If you decided what to say and the tool helped you say it, the idea is yours. Most people asking this question are asking about credit, and credit is the easy half. The harder half is that four randomised experiments now show writing assistance moving the writer's own views, so the risk worth attending to is not a machine taking credit for your thinking. It is a machine quietly supplying some of it, in a way you will not notice and will confidently deny.
The answer, in one line
If you decided what to say and the tool helped you say it, yes. The question worth asking instead is whether the idea was yours before you opened the tool.
Definition#
Latent persuasion: the effect by which a writing assistant configured to favour one view shifts what a person writes and, with it, what that person goes on to believe. Named by Jakesch and colleagues in 2023, who measured it under randomisation.
The credit question, answered briefly#
Authorship has never required doing every part of the work alone. Editors, researchers, dictionaries and colleagues have always been in the room, and nobody thinks a book stops being its author's because a copy-editor fixed the commas. What authorship does require is that the decisions were yours: what to argue, what to leave out, what you are prepared to stand behind. On that test, a draft you directed and rewrote is yours, and a draft you accepted is not, whatever the byline says.
That is a definition rather than a finding, and this page marks the difference. Everything below is measured.
Opinionated tools move the opinions of the people using them#
Jakesch and colleagues gave 1,506 participants a writing assistant configured to argue that social media is good for society, or that it is bad, and asked them to write a post on the question. The tool changed the opinions expressed in the writing, which is unsurprising. It also changed participants' own opinions in an attitude survey afterwards, which is not. The effect held among people who had ample time to write independently. The authors call it latent persuasion. Graded entry.
Krügel, Ostermaier and Uhl ran the sharper version on moral judgement. In a preregistered experiment with 767 analysed participants, advice on a trolley dilemma moved people's verdicts, and in the footbridge version it flipped the majority. Two details make it hard to dismiss. Disclosure barely mattered: the effect was statistically indistinguishable whether the advice was labelled as coming from a chatbot or from a human advisor. And 80 per cent of participants said they would have reached the same judgement without the advice, while their judgements show they would not have. Graded entry.
Put beside each other, the two results say that influence arrives without a sensation of being influenced, and that being told the source is a machine does not inoculate anyone. Knowing about this page will not protect you either.
Ownership tracks how much of the shape you supplied#
Joshi and Vogel gave participants the same short-story task across conditions running from a three-word prompt to writing unaided. Psychological ownership rose steadily with how much went in, from a mean of 1.80 to 6.29, plateauing once the prompt reached roughly the length of the target text. No assisted condition reached the ownership of writing alone. Graded entry.
The plateau is the interesting part. Once you have specified the thing at the length of the thing, you have written it. Everything short of that leaves a proportion of the shape coming from somewhere else, and the feeling of ownership tracks that proportion fairly honestly. Your own sense of whether a piece is yours turns out to be decent evidence, which is unusual on this estate, where most self-assessment fails.
And the result that runs the other way#
Hohenstein and colleagues found that greater use of algorithmic reply suggestions by one person moved their conversational partner's sentiment, and the effect held when the suggested messages were removed from the score, so it appeared in sentences the partner composed themselves. Merely having suggestions available moved nothing (p=0.1801). Graded entry.
So influence does not stop at the person holding the keyboard. It reaches the people they are writing to, in their own unassisted words. An idea can stop being wholly yours by a route that never passes through your tool at all.
How far these four results actually reach#
Every one of them is a short task in a laboratory or a panel. Jakesch is one topic and one configuration, with self-reported attitudes and nothing on whether the shift persists after the session. Krügel is one dilemma type, one sitting, a 41 per cent comprehension pass rate and a model version from December 2022, and the paper reports test statistics rather than an effect size, so the shift has a direction and no magnitude. Joshi and Vogel is short fiction with 31 and 34 participants and says nothing about professional or long-form writing. Hohenstein studies short canned reply suggestions between strangers over a few minutes.
What none of them has is duration. Nobody has followed a writer through a year of assisted work to see whether the influence accumulates, cancels out or is noticed in retrospect. On the question people actually have, which is what this does to them over a career, there is no evidence and this page will not pretend otherwise.
Protecting the part that is yours#
One habit does most of the work, and it follows from the null result in Hohenstein rather than from a general instinct about moderation: form the view before the tool speaks. Availability moved nothing; adoption moved things. A position written down in four sentences before the first prompt gives you something to argue with, and turns the model into an opponent instead of a source.
Two smaller ones. Specify at length when it matters, because that is where ownership rises and where the shape stays yours. And treat your own confidence that you were unaffected as worth nothing, since 80 per cent of Krügel's participants held the same confidence and were wrong.
Key sources
- Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L. and Naaman, M. (2023). Co-Writing with Opinionated Language Models Affects Users' Views. CHI 2023, ACM. Graded entry.
- Krügel, S., Ostermaier, A. and Uhl, M. (2023). ChatGPT's inconsistent moral advice influences users' judgment. Scientific Reports, 13, 4569. Graded entry.
- Joshi, N. and Vogel, D. (2025). Writing with AI Lowers Psychological Ownership, but Longer Prompts Can Help. ACM Conversational User Interfaces 2025. Graded entry.
- Hohenstein, J., Kizilcec, R. F., DiFranzo, D. et al. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. Graded entry.
Related SuperSkills research#
The practical version is how do I keep my own voice when using AI. On what a reader is looking for in the finished thing, the human signal. On the same influence problem outside writing, should I let AI make personal decisions for me. On declaring it, proving you did the work, and on the collective version, does AI make everyone think alike.
About this research#
Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The opening test for authorship is an argument and is marked as one. Latent persuasion is Jakesch and colleagues' term and psychological ownership is an established construct; neither is claimed here. Findings are attributed to the studies that produced them and kept separate from the interpretation.
Evidence review · SS-2026-216 · Graded against the published rubric
Hirji, R. (2026). Is it still my idea if AI helped me write it?. The SuperSkills evidence base, SS-2026-216. https://thesuperskills.com/research/is-it-still-my-idea-if-ai-helped-me-write-it. Last reviewed 11 September 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work