Text Generation, Summarisation & Paraphrasing

Generation,Generation, summarisation, and paraphrasing look like three separate buttons in most AI tools, but underneath they are the same mechanism doing three different jobs: predicting plausible next text, given whatever instruction and source material it has been given. That shared mechanism is exactly why all three share a family of failure modes that are easy to miss if each task is learned as an isolated skill rather than as a variation on one underlying risk — a model producing fluent, confident text with no built-in signal for how well-grounded that text actually is.

Generation: Creative Versus Factual

WriteWrite me something" covers two very differently risky requests. Creative generation — a poem, a marketing tagline, a story opening — has no external ground truth to be wrong about, so the main risks are stylistic (tone, originality, appropriateness) rather than factual. Factual generation — "explain how X works," "list the key dates in Y," "summarise the current consensus on Z" — is a different proposition entirely, because the model produces text with exactly the same fluent, confident register whether or not the underlying content is accurate. This is the core finding running through the wider research literature on language-model hallucination: models generate content that is fluent and internally coherent but not reliably grounded in verified fact, and — critically — nothing about the surface style of the output signals which category a given sentence falls into [1]. A fabricated citation, a wrong date, or an invented statistic reads, on the page, exactly as confident as a correct one. The practical implication is procedural rather than technical: any AI-generated factual claim that matters needs to be checked against an independent source before it is relied on, precisely because the model's own tone gives no reliable warning when it is wrong.

Summarisation: Abstractive Versus Extractive

ExtractiveExtractive summarisation selects and rearranges sentences that already exist in the source text; abstractive summarisation generates genuinely new sentences that paraphrase and compress the source's meaning. Modern AI tools default almost entirely to the abstractive mode, because it reads far better — but abstraction is exactly where faithfulness risk is introduced, since every rewritten sentence is a new opportunity for the model to add, drop, or subtly alter content relative to the source. A detailed human-annotation study of summaries produced by several leading abstractive systems found substantial amounts of hallucinated content — material not actually supported by the source document — present across every system evaluated, not as an occasional edge case but as a structural property of how abstractive summarisation currently works [2]. That finding is the direct evidence behind a rule worth teaching as a habit rather than a caveat: a fluent AI-generated summary is not, by default, a faithful one, and the two need to be checked separately.

The Caveat-Dropping Failure Mode

OneOne specific, easy-to-miss pattern deserves calling out on its own: a summary can be factually accurate in every sentence and still be badly misleading, because compression silently removed a qualifier that changed the source's actual claim. "The trial showed a benefit in a specific subgroup, under specific conditions, and the authors caution the result needs replication" is not faithfully compressed by "the trial showed a benefit" — every word of the shortened version is technically present in the source, and yet the caveats that made the original claim honest have vanished. This is a harder failure to catch than an outright fabrication, because nothing in the summary looks wrong in isolation; it only becomes visible by comparing the summary against the source sentence-by-sentence and asking, specifically, "what qualified this claim in the original, and is that qualifier still here?" — the same discipline the annotators in faithfulness research apply systematically [2], adapted here into something a student can do by hand on a single document.

Paraphrasing: Legitimate Uses and Integrity Risk

ParaphrasingParaphrasing tools have real, defensible uses: adapting dense academic language for an accessibility need, adjusting register for a different audience, or rephrasing a technical explanation into something a specific reader will actually understand. The same mechanism has a second, illegitimate use: running one's own or someone else's text — or AI-generated text — through a paraphraser specifically to defeat plagiarism checking or AI-content detection. This is not a hypothetical risk. Purpose-built paraphrasing systems have been shown to sharply reduce the accuracy of AI-text detectors — one widely cited study built an academic paraphrasing model and demonstrated it could drive a detector's accuracy down dramatically while barely affecting the text's readability [3]. Compounding the problem, a large multi-institution evaluation of commercially available AI-text detection tools found none of them reliably accurate, with real risks in both directions: false negatives that miss AI-generated or AI-paraphrased text, and false positives that wrongly flag genuinely human-written work [4]. Taken together, these findings point to a conclusion worth stating plainly rather than left implicit: because detection is demonstrably unreliable in both directions, the honest position on paraphrasing is not "don't get caught" — it is that submitting AI-generated or AI-paraphrased text as one's own unaided work is an integrity violation regardless of whether any tool happens to catch it, and institutional policy on academic integrity should be read and applied on that basis, not treated as a detection arms race to be won.

References

1. Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12), Article 248. https://doi.org/10.1145/3571730
2. Maynez, J., Narayan, S., Bohnet, B., & McDonald, R. (2020). On Faithfulness and Factuality in Abstractive Summarization. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 1906–1919. https://aclanthology.org/2020.acl-main.173/
3. Krishna, K., Song, Y., Karpinska, M., Wieting, J., & Iyyer, M. (2023). Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense. Advances in Neural Information Processing Systems, 36.
4. Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, 26. https://doi.org/10.1007/s40979-023-00146-z