I keep telling teachers that detection software will never solve the AI writing problem, and that a culture of disclosure and judgment might. A new essay from Petar Jandrić (2026), editor-in-chief of Postdigital Science and Education, shows what happens when that culture breaks down.
From the editor’s chair, he estimates that nine of every ten submissions now arrive cowritten with some form of GenAI, and that a rising share fail basic scholarly standards. The unsettling detail is that among the senders are established academics with years of solid work behind them.
Jandrić’s reframe is key here. He grounds this in two old Greek ideas, episteme, or justified true belief, and doxa, opinion that may not be true. People and machines both produce a blend of the two, yet only humans can tell them apart. As he argues, “GenAI can offer useful suggestions, yet for as long as it is structurally unable to distinguish between doxa and episteme, judgement of GenAI suggestions remains with humans” (p. 382).
Generative AI and Academic Writing
Jandrić lays out four challenges from daily practice. Detection is the first, and his verdict matches the research: his editor’s hunch is fallible and the automated detectors are worse, so unacknowledged GenAI use can’t be proven the way old-fashioned plagiarism can. That tracks with what Bassett and colleagues (2026) documented about AI detectors failing in education. Sourcing is the second, where fabricated references force him to cross-check every citation, and his rule is firm: an argument you can verify cannot rest on sources you can’t.
The writing challenge is quieter and more corrosive. At sentence level, GenAI is invisible. At scale it produces convincing prose that says very little, circular and repetitive under the polish. With enough human editing, some of these papers pass review and even read as original. Jandrić admits he published one such piece and later could not retract it, because superficial paraphrasing is not provable misconduct.
The fourth challenge, attitude, is where the essay turns personal. He recounts an author who submitted three drafts, including two replies to his feedback, all written by GenAI, after insisting the work was his own. Jandrić now keeps a blacklist, and he describes editors brought to tears by the same betrayal. His summary of the toll is hard to argue with: “as GenAI articles proliferate, editing becomes less and less focused on concepts and ideas and more and more focused on detection and management of fraudulent behaviour” (p. 383).
Why the Rulebook Fails
If you expected publisher guidelines to help, Jandrić’s analysis will disappoint you. He reads Springer Nature’s rules as unenforceable and self-contradictory. No one can separate “AI used for grammar” from “AI used to build the argument,” so the box authors tick to declare light use mostly absolves the publisher of responsibility.
He calls these “cynical attempts at moral and legal whitewashing with little or no epistemic consequence” (p. 380). His sharpest example is the instruction that editors must not feed manuscripts into AI tools, printed one line above the instruction to declare the AI tools they used in evaluation.
I think he’s right about the guidance, and it points to a gap the field keeps dodging. Real disclosure norms, the kind Cleland and colleagues (2025) worked through for AI use in academic publishing, ask more of everyone than a checkbox does. A culture of acknowledgment is harder to build than a policy is to post, and far more useful.
Where I’d Land
Here is where I part ways with the essay, gently. Jandrić ends by announcing a zero-tolerance policy: improper or unacknowledged GenAI use means rejection and a blacklist. I share his exhaustion and I understand the impulse. He has just spent the whole paper explaining that this use usually cannot be proven, that his hunch fails, and that the cowriting-versus-word-processing line is dissolving. A zero-tolerance rule aimed at something you concede you often can’t detect risks punishing the clumsy while the genuinely deceptive keep slipping through.
The more durable path is the one Jandrić models in his better moments: rebuild authorship and disclosure as shared norms, and judge the work by whether it holds episteme, his own test. That comes close to what Eaton (2023) called postplagiarism, a world where the old categories of original and copied no longer map cleanly onto how text gets made, and where integrity has to be taught and negotiated, not just enforced.
His best image is a bicycle: a well-built human-GenAI partnership should extend what we can do within real limits, as long as we understand its strengths and its constraints. That is the version of this debate worth having. Detection will keep losing the arms race. What lasts is a scholarly culture that values how knowledge is made, names its tools openly, and keeps human judgment in the chair.
References
- Bassett, M. A., Bradshaw, W., Bornsztejn, H., Hogg, A., Murdoch, K., Pearce, B., & Webber, C. (2026). Heads we win, tails you lose: AI detectors in education. Journal of Higher Education Policy and Management. https://doi.org/10.1080/1360080X.2026.2622146 //
- Cleland, J., Driessen, E., Masters, K., Lingard, L., & Maggio, L. A. (2025). When and how to disclose AI use in academic publishing: AMEE Guide No. 192. Medical Teacher. https://doi.org/10.1080/0142159X.2025.2607513
- Eaton, S. E. (2023). Postplagiarism: Transdisciplinary ethics and integrity in the age of artificial intelligence and neurotechnology. _International Journal for Educational Integrity_, 19(23). https://doi.org/10.1007/s40979-023-00144-1
- Jandrić, P. (2026). GenAI and academic writing. Postdigital Science and Education, 8(2), 372–386. https://doi.org/10.1007/s42438-026-00641-9