What 41 heavy AI users said about tasks that felt like theirs and tasks that didn't, read as objectively as a small, self-selected sample allows.
Respondents locate ownership in the judgment they exercised, not in who produced the artifact. Ownership was reported most strongly for work they could not have produced alone, and withheld most often for work they could easily have done themselves.
The survey asked each person for one AI-assisted task that "felt very much like yours" and one that "did not really feel like yours," then asked directly what made the difference. Three patterns hold across the free text. First, ownership follows effort spent evaluating and reshaping output, and it collapses when the person cannot evaluate the output at all. Second, tasks people disown are usually short-form writing in their own voice (emails, LinkedIn posts, captions, a conclusion) where AI's voice mismatch is audible, or chores they never wanted. Third, a non-trivial minority rejected the owned/not-owned framing altogether, and their reasons are themselves a finding.
Ability. Each bar is 100% of the people who answered for that task (owned n=37, disowned n=38). The direction is the reverse of a naive expectation: owned work is more AI-dependent.
Motivation. Same two rows. Eleven of 37 owned tasks would not have happened at all; for disowned tasks the figure is 8 of 38.
Owned tasks that only exist because of AI: P14's graphic novel from his own books, P21's children's-book illustrations after a decade on the shelf, P24's video game, P2's product, P26's document-filing automation, P1's curated opinion corpus.
The design assumes everyone has both kinds of task. Fourteen people did not produce a clean pair. P3 ("I never had that feeling"), P8 ("There are none. I own all of the thinking"), and P2, P4, P25 had no disowned task. P5, P30, P38 and P17 had no owned task or lost it to a mis-click. P16, P18, P29 said the two felt the same. P10, P21, P23, P28 and P39 used the "not yours" slot for something they in fact owned, in P10's case because "AI failed so horribly." These are treated here as a finding about how ownership is construed, not as missing data. The pairing also appears to smuggle in a second variable: several people answered it as "something I chose" versus "something I was assigned" (P14, P34, P35), which is a question about the task, not about AI.
Self-selected, expert, and mostly friendly to the researcher. 29 of 41 live in the US; 27 are 45 or older; 8 are students aged 18–24, and one of them (P35) says the survey was posted to MIT Dormspam. Several respondents reference the researcher's own work (P7: "Helen's authorship concept"; P21: Staying Human; P25: "cognitive sovereignty"). Findings describe how sophisticated heavy users reason about ownership, not the general population.
Six bands, n=41
n=41
n=41
n=41; "Not applicable" is mostly students and retirees
P17 (never uses AI) answered no task questions. P5, P30 and P33 are rare users; P5 and P30 reported no owned task. P15 saw the owned-task page as a repeat and wrote "same as before" in every field. P38 advanced past the second task by accident and described it only in the closing question. Effective n for paired contrasts is roughly 27.
This is a heavy-use sample. Half use an assistant many times a day, 13 describe their work use as "constantly," and 20 pay out of pocket (five in the $100–199 band, three at $200 or more). Claude is the most-used tool, which is consistent with recruitment through the researcher's network.
n=41; 5 skipped
n=41
Multi-select, share of 41
Multi-select, share of 41. 11 use tool calls or MCPs.
Multi-select, share of 41. Learning and lookup lead; creative writing and casual chat trail.
Multi-select, n=41
Each described task was hand-coded into one of seven types. Owned tasks cluster in long-horizon building and code; disowned tasks cluster in writing that is supposed to sound like the person, and in chores. Coding is single-coder and provisional.
Count of described tasks. Owned n=37, disowned n=36 (P8 and P25 declared none; P17, P2, P4 blank).
Build = creative or system projects (graphic novel, game, product, agent harness, filing automation). Own-voice writing = anything meant to sound like the person (emails, posts, resume, blog, sermon, eulogy).
Coded from the direct question ("what made the biggest difference in how much ownership you felt?") plus supporting narrative. Multi-label; 41 respondents.
"Judgment invested" and "voice fidelity" are the two most frequent and are two sides of one mechanism: people own what they reshaped, and disown what came back sounding like someone else.
Would you be comfortable telling others you used AI? Coded from free text, n=36 with a codable answer.
Conditional answers hinge on audience and on being allowed to explain the extent of use (P41, P4, P16). Withholding concentrates on own-voice outputs: emails (P12), captions (P20), a party letter (P37).
Cases where the two diverge
Ordered by how much of the sample each one draws on. Respondent IDs are given so each claim can be traced to individual responses in the underlying data.
P14 did not draw a panel of his graphic novel and calls it fully his. P21 cannot illustrate and feels "complete ownership" of her book images. P24's game is "almost all mine, with AI acting as a subordinate." P26's filing system was built by a technician with Claude, "but it's my work conceptually." The literal-authorship view (ownership = typing the text) appears only among the 18–24 coders, P31 and P41.
P19 states it directly: ownership "comes less from who generated the words and more from how much judgment, thought, and discernment I put into shaping the final result," and she calibrates that investment ("C work" when energy is low, "A+" when she goes line by line). P12, P13, P9, P20, P26, P28, P29 and P38 describe the same mechanism through iteration count and rewriting.
Where the person drafted first and used AI to refine, ownership held (P9, P25, P31, P35, P37). Where they handed AI an idea and asked it to generate, voice mismatch followed and the task was disowned (P10 flyer, P24 and P27 LinkedIn posts, P37 party letter, P40 essay conclusion). This is a behavioral lever rather than an attitude.
Bad AI output clarifies what the person wants, and they own the corrected result more strongly. P10 felt more ownership of the task where "AI failed so horribly." P23 used generic output "to decide what I didn't like in order to identify what I did." P7 went "180" against AI's landing-page advice. P40 rejected the ChatGPT conclusion and wrote her own.
Several people sorted tasks before AI entered. P14: a task "I wanted to have ownership of" versus one "I would have rather not have had to do in the first place." P34: self-directed research versus committee duty ("like eating vegetables"). P35: a task "consciously thought of as creative" versus one where productivity was all that mattered. Obligation lowers ownership independent of AI.
When the respondent cannot evaluate the output, they disclaim it even when effort was high. P32 cites "the difference in competence between me and the AI" and does not feel responsible for a working website. P38 feels less ownership of a metadata-merge task she "could not have easily gotten my head around without AI." P12 left statistical method choice to AI and says so.
A subgroup of heavy users place ownership in the harness, context folder, trained agent or interaction rules they built, not in any single output. P1's "coral reef" of curated opinions; P11's "AI OS" and "second brain"; P21's project brief instructing Claude to offer no creative suggestions and to call her the "founding voice"; P25's self-designed Copilot modes; P39's AI "trained with quite a bit of writing." For these users, per-task ownership questions were partly meaningless.
Two types. Tool-instrumentalists say AI is a calculator and identity is untouched (P3, P8, P18, P28, P33). P1 is a different case: he does not want output that reflects only him ("'I' do not need to be purely expressed"). P3 goes further and wants AI to be unlike him: "Diversity is where the power comes from."
P14 is "very proud" of a refund he does not own. P32 disclaims both ownership and responsibility for a site that works. P28 owns the eulogy more than the genealogy but is proud of neither. P13 says outcome matters and process does not. Ownership as measured here is not the same construct as pride or accountability.
Most say they would disclose, with qualifiers. P9 names a stigma "that I didn't do any work." P16 worries "people will hear delegation rather than oversight." P41 would tell coworkers but not Tumblr. P37 would be "ashamed" about the party letter. P12 and P20 quietly do not disclose emails and captions. P5 fears a "pile on" from colleagues. Anxiety is highest where output is in the person's own voice and the audience is non-technical.
P40: "I was just a little lazy. I always admire other people who make things a little harder for themselves." P36: "sad that i didn't have the patience to do it myself." P41 enjoys debugging less because Claude now diagnoses problems. P4 notices dependence forming. Absent among the 45–64 professional majority; P28 (65+) voices it as a general worry.
"Slop," "vanilla," "pablum," "generic," "artificial," "over aggregated" recur. P7 reports "AI sickness" and has moved to "hyper simplification." P38 names the tells (comma-separated triplet of benefits, a repetitive wrap-up sentence) and strips them. Loss of ownership is experienced as the output sounding like everyone else.
Grouping by the logic respondents used, not by demographics. Membership is from free text and is provisional; some people fit two.
Ownership is proportional to revision, questioning and rewriting. Several scale effort to stakes deliberately. Mostly 45–64 professionals. Disclose readily but notice stigma.
Ownership lives in the harness, corpus or rules they built. Per-output questions feel beside the point. Heaviest users and highest spenders in the sample. Comfortable disclosing; several teach others.
AI is a calculator or search engine; identity is not in play. Report no difference between tasks or reject the question. Skew older and male. P3 actively wants AI to be unlike him.
Ownership tracks whether they typed it or understood the code. All 18–24 students. Most express some regret about skipped learning, and want to disclose with an explanation so their contribution is not discounted.
Ownership is whether it sounds like them. Disowned tasks are almost all own-voice writing that came back generic. Often recover ownership by rejecting the output and writing themselves.
Rare or non-users (P5, P17, P30), one-sided responders (P2, P4), or a damaged record (P15). P34 is a clear case of obligation-versus-volition rather than any of the above.
Points in this data most likely to be additive or corrective to an existing ownership framework.
Self-selected, expert, researcher-adjacent sample (n=41), with heavy attrition on the paired design. The dichotomous prompt was rejected or subverted by about 14 people, so within-person contrasts are only partially valid. Compound questions (flagged by P11) make it impossible to cleanly separate pride, disclosure and outcome satisfaction. There is no ownership scale; every quantitative statement rests on the two counterfactual items and single-coder hand-coding of free text. Encoding damage in the export (curly quotes, en dashes) was repaired before analysis and did not affect counts.