough to judge whether a new example fits it. You have an argument in your head to be tested on. You wrote the text being commented on. In each case the AI comes after your own effort. The replacing uses are the same activities with your effort taken out.

The rules for your paper

The policy appears twice: in De Haro’s course policy, and in the section “About the use of AI tools” of the Project Description, which is written for the paper and is more detailed and formal. The two are consistent.

AI use is permitted but not encouraged. All conceptual analysis, interpretation and philosophical argument must be yours.

Allowed:

• explaining a term after you have tried to understand the text (clarify)
• asking for alternative examples, then judging whether they fit (compare)
• having AI question you on an argument you have read (test)
• suggestions on a text you wrote, where you decide which to take (revise)
• improving clarity and grammar, refining phrasing
• brainstorming and putting ideas in context
• finding relevant papers or concepts, as you would with a search engine

Not allowed:

• outsourcing the assigned reading
• generating a text for submission, or submitting AI-generated text as your own
• generating or structuring your arguments or analysis
• using text you cannot assess for accuracy, quality and argumentation
• fabricating references or quotations

Each of the first four carries a qualifier (after you tried, then judge, a text you wrote, you decide), and the qualifier is what makes it allowed: without it, the same use lands in the second list. Brainstorming is allowed but structuring your argument is not, so the practical line is: AI may help you find a topic and the literature, but the argument and its outline have to come from you.

The AI use statement, and the rest of the small print

• If you used AI: add an AI use statement at the end of the paper naming the tool(s), the purpose(s) and the extent of the help.
• De Haro’s policy asks written work to state what you did yourself, explain whether and how AI was used and for which purpose, and say which choices were your own. “Your submitted work must make your own reasoning and responsibility visible.” My reading: if you used no AI, one line saying so meets this.
• You are responsible for checking every claim and every citation. An AI-suggested paper that does not exist, or does not say what the AI claims, ends up in your paper as a fabricated reference.
• The vault’s lecture notes are AI-written too. For the paper, read the texts you discuss in the original, and take every quotation and page number from the original text, never from a note.
• Put no personal or confidential data in prompts.
• Misuse is academic misconduct under the UvA Regulations Governing Fraud and Plagiarism (2019).
• De Haro, on the Jetten case: “Pangram detection has become very good.” (Pangram is a commercial AI-text detector.) Whether it is run on papers is not stated; do not count on undeclared AI text going unnoticed.

De Haro’s own slides carry an AI use statement, which is one example of the form: “AI was used to produce a first draft of these slides, based on an outline, detailed instructions, and texts that I supplied to AI Chat. Afterwards, I edited and adapted that draft to produce the final slides.” Copy the form only. The use it describes, drafting text, is not allowed in your paper.

The Project Description’s AI section speaks of “your submission”, meaning the paper. The course policy covers all submitted work, and the presentation slides are submitted on Canvas, so the safe reading (mine; no source says it outright) is that the same rules apply to the slides, with a one-line AI use statement on the last slide if you used AI for them.

Requirements and writing advice for the paper

The paper is marked on five criteria (from the Project Description):

1. Understanding and explanation of the ideas and arguments in the original texts.
2. Structure: the structure of the paper reflects the structure of your argument.
3. Development of the thesis: argumentation without logical gaps.
4. Clarity and precision of the thesis and the argumentation.
5. Difficulty and originality.

What else the Project Description and the proposal guidelines set:

• Length. Around 4,000 to 7,000 words per student (8 to 15 pages), everything included. This is only a guide.
• Minimum requirements. On time, correct spelling and grammar, no plagiarism.
• Formative part. 10% of the paper mark comes from the project idea discussion, the proposal, how you use the feedback, and the first draft. The Project Description also calls the draft “not graded”. My reading: the draft gets no grade of its own, but it counts towards that 10%. Hand it in.
• The draft may be partial, but should have a title, abstract, table of contents with a short description and references under each item, a reference list, and substantial text.
• Stay near the proposal. Small changes are allowed, but stay as close to the approved proposal as you can.
• Submission. On Canvas, one submission per group, for both the draft and the final version.
• One question. The paper should pursue a single, unified question in depth. A paper whose scope is too wide, with too many claims not properly justified, cannot get above a 7.
• Originality. Explaining an existing question, debate or line of reasoning in your own words, clearly and insightfully, can earn an 8. Higher needs a contribution of your own, even a small one, made by engaging the key arguments in the literature: explicate a difficult problem one of the authors raised, or apply their view to a new case and draw a surprising conclusion. Doing things your own way without engaging the literature does not count.
• Audience. Write for fellow students who did not take this course or do the readings, but are generally engaged in philosophy of science.
• Starting point. An existing discussion in the philosophical literature. The Project Description recommends the Stanford Encyclopedia of Philosophy page on your topic and journals such as Synthese or Philosophy of Science. Popular science, articles for a broadly interested philosophical audience, and scientists writing about philosophy may be used and cited, but should not be your main topic.
• Citations. APA or similar, with page numbers in the text: “Jones (1985: pp. 175-176)”, never just “Jones (1985)“.
• Scope. State the limits of your argument explicitly.
• Argument reconstruction. Premises and conclusion (P1, P2, C) can help, e.g. “I accept P1 but argue against P2”. If you use one, state it in a valid form, for example modus ponens or modus tollens.

Luka Vinck’s writing workshop (tutorial 3) covers how to build the paper section by section: thesis, argumentation scheme as the outline, what goes in the introduction and at the start and end of each argument, and plain language without overstatement. It is written up in PhilSci - Tutorials.

For the presentation, the obvious move (my suggestion; no source says it) is the same skeleton: thesis first, then the argumentation scheme. The presentation’s criteria and evaluation form are under Canvas Files (per the course overview) and are not in the vault.

Why he draws the line there

Generating an answer is the cheap part

De Haro is not sceptical across the board. In mathematics, he grants, AI may help find relations across technical literatures and suggest hypotheses for someone to check. He adds two caveats: access depends on a few commercial companies, and

De Haro

Generation is not verification, proof, or understanding.

Producing a candidate answer and establishing that it is correct are different jobs. AI has made the first cheap; the second still takes someone who understands the material.

The evidence: more writing, worse writing

A 2026 study by Gartenberg and colleagues looked at five years (from January 2021) of submissions and peer reviews at Organization Science, a management journal. They measured writing quality by readability, jargon, passive voice, hedging and similar markers.

Their central claim: current AI tools, combined with incentives, are pushing research towards more, rather than better, writing.

• More. Submissions rose by 42% after ChatGPT appeared in late 2022.
• Worse. Flesch Reading Ease (a standard readability score: longer sentences and longer words give a lower score) fell by 1.28 standard deviations between January 2021 and January 2026. It had been flat for a decade and dropped sharply only after ChatGPT.
• Manuscripts with more than 30% AI use were desk-rejected (turned down by the editor without going to reviewers) 30% more often, and rarely got a revise-and-resubmit decision (the usual route to acceptance).
• Reviews were affected too: more than 30% used some AI, and those were harder to read and narrower in focus.

The concern De Haro draws out: easy text generation shifts costs to editors and reviewers. Writing a submission got cheaper; reading and judging one did not.

Why “more rather than better”? The authors’ diagnosis is institutional:

Universities reward publications in top journals
  -> researchers are under pressure to increase output
  -> GenAI reduces the cost of producing text
  -> journals receive more weak submissions

The study has limits. The finding is about the system as a whole, and says nothing about whether a particular author writes worse with AI. It comes from one journal. And it does not show that every use of AI makes text worse.

The authorship problem

The second example is public communication. The Dutch newspaper NRC reported that 232 social media posts by Prime Minister Rob Jetten and his party had been AI-generated. It was most sensitive where the posts were apologies to the Moluccan community, and public debate turned on authenticity and responsibility.

De Haro’s question: AI can make things better, but who is speaking, and who is responsible for the message? His moral: using AI is not the whole issue; context, transparency and authorship matter.

The link to your paper is my reading, and no source states it: an apology only counts if the person apologising made it, and a submitted paper is the same kind of text.

His own experiment: 500 pages and still unfinished

De Haro tried writing a review chapter on scientific theories with AI, for an Elsevier reference work. A review paper surveys an existing debate rather than proposing new research. So he did not start by asking AI for original philosophical ideas, only for text around an argument he supplied. He gave it a detailed outline with a word count per section and the LaTeX format of an earlier chapter, and worked through it section by section.

The division of labour:

• He did: the detailed outline and the argumentation, detailed instructions on style and literature, and endless rewriting, feedback and writing of his own.
• The AI did: produce text to his instructions, and rewrite it after his feedback.

Expert guidance was needed throughout: prompting, checking, selecting, writing and rewriting. Result: 500 pages of conversation, and the paper is still unfinished. The obvious lesson for a student, though De Haro leaves it unstated: if it cost an expert who supplied the whole argument this much, it will cost more for someone who does not yet have the argument worked out.

Writing with AI requires expertise

The general conclusion is a pair of conditions, and both have to hold.

When you can responsibly use AI to help write

You can responsibly use AI to help write only if both conditions hold:

(A) You are already an expert. You can check claims, detect omissions, and revise to improve the argument.

(B) You can already write philosophy. You recognise unnatural text and weak or incomplete arguments.

“Before that stage, the priority is to develop your own philosophical voice and judgement.”

Both conditions are mainly about judging what the AI produced: checking claims, spotting gaps, recognising weak arguments. So AI-assisted writing is safe only for someone who could have spotted the problems anyway. Until you meet the conditions, your own practice comes first.

The mismatch is sharpest in this subject. De Haro’s assessment is that GenAI is good at surface summaries but poor at argumentative nuance, and the course is about exactly the nuance: reading and reconstructing arguments carefully, judging whether they succeed, and writing and criticising text by the same standards.

What to remember

• The test is whether a use strengthens or replaces your learning. The four study uses De Haro allows (clarify, compare, test, revise) all come after your own effort.
• Bright line for the paper: AI must not generate or structure your arguments or analysis, or write text you submit.
• AI use statement if you used AI: tool, purpose, extent, what you did yourself and which choices were yours. If you used none, one line saying so (my reading of De Haro’s transparency rule).
• Take every quotation and page number from the original text. A hallucinated reference in your paper is a fabricated reference.
• The draft feeds the 10% formative part of the paper mark: hand it in.
• Writing with AI needs both expertise and the ability to write philosophy already, because both are mainly about judging the output.
• De Haro’s closing advice: use AI to test your understanding; develop your own philosophical voice before relying on AI to draft text; be transparent about AI use and responsible for every claim you submit.

Links

• Course: Course overview
• Same session: PhilSci-L03 - Under-determination was the first half of the 15 September flipped classroom
• Writing workshop: PhilSci - Tutorials, tutorial 3
• Sources: 3b GenAI.pdf (lecture slides, 15 September 2026) · 2026 Project Description Philosophy of Science.pdf, sections “Project timeline”, “The Paper”, “Further comments” and “About the use of AI tools” · 2026 Guidelines Project Proposal Philosophy of Science.pdf
• Further reading: Gartenberg et al. (2026), via the summary document provided for the class · Denise Retera, “AI-gebruik van premier Jetten op social media ligt onder de loep”, NRC, 11 September 2026