School, work honesty, and citations

You are the author of any file that carries your name, including the parts an AI tool wrote. AI models invent citations that lead nowhere, so open every source link yourself, keep only the sources that really say what you claim, and follow the honesty rules your school or employer has already written down.

Say you have a lab report due at midnight. At 11:20 pm the bibliography still needs four papers on primer design, which is how scientists pick the short DNA pieces that start a copying reaction. You asked ChatGPT for “four real papers with DOIs, APA format.” A digital object identifier (DOI) is a permanent web address for a paper, and American Psychological Association (APA) style is a common way to format citations. Two of those addresses open a “not found” error page. A third opens a real article, but it is about soil fungi and not the title in your list. Your school’s portal already holds an earlier export named Lab4_FINAL.pdf. You have forty minutes, two dead links, one wrong paper, and a file name that already says final.

The rule lives on a page you can quote

In that story, the biochemistry syllabus has one line: “Use of unauthorized materials or automated tools on graded work is academic dishonesty.” You have not opened that PDF since week one. The lab chat thread says “ask ChatGPT for papers, everyone does,” but a chat thread is not the syllabus. If the official page is silent, send a one-sentence email and keep the reply with the assignment.

Universities have started writing these rules down. Notre Dame’s Honor Code office tells instructors to be explicit and to treat unauthorized generative AI as an Honor Code issue, and its sample syllabus language is useful even if you have never been near South Bend. When AI is banned, the ban can include writing and editing tools, and they name Grammarly. When AI is allowed, the credit you give should be reproducible, meaning someone else could repeat it: the date you accessed the tool, the URL, and a description of the prompt. The student is still responsible for quality, completeness, and accuracy, because the model does not sit in the hearing.

Work follows a different rulebook. Your company may push Copilot, Gemini for Workspace, or Claude for work, and still fire you for dropping a customer file into a personal ChatGPT tab. So find the policy page, quote the line, and only then open the tool. If there is no page, ask your manager in writing. A specific question like “Can I use Claude to outline this weekly memo if I check every number?” gets a better answer than “Are we allowed to use AI?”

Rule of thumb: Quote the policy in one sentence before you paste the prompt. If you cannot find a sentence, you do not have permission yet, so ask.

The post on using AI for writing and questions is about getting a draft onto the page. This one is about who owns that draft once your name goes on it. If you still need to pick a product, the product chooser and the Learn page stay useful. Honesty does not depend on which logo is in the browser tab.

Four rooms, four honesty lines

People lump “school” and “work” into one worry, but there are really four different settings. Homework is not a take-home exam, and a messy first draft in Google Docs is not the PDF you email to a client.

The word “typical” in that table is doing a lot of work. Your lab may ban even spellcheck, and your newsroom may require an AI line on every reported piece. So copy the four columns onto the assignment or the ticket, and fill them in from the actual policy page rather than from this post. One line does not change from room to room: a source you did not open is not a source.

Why two of those DOIs led nowhere

Language models work by predicting the next small piece of text. They are very good at the shape of a citation, with author names that sound like real authors and a 10. prefix that looks like a DOI. They are not a library catalog. When their training material is thin on your topic, the shape still arrives, but the paper may not exist.

This has been measured, and it is more than folklore. In 2025, Jake Linardon and colleagues asked GPT-4o to generate citations for six mental-health literature reviews. The model produced 176 citations, and 35 of them (19.9%) were fabricated. Of the 141 real citations, 64 (45.4%) still contained errors, most often a bad or invalid DOI. When the model attached a DOI to a fabricated citation, 21 of those 33 DOIs were valid identifiers that pointed at unrelated articles, and 12 of the 33 were dead. In the lab report story you would hit both kinds in one night: a “not found” page, and a live DOI that belongs to soil fungi.

A link that opens is not a pass. The check is three questions in order: does it open, is it the paper you named, and does it contain the claim you used? Asking for “four real papers with DOIs” is a request for the shape of a citation, so ask for search terms instead and open the PDFs yourself. Asking the same chat “are these citations real?” is not a check either, since models tend to confirm their own fiction.

You are the author of the file you submit

Honor codes and workplace policies care about the name on the work. ChatGPT does not have a student ID, and your name is the one on the PDF. If a DOI is fake, you submitted a fake DOI. If a paragraph is generated and the syllabus required your own prose, you submitted generated prose. The tool can be allowed, required, or banned, but authorship never moves to the tool.

Hiding your use is the usual way trouble starts, even in settings where the tool is fine. On a take-home exam that bans all help, opening the tab was already the violation. On a homework set that allows an outline if you say so, the outline is fine but secretly pasting in the discussion section is not. A manager who said “use Copilot, then you check the numbers” will still treat a made-up analyst quote as your quote.

A disclosure that someone else could repeat is boring on purpose: the tool, the date, what you asked, what you kept, and what you checked. Notre Dame’s “date accessed, URL, prompt” is the school version, and at work it can be five lines at the top of the document. A bare “I used AI” is not repeatable. Privacy sits next to honesty here, because a customer contract in a personal chat is a leak even if every citation is real. The longer version is in the post on privacy and running AI yourself.

A source check you can run in ten minutes

Make one row per citation you plan to keep, and fill it in before you export the PDF. The rows below are made up to match the lab report story. A “not found” page means delete the citation, and so does a live DOI with the wrong title. Never “fix” an identifier by editing digits until something loads, because whatever loads will be someone else’s paper. The short script below is optional. It just keeps the same checklist in one place, one entry per source.

# Citation gate. Grain: one dict per source in the file you will submit.
# You are the author. A 404 is a delete, not a rewrite.

sources = [
    {"n": 1, "title": "Primer design review (toy KEEP)",
     "doi": "10.0000/ex-keep-1", "opened": "yes",
     "title_matches": "yes", "claim_in_pdf": "yes", "status": "KEEP"},
    {"n": 2, "title": "Novel PCR kinetics in undergrads (toy 404)",
     "doi": "10.0000/ex-404-2", "opened": "404",
     "title_matches": "n/a", "claim_in_pdf": "n/a", "status": "DELETE"},
    {"n": 3, "title": "High-fidelity polymerases for lab courses (toy WRONG PAPER)",
     "doi": "10.0000/ex-wrong-3", "opened": "yes",
     "title_matches": "no (soil fungi)", "claim_in_pdf": "no", "status": "DELETE"},
    {"n": 4, "title": "Campus PCR protocol, methods note (toy KEEP)",
     "doi": "10.0000/ex-keep-4", "opened": "yes",
     "title_matches": "yes", "claim_in_pdf": "yes", "status": "KEEP"},
]

print(f"{'n':<3} {'status':<8} {'opened':<8} {'match':<18} doi")
for s in sources:
    print(f"{s['n']:<3} {s['status']:<8} {s['opened']:<8} {s['title_matches']:<18} {s['doi']}")

deletes = [s for s in sources if s["status"] != "KEEP"]
print("Do not submit until deletes == 0. Count:", len(deletes))

# Markdown you can paste under the assignment (fill the brackets).
disclosure = """
## AI use
- Tool and surface: [ChatGPT web / Claude / Copilot / none]
- Date: [YYYY-MM-DD]
- What I asked: [outline / search terms / grammar / none]
- What I wrote myself: [analysis, citations, final prose]
- Citations in this file: [N]. Each DOI opened. Deletes: 0.
- Policy I followed: [syllabus line / manager Slack / journal AI note]
"""
print(disclosure)

This table shows what that code prints for the four made-up rows:

The count of deletes is 2, so you do not submit yet. You remove papers 2 and 3 from the list and from any sentence that leaned on them, and you fill in the disclosure block only after the count reaches 0. The 10.0000/ex-… strings are labels for this teaching table, so do not paste them into a real paper. For live work, type the DOI after https://doi.org/. A library search is how you find papers. The chat may suggest search terms when policy allows help, but it is a bad place to collect finished reference lists.

Worked example: four citations, two keepers

Back to the lab report, on a Thursday night with the deadline close. You already wrote a methods paragraph from your own bench work, with the primer sequences, an annealing temperature of 58 C (the temperature at which the primers stick to the DNA), and 35 cycles. That paragraph is yours. What caused the panic is the “background” section the grading rubric wanted, with citations, so you pasted the methods into ChatGPT and asked for four papers.

The one thing you should not do at 11:32 is ask the model for two more papers to fill the holes, because that is how you get a fifth dead link. Keep citations 1 and 4, rewrite the background around the protocol you actually ran and the one review you opened, and add the disclosure the syllabus required once you find the line. If the rubric demanded four sources and you only have two real ones, turn in two with an email: “I discarded two model-suggested citations that did not resolve. Happy to add sources tomorrow if you want a revision.” That is ugly, but it is better than two dead links in a PDF the teaching assistant will open on a phone.

At work the same problem looks like a slide titled “Industry benchmarks (2025).” It has four bullets and four studies, and two of them are made up. Pull the made-up bullets, and put “source: our pipeline, last 12 weeks” on the real numbers. A sales deck counts as published analysis the moment it leaves the building.

Common mistakes

• Pasting a chat bibliography without opening a single DOI.
• Using “are these citations real?” as the only check, when the same model wrote them.
• Treating a live DOI as a pass when the title or topic does not match.
• Writing “I used AI” with no tool, date, or prompt.
• Assuming Grammarly is fine when the syllabus banned automated editing tools.
• Asking for “hints only” on a take-home exam that is closed to outside help.
• Dropping a customer spreadsheet into a personal free chat because the office Copilot was slow.
• Replacing a dead link by editing the DOI until a page loads, which gives you someone else’s paper.

How to practice this week

Pick one file with your name on it. Write the four-column table for that setting from the real policy page. Run the citation check on every source, even if there are only two, and then fill in the disclosure block or write “no AI used, citations opened on [date].” If a DOI leads to an error page, delete it today.

If the policy page is blank, send the one-sentence question. For school: “For lab 4, may I use a chatbot for search terms if I open every paper myself and add a disclosure?” For work: “For the Friday client memo, may I use Claude to outline if I verify every figure and cite?” Save the reply next to the file.

The next post covers health, legal, and money: hard limits, where “disclose and proceed” is not enough. The earlier post on scams and deepfakes covers the other side of everyday AI safety. The hub is AI safety for everyday life. If you still need setup help, see AI setup from zero and Practical AI.

Quick recap

• Quote the syllabus or the workplace AI page before you paste the prompt. If there is no page, ask in writing.
• Homework, take-home exam, workplace draft, and published analysis are four settings. Fill in allowed, disclose, and forbidden from the real policy.
• You are the author of the file you submit. The model does not take the hearing or the client call.
• Open every DOI. A dead link is a delete, and so is a live DOI that leads to the wrong paper.
• Keep a repeatable note: tool, date, what you asked, what you kept, deletes at zero.

Series notes

This is Part 2 of AI safety for everyday life. Previous: Scams and deepfakes. Next: Health, legal, and money.

Sources

Pages used for this article. Course and company rules still beat any of these.

• University of Notre Dame, Undergraduate Academic Code of Honor: AI recommendations for instructors (unauthorized generative AI as an Honor Code issue; sample syllabus language; reproducible attribution: date, URL, prompt; student remains responsible for accuracy)
• Linardon J, et al. Influence of Topic Familiarity and Prompt Specificity on Citation Fabrication in Mental Health Research Using Large Language Models: Experimental Study. JMIR Mental Health (2025) (GPT-4o: 176 citations, 35 fabricated (19.9%); 64 of 141 real cites had errors; fabricated cites with DOIs often resolved to unrelated articles)
• The DOI Foundation: doi.org (resolve an identifier; a 404 or a mismatched title is a delete)
• OpenAI usage policies and Anthropic usage policy (vendor rules are not your syllabus or your employment contract)
• AMS: AI safety for everyday life (this series) and AI for writing and questions (drafts, not authorship)

Merch on Etsy

Open store →

Git Version Control Shirt View on Etsy → Standard Deviation White Ceramic Mug View on Etsy → You're On Mute Coffee Mug View on Etsy → Apache Hadoop T-Shirt View on Etsy →