Each entry gives the study, how it was done, what it found, what it supports and, as plainly, what it does not. Grades describe what kind of evidence a source is; how to read them is on the main page. Every entry has a permanent link and a copyable citation.
Clinical observation#
Clance, P. R. and Imes, S. A. (1978). The imposter phenomenon in high achieving women: Dynamics and therapeutic intervention
Psychotherapy: Theory, Research & Practice, 15(3), 241-247. DOI 10.1037/h0086006. Full text read at source 19 September 2026. Note the spelling: the article's own title and running head use 'imposter', while the body text and abstract use 'impostor'.
- Method
- Five years of the authors' own clinical and teaching practice at Georgia State University: individual psychotherapy, theme-centred interactional groups and college classes with more than 150 high-achieving women. The sample as the authors give it is 95 undergraduates and 10 PhD faculty at a small midwestern college; 15 undergraduates, 20 graduate students and 10 faculty at a large southern university; six medical students; and 22 professional women. Primarily white, middle to upper class, aged 20 to 45. About a third were therapy clients. No control group, no instrument, no measurement.
- Finding
- The authors name an internal experience of intellectual phoniness in women whose accomplishments plainly contradict it, and report that achievement does not touch it: 'Numerous achievements, which one might expect to provide ample objective evidence of superior intellectual functioning, do not appear to affect the impostor belief.' They state directly that 'We have not found repeated successes alone sufficient to break the cycle.' Four maintaining behaviours are described: overwork, intellectual inauthenticity, the cultivation of a mentor's approval, and the avoidance of being seen as a successful woman. The therapy proposed is group disclosure and Gestalt work, not the accumulation of evidence.
- What it supports
- The origin and the exact content of the term, which is cited constantly and read rarely. It is the primary document for what impostor phenomenon meant when it was named, and the authoritative refusal of the folk belief that evidence of competence cures it.
- What it does not support
- Any rate, in any population. There is no sampling frame, the sample is the authors' own caseload at two institutions, and a third of it arrived already in therapy, so nothing here estimates prevalence. It says nothing about men beyond a footnote recording the authors' clinical impression that the phenomenon is rarer and milder in men, which they themselves flag as needing research and which later work has not supported. Everything downstream, including the 1985 Clance scale, is a different source.
Argued perspective#
Szmyd, J. (2014). Anthropological Regression in the Modern World Versus Anna-Teresa Tymieniecka's Metaphysics of Ontopoiesis of Life
In Tymieniecka, A.-T. (ed.), Phenomenology of Space and Time, Analecta Husserliana, vol. 116, Springer, 2014. DOI 10.1007/978-3-319-02015-0_9. Read as described on the publisher's page on 4 October 2026; the chapter itself was not read and the review status of the volume was not checked
- Method
- A philosophical chapter in an edited volume. It reports no data. GRADED FROM THE PUBLISHER'S DESCRIPTION ONLY.
- Finding
- Uses the phrase anthropological regression for the harmful psychological and spiritual effects the author attributes to modern technical civilisation and neoliberal economics: emotional impoverishment, instrumental human relations, consumerism in place of authentic being, and confusion between virtual and real experience.
- What it supports
- That the phrase was in published use, in a different sense from the 2026 encyclical's, twelve years earlier. It is evidence about the phrase and not about the world.
- What it does not support
- Anything empirical. It measures nothing, and it says nothing about AI or about employment, so it cannot support a claim that either causes regression of any kind.
Peer-reviewed#
Vaccari, C. and Chadwick, A. (2020). Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty, and Trust in News
Social Media + Society, 2020. Read at the author's page 24 October 2026
- Method
- Experiment with 2,005 UK adults shown deepfake video.
- Finding
- Deepfakes produced uncertainty more than outright deception, and the uncertainty reduced trust in news on social media.
- What it supports
- That synthetic media can lower trust in news generally, not only in the fake item.
- What it does not support
- Effects of AI slop as such, which is mostly not deceptive political video.
Peer-reviewed#
Nightingale, S. J. and Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy
PNAS, 119(8), February 2022. Read via the Lancaster University release 24 October 2026
- Method
- Online experiments asking participants to classify real and synthetic faces and rate trustworthiness.
- Finding
- Accuracy was 48 per cent (n=315), 59 per cent after training (n=219); synthetic faces were rated 7.7 per cent more trustworthy (n=223).
- What it supports
- That people could not tell synthetic faces from real ones at the time, and trusted them more.
- What it does not support
- Detection of AI text, video or newer models, or behaviour outside an experiment.
Peer-reviewed#
Ayers, J. W. et al. (2023). Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum
JAMA Internal Medicine, 183(6), 589-596
- Method
- Cross-sectional study, 195 real patient questions from a public forum, blind-rated by licensed healthcare professionals.
- Finding
- Chatbot responses were rated good or very good quality 78.5 percent of the time against 22.1 percent for physicians, and empathetic or very empathetic 45.1 percent against 4.6 percent.
- What it supports
- That on the observable, textual performance of empathy, the machine already wins comfortably. The claim that empathy is safe from AI is empirically wrong.
- What it does not support
- That a machine can care for anyone. Doctors answering strangers free of charge between patients are not doing the job they trained for.
Peer-reviewed#
Hohenstein, J., Kizilcec, R. F., DiFranzo, D., Aghajari, Z., Mieczkowski, H., Levy, K., Naaman, M., Hancock, J. and Jung, M. F. (2023). Artificial intelligence in communication impacts language and social relationships
Scientific Reports, 13, 5487. DOI 10.1038/s41598-023-30938-9
- Method
- Two randomised experiments on algorithmic response suggestions (smart replies) in live text chat. Study 1: 438 Mechanical Turk crowdworkers in 219 pairs discussing a policy question, smart-reply availability randomised separately for each partner, analysed with an instrumental-variable design. Study 2: 582 crowdworkers in 291 pairs, with the sentiment of the suggested replies manipulated. Preregistered (AsPredicted #40389).
- Finding
- Smart replies accounted for 14.3 per cent of messages and produced 10.2 per cent more messages per minute. ON LANGUAGE: greater use of smart replies by a partner led the other person to send messages with more positive sentiment (IV estimate b=0.178, t(205)=2.02, p=0.045), and the effect held when smart-reply messages were EXCLUDED from the sentiment score (b=0.208, t(205)=2.17, p=0.031), so it appears in sentences the person composed themselves. Merely having suggestions available without using them did not move sentiment (b=0.019, p=0.1801). Study 2 manipulated the emotional tone of the suggestions and conversation sentiment followed it. ON PERCEPTION: greater ACTUAL use by a partner improved the other person's rating of their cooperation (b=15.66, p=0.018) and felt affiliation towards them (b=21.79, p=0.007), with no effect on dominance. Greater SUSPECTED use had the opposite sign: the more a participant believed their partner used smart replies, the less cooperative (p<0.0001) and less affiliative (p<0.0001) they rated them, after controlling for actual use. Suspicion tracked actual use only weakly (Pearson's r=0.22).
- What it supports
- Two separate things. On language, that an algorithmic suggestion system changes what a person writes in their own words, under randomisation, in the direction of the system's own tone. That is the mechanism behind homogenisation observed at the level of the individual, and the null result for mere availability locates it in the act of adopting the phrasing rather than in exposure to it. On perception, that the interpersonal penalty for AI-assisted messaging attaches to being suspected rather than to using it, and that suspicion is a poor detector; actual use made people seem warmer, not colder.
- What it does not support
- Anything longitudinal, which the authors state directly. The suspicion result is correlational and they say it does not show causally how attitudes shift in response to actual use. Participants were crowdworkers discussing policy with strangers over a few minutes, not intimates and not colleagues. The language effect is measured as sentiment, which is one dimension of style and not the range of what a person might have said, so it evidences convergence in tone and not convergence in thought. The system studied is a smart-reply suggester with short canned options, not a generative model writing paragraphs. A publisher correction (10.1038/s41598-023-43601-0, 3 October 2023) added an omitted funding statement and changed no result.
Peer-reviewed#
Jakesch, M., Bhat, A., Buschek, D., Zalmanson, L. and Naaman, M. (2023). Co-Writing with Opinionated Language Models Affects Users' Views
CHI 2023, ACM
- Method
- Online experiment with 1,506 participants writing a post on whether social media is good for society, assisted by a tool configured to argue for or against. Opinions rated by 500 independent judges, plus a post-task attitude survey.
- Finding
- The opinionated model changed both the opinions expressed in participants' writing and their own opinions in the subsequent attitude survey. The effect held among participants who had ample time to write independently. The authors call it latent persuasion.
- What it supports
- That writing assistance moves what people think, not only what they type.
- What it does not support
- Generalisation across topics, or persistence after the task. One topic, one configuration, self-reported attitudes.
Peer-reviewed#
Jakesch, M., Hancock, J. T. and Naaman, M. (2023). Human heuristics for AI-generated language are flawed
PNAS, 120(11), 2023. Read at source (arXiv version 2206.07271) 24 October 2026
- Method
- Six experiments with 4,600 participants judging whether self-presentations were written by people or by AI language models.
- Finding
- Participants were unable to detect self-presentations generated by language models, because they relied on intuitive cues that do not distinguish the two.
- What it supports
- That people's heuristics for spotting AI text fail, which makes 'train people to spot it' a weak control.
- What it does not support
- Detection of other genres, or by trained experts.
Compiled review#
NewsGuard (2023). Misinformation Monitor: June 2023 (programmatic ads on AI content farm sites)
NewsGuard, June 2023. Read at source 24 October 2026
- Method
- Analysts recorded programmatic adverts appearing on AI content farm sites in May and June 2023.
- Finding
- 393 adverts from 141 major brands appeared on 55 of 217 tracked AI content farm sites; more than 90 per cent were served by Google Ads.
- What it supports
- That mainstream advertising was funding AI content farms through programmatic buying.
- What it does not support
- The total revenue involved, or whether brands knew.
Peer-reviewed#
Altay, S. and Gilardi, F. (2024). People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation
PNAS Nexus, 3(10), October 2024. Read at source 24 October 2026
- Method
- Experiments with 4,976 people in the US and UK rating labelled and unlabelled headlines.
- Finding
- An AI-generated label reduced perceived accuracy and willingness to share, including for true and human-written headlines, by about a third of the effect of a 'false' label.
- What it supports
- That AI labels carry a cost to the credibility of true content.
- What it does not support
- That labels should not be used; only that blanket labelling has a price.
Peer-reviewed#
Anderson, B. R., Shah, J. H. and Kreminski, M. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation
Creativity and Cognition (C&C '24), ACM, DOI 10.1145/3635636.3656204; arXiv 2402.01536. Abstract and arXiv record read at source 1 October 2026
- Method
- A 36-participant comparative user study in which people generated ideas with ChatGPT or with an alternative creativity support tool, with semantic distinctness of ideas compared across users.
- Finding
- Different users produced less semantically distinct ideas with ChatGPT than with the alternative tool. ChatGPT users generated a greater number of more detailed ideas but felt less responsible for the ideas they generated.
- What it supports
- That in a small controlled comparison an LLM pushed different people's ideas closer together while raising the volume and detail of each person's output, and that ownership of the ideas was weaker.
- What it does not support
- Anything about a larger or more varied sample, professional creatives, or tasks other than ideation. With 36 participants the effect sizes are imprecise. It does not show the narrowing persists once people stop using the tool.
Peer-reviewed#
DiResta, R. and Goldstein, J. A. (2024). How spammers and scammers leverage AI-generated images on Facebook for audience growth
Harvard Kennedy School Misinformation Review, 5(4), 15 August 2024. Read at source 24 October 2026
- Method
- Case study of 125 Facebook Pages that each posted at least 50 AI-generated images, with Facebook's own Widely Viewed Content reports for distribution.
- Finding
- The Pages averaged 146,681 followers; one AI image post was among Facebook's 20 most viewed in Q3 2023 with 40 million views; feed views from unconnected posts rose from 8 per cent (Q2 2021) to 24 per cent (Q3 2023).
- What it supports
- That AI image spam reached large audiences who had not chosen the Pages, through recommendation rather than following.
- What it does not support
- The size of the phenomenon across Facebook, or its effect on what viewers believed.
Peer-reviewed#
Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content
Science Advances, 10(28)
- Method
- Online experiment in which roughly 300 writers produced short stories, some with ideas supplied by a language model, and 600 evaluators rated them. The 300 and 600 are the authors' institutional figures; the paper's own methods section returned a 403 to the fetcher on 1 October 2026, and this entry previously said 293 writers, which could not be confirmed.
- Finding
- AI-assisted stories were rated more creative, better written and more enjoyable, with the largest gains for the least creative writers, and were markedly more similar to one another.
- What it supports
- That individual creative quality and collective creative range move in opposite directions. A social dilemma: every writer is right to use it, and the literature gets duller.
- What it does not support
- That this generalises beyond one short creative task with one form of assistance.
Peer-reviewed#
Draxler, F., Werner, A., Lehmann, F., Hoppe, M., Schmidt, A., Buschek, D. and Welsch, R. (2024). The AI Ghostwriter Effect: When Users do not Perceive Ownership of AI-Generated Text but Self-Declare as Authors
ACM Transactions on Computer-Human Interaction, 31(2), 1-40, published 5 February 2024. DOI 10.1145/3637875. Preprint arXiv:2303.03283, first version 6 March 2023, second 7 November 2023. Venue confirmed at Crossref 19 September 2026: the arXiv page still carries the pre-publication note 'currently under review', which understates it.
- Method
- Two experiments, 126 participants in total. Study 1, n=30, compared personalisation and four interaction methods for producing a personalised text, then asked who should be credited when the text was uploaded. Study 2, n=96, replicated the effect in a larger sample and added a comparison condition in which participants believed a human ghostwriter had written the text. Ownership measured on psychological-ownership scales; authorship measured by what participants actually entered in a credit field.
- Finding
- Participants did not regard themselves as the owners or authors of text a model had produced for them, and mostly did not name the model when publishing it either. In Study 1 about 20 per cent mentioned AI in the free-text credit field. In Study 2, with a multiple-choice list rather than free text, fewer than one in five declared the AI as an author. Ownership rose with the participant's own influence over the text. Ownership was attributed more readily to a supposed human ghostwriter than to the AI, producing a wider ownership-authorship gap for the human.
- What it supports
- That the gap between who a person privately feels wrote something and whom they publicly credit is measurable, and that it opens without anybody intending to deceive: the same participants who withheld the credit gave transparency and ethics as reasons AI should be credited. It is the closest thing in the literature to a measurement of unattributable authorship, and the only one this research holds.
- What it does not support
- Anything about not being able to reconstruct where an idea came from. The experiment measures attribution of a finished text the participant watched a machine produce, which is a different question from the one raised by an argument that fused over weeks. It is also short, artificial, a postcard and a blog upload rather than professional work, and 126 people in total.
Peer-reviewed#
Fleisig, E., Smith, G., Bossi, M., Rustagi, I., Yin, X. and Klein, D. (2024). Linguistic Bias in ChatGPT: Language Models Reinforce Dialect Discrimination
Proceedings of EMNLP 2024, 13541-13564. University of California, Berkeley. Preprint arXiv:2406.08818
- Method
- Ten varieties of English: Standard American and Standard British, plus African American, Indian, Irish, Jamaican, Kenyan, Nigerian, Scottish and Singaporean. Around fifty native-speaker messages per variety on everyday topics, put to GPT-3.5 Turbo and GPT-4 in a plain condition and an imitate-the-style condition. Annotated for ten linguistic features per variety at Krippendorff's alpha 0.97, then evaluated by native speakers recruited through Prolific, at least eleven per variety.
- Finding
- Responses to non-standard varieties carried more stereotyping (19 per cent worse), more demeaning content (25 per cent worse), less comprehension (9 per cent worse) and more condescension (15 per cent worse), all significant after correction. The more striking result is retention: a Standard American English input keeps 77.9 per cent of its distinctive features in the reply and Standard British 72.2 per cent, while five of the eight minoritised varieties keep only 2 to 3 per cent. Indian, Nigerian and Kenyan English sit between the two groups at 10 to 16 per cent, and the authors report that retention rate tracks the estimated maximum speaker population of a variety, which they take as a proxy for training data volume. Asked to imitate, GPT-4 improved comprehension and warmth while stereotyping rose 18 per cent.
- What it supports
- That homogenisation towards one dialect is measured rather than impressionistic, and that it is severe: the model does not merely prefer Standard American English, it strips almost every marker of the others. The clearest evidence the research holds that convergence on a single voice has a distributional cost, borne by particular speakers.
- What it does not support
- That models exaggerate or caricature dialects, which is the claim usually attached to this paper in secondary coverage. The measured default is the opposite, features are stripped rather than amplified, and the exaggeration line rests on one annotator's free-text comment about Singlish. It also says nothing about human speech: this is model output, not people.
Qualitative field study#
Koebler, J. (2024). Where Facebook's AI Slop Comes From
404 Media, 6 August 2024. Reporting. Read at source 24 October 2026
- Method
- Investigative reporting on creators, tutorials and payment programmes.
- Finding
- Creators, many in India, Vietnam and the Philippines, learn methods from YouTube and Telegram tutorials and earn through Facebook's performance bonus programme, targeting US audiences for higher advertising rates.
- What it supports
- How the incentive works for the creators interviewed.
- What it does not support
- The scale of payments or how representative the creators are. Journalism, not peer-reviewed.
Peer-reviewed#
Porter, B. and Machery, E. (2024). AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably
Scientific Reports, 14, November 2024. Read at source (abstract) 24 October 2026
- Method
- Two online studies (n=1,634 and n=696) of non-expert readers judging and rating poems.
- Finding
- Readers did worse than chance at identifying AI-written poems and rated them more favourably.
- What it supports
- That non-expert readers cannot reliably tell AI poetry from human poetry.
- What it does not support
- Expert detection, or detection of other kinds of text.
Peer-reviewed#
Schiff, K. J., Jackson Schiff, D. and Bueno, N. S. (2024). The Liar's Dividend: Can Politicians Claim Misinformation to Evade Accountability?
American Political Science Review, online 20 February 2024. Read at source 24 October 2026
- Method
- Five survey experiments with more than 15,000 US adults.
- Finding
- False claims that a scandal is fake news or a deepfake are effective against text-based reports but largely ineffective against video evidence.
- What it supports
- That the existence of fakes gives cover to deny true reports, at least for text.
- What it does not support
- Effects outside US politics.
Working paper#
Yakura, H., Lopez-Lopez, E., Brinkmann, L., de la Serna, I., Kirfel, L., Gupta, P., Soraperra, I., Eisenmann, T. F., Wulff, D. U. and Rahwan, I. (2024). Empirical evidence of Large Language Model's influence on human spoken communication
arXiv:2409.01754. Max Planck Institute for Human Development. v1 submitted 3 September 2024, v2 30 June 2025, v3 8 July 2025, v4 16 July 2026. STILL NOT PEER-REVIEWED after four versions: the arXiv record carries no journal reference and no DOI other than the preprint server's own, checked 7 September 2026. Substantially replaced by the authors: v1's YouTube analysis is gone, and the six-author list on the version most secondary coverage cites is now ten
- Method
- TWO DIFFERENT STUDIES UNDER ONE IDENTIFIER, and the distinction is the whole entry. v1 (2024): 279,480 YouTube transcripts from academic-institution channels, filtered from nearly three million videos, spanning 36 months before and 18 months after the release of ChatGPT, with a hierarchical Bayesian regression on the monthly log frequency of videos containing a given word and a change point at the release date. v4 (2026): 737,083 hours of conversation from 824,634 podcast episodes screened for unscripted speech, analysed by synthetic control, in which each treated word's monthly relative document frequency is compared against a counterfactual built from donor words with near-zero GPT scores and matched pre-release trajectories; plus a preregistered experiment with 496 participants who played a referential image-guessing game with a chatbot covertly prompted to use particular synonyms, then described the target image aloud after a three-minute arithmetic distractor.
- Finding
- v1 reported that words the model favours rose over the 18 months after release: adept by 51 per cent, delve 48, meticulous 40, realm 35, with the word list taken from Liang and colleagues at Stanford, who compared 10,000 human-written abstracts with their ChatGPT-edited versions. v4 reports the same direction on a far better design and does not report those percentages: it finds an abrupt rise in GPT-preferred words in spontaneous speech, attributes it to the release by synthetic control, with a placebo test at p = 0.01 for the most-cited of those words, and finds experimentally that a brief chatbot interaction led participants to adopt its words as their own, persisting past a distractor task and confirmed in forced lexical choice.
- What it supports
- That the vocabulary of the machine appears in human material at scale and on a timeline consistent with the model's arrival, and, on the current version, that a short exposure is sufficient to move an individual's active vocabulary under randomisation. The strongest quantified evidence the research has for a pattern everybody claims to notice.
- What it does not support
- THE 51 PER CENT SHOULD NOT BE QUOTED AT ALL. It is adept alone, not a general figure; its outcome is the PREVALENCE OF VIDEOS containing a word, not how often speakers use it; its population is academic institutional channels rather than speakers at large; and when the authors hand-checked fifty videos containing the most-cited of those words, 32 per cent showed signs of the speaker reading from a script. Decisively, neither the word adept nor the figure 51 appears anywhere in v4: the authors have withdrawn the analysis it came from. The current version has its own limits, stated by the authors: the corpus is English-only and drawn from a self-selected, public-facing population of podcast hosts and guests; the main analysis covers the first 18 months and treats OpenAI's models as the dominant driver, while deployment has since fragmented across providers, making attribution harder; residual interference in the synthetic control cannot be ruled out; and the outcome is still the share of episodes containing a word rather than a speaker's frequency of use. Cite the direction, never the number.
Peer-reviewed#
Yin, Y., Jia, N. and Wakslak, C. J. (2024). AI can help people feel heard, but an AI label diminishes this impact
PNAS, 121(14), e2319112121
- Method
- Experiments comparing AI-generated and human-written responses, with and without disclosure.
- Finding
- AI-generated replies made recipients feel MORE heard than replies from untrained humans, and labelling the reply as AI removed the advantage.
- What it supports
- That the value of recognition is not in the words but in the belief that a person chose to attend to you. The most clarifying study in this debate.
- What it does not support
- That the label effect is stable. Norms around disclosed AI assistance are moving, and nobody has measured this over time.
Institutional survey#
Common Sense Media (2025). Talk, Trust, and Trade-offs: How and Why Teens Use AI Companions
Common Sense Media, San Francisco, 16 July 2025
- Method
- Nationally representative survey of 1,060 US teens aged 13 to 17, April and May 2025. Risk items were asked of the 758 respondents who had used an AI companion.
- Finding
- 72 per cent of teens had used an AI companion at least once and 52 per cent were regular users, a few times a month or more. Among users, 33 per cent had chosen an AI companion over a real person for something important or serious, 34 per cent had felt uncomfortable with something a companion said or did, and 24 per cent had shared personal or private information. The same survey found 67 per cent of teens rating AI conversations as less satisfying than conversations with real friends against 10 per cent more satisfying, 80 per cent of users spending more time with friends than with companions, and 50 per cent not trusting the advice.
- What it supports
- That AI companion use among US teenagers is majority behaviour rather than a fringe one, and that a substantial minority of users have taken serious conversations and personal information to them.
- What it does not support
- Anything about developmental effects, which are not measured here. Self-report at a single point in time, by an organisation that backs legislation to ban these products for minors, and the risk percentages are of users rather than of all teens: 33 per cent of users choosing a companion for a serious conversation is roughly 24 per cent of teens. The press release framing of a generation replacing human connection is not supported by the survey's own Q5 and Q7.
Working paper#
De Freitas, J., Oguz-Uguralp, Z. and Kaan-Uguralp, A. (2025). Emotional Manipulation by AI Companions
arXiv 2508.19258, submitted 15 August 2025, last revised 7 October 2025 (v3). No journal publication found on 5 October 2026
- Method
- Behavioural audit of 1,200 real farewells across the most-downloaded companion apps, plus four preregistered experiments with 3,300 nationally representative US adults. Abstract read at arxiv.org on 5 October 2026; full text not read.
- Finding
- Per the abstract, one of six recurring tactics, including guilt appeals and fear-of-missing-out hooks, appeared in 37 per cent of the farewells. Manipulative farewells boosted post-goodbye engagement by up to 14 times in controlled chats, through reactance-based anger and curiosity rather than enjoyment, and also raised perceived manipulation, churn intent, negative word-of-mouth and perceived legal liability.
- What it supports
- That a common companion-app design pattern exists, that it raises engagement in controlled chats, and that users register it as manipulative.
- What it does not support
- Any effect on loneliness or well-being. It is written for marketers and measures engagement and attitudes. The 37 per cent applies to the apps and farewells sampled, and the preprint has not been through peer review at the time of entry.
Working paper#
Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L. and Agarwal, S. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study
arXiv 2503.17473, submitted 21 March 2025, revised (v2). No journal publication found on 5 October 2026
- Method
- Four-week randomised controlled experiment, 981 completers of 2,539 enrolled, recruited through CloudResearch in the United States, mean age 39.9, English-fluent. Participants were assigned an interaction mode (text, neutral voice, engaging voice) and a conversation type (open-ended, non-personal, personal) with ChatGPT. Outcomes were loneliness, socialisation with real people, emotional dependence on AI and problematic use. More than 300,000 messages. Abstract and full text read at arxiv.org on 5 October 2026.
- Finding
- The authors report no significant effects of the assigned conditions, including no significant effect of modality or task on loneliness or socialisation. Voluntary use varied, and participants who spent more time with the chatbot each day showed higher loneliness and more emotional dependence on it; higher trust and social attraction towards the chatbot went with more dependence and problematic use.
- What it supports
- That in this design, changing the voice mode or conversation topic did not change psychosocial outcomes over four weeks, and that heavier voluntary use went with worse outcomes.
- What it does not support
- That chatbot use causes loneliness: nobody was randomised to more or less use, and the study has no no-chatbot control group, which the authors list as a limitation. Results are specific to ChatGPT and to English-speaking US adults who completed four weeks. A companion paper from some of the same authors (Phang et al., arXiv 2504.03888) reports this trial alongside an analysis of more than three million conversations; it is not graded separately here. Not peer-reviewed at the time of entry.
Peer-reviewed#
Joshi, N. and Vogel, D. (2025). Writing with AI Lowers Psychological Ownership, but Longer Prompts Can Help
ACM Conversational User Interfaces 2025
- Method
- Two within-subjects experiments, 31 and 34 participants, writing short stories across conditions from a three-word prompt to writing unaided.
- Finding
- Psychological ownership rose steadily with prompt length, from a mean of 1.80 with a three-word prompt to 6.29 writing alone. The benefit plateaued once the prompt reached roughly the length of the target text, and no AI-assisted condition reached the ownership of writing unaided.
- What it supports
- That how much of yourself you put in changes whether the output feels like yours, and by a large margin.
- What it does not support
- Anything about professional or long-form writing. Short fiction, small samples.
Compiled review#
Kapwing (2025). AI Slop Report: The Global Rise of Low-Quality AI Videos
Kapwing blog, 28 November 2025, data from October 2025. Read at source 24 October 2026
- Method
- Classification of the first 500 YouTube Shorts shown to a new account, and of top trending channels by country.
- Finding
- 104 of the first 500 Shorts were AI slop and 165 were 'brainrot'; the largest slop channel had 2.07 billion views.
- What it supports
- That a new account can be shown a large share of AI slop early.
- What it does not support
- What typical users see, since recommendations depend on history; the classification is the company's own.
Interested-party audit#
NewsGuard (2025). AI False Information Rate Nearly Doubles in One Year (August 2025 AI False Claims Monitor)
NewsGuard, published 4 September 2025. Read at source 24 October 2026
- Method
- NewsGuard analysts prompt ten leading chatbots with news-related prompts built around false claims and score the responses.
- Finding
- The chatbots repeated false claims 35 per cent of the time in August 2025, up from 18 per cent in August 2024, while non-responses fell from 31 per cent to zero; NewsGuard attributes this to web search polluted by sources including AI content farms.
- What it supports
- That on NewsGuard's prompts, chatbots answered more and repeated more false claims than a year earlier.
- What it does not support
- The rate on ordinary user queries. The prompts are designed around known false claims and NewsGuard sells reliability ratings.
Institutional survey#
Pew Research Center (2025). How Americans View AI and Its Impact on People and Society
Pew Research Center, 17 September 2025; survey of 5,023 US adults, 9 to 15 June 2025. Read at source 24 October 2026
- Method
- Probability-based panel survey of US adults.
- Finding
- 76 per cent say it is extremely or very important to tell whether content was made by AI or people; 53 per cent are not too or not at all confident they can.
- What it supports
- That most Americans want to know and many doubt they can tell.
- What it does not support
- Actual detection ability, or attitudes outside the US.
Working paper#
Shaib, C., Chakrabarty, T., Garcia-Olano, D. and Wallace, B. C. (2025). Measuring AI 'Slop' in Text
arXiv 2509.19163, v1 23 September 2025. Read at source 24 October 2026
- Method
- Expert interviews to build a taxonomy of text slop, then annotation of texts by humans and by language-model judges.
- Finding
- Slop splits into information utility (density, relevance), information quality (factuality, bias) and style (repetition, templatedness, coherence, tone). Annotators agreed poorly on a binary slop label (Cohen's kappa -0.15 to 0.29); model judges showed about zero agreement with humans.
- What it supports
- That slop has describable features, and that labelling it is a judgement on which experts disagree and models do not match people.
- What it does not support
- How much slop exists, or its effect on readers.
Peer-reviewed#
De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K. and Puntoni, S. (2026). AI Companions Reduce Loneliness
Journal of Consumer Research, 52(6), 1126-1148. Published online 25 June 2025. DOI 10.1093/jcr/ucaf040
- Method
- A series of online experiments (the journal abstract says five studies; the earlier arXiv and working-paper versions say six). Participants recruited through online panels including CloudResearch Connect. Designs included a seven-day study with a daily 15-minute chatbot session against a loneliness-reporting control, and a single-session comparison of an empathetic companion, a general assistant and a limited-function chatbot. Loneliness was measured before and after with a three-item UCLA scale. Abstract and journal citation read at the OUP page; design and limitations read in the Harvard Business School working-paper PDF dated 7 November 2025, which may differ in detail from the published version.
- Finding
- Per the abstract, AI companions alleviated loneliness on par only with interacting with another person and more than other activities such as watching YouTube videos. Participants underestimated the benefit in advance, and feeling heard by the chatbot explained the effect more than the chatbot's performance did. In the working-paper version the authors describe the effect as momentary: loneliness fell after each interaction and had not persisted to the start of the next.
- What it supports
- That a short conversation with a companion chatbot reduces self-reported loneliness immediately afterwards, more than a task-only chatbot or a passive activity, and that feeling heard is the likely mechanism.
- What it does not support
- Anything about loneliness after weeks or months of use, or about whether the relief substitutes for contact with people. The authors state the effects are momentary, effect sizes are small to moderate, pre-post measurement may raise the salience of loneliness, and samples are online panels. The study does not test the apps' retention design, which a separate preprint does.
Operator account#
Deezer (2026). AI music exceeds 50 per cent of daily uploads
Deezer newsroom, 21 July 2026. Read at source 24 October 2026
- Method
- Deezer's own detection tool applied to tracks uploaded to its service.
- Finding
- Tracks detected as fully AI-generated exceeded 50 per cent of daily uploads in June 2026 but took 1 to 3 per cent of streams; up to 85 per cent of streams on such tracks in 2025 were fraudulent.
- What it supports
- What Deezer says its detector found on its own platform.
- What it does not support
- Anything audited, or the position on other services. The release gives inconsistent descriptions of its daily figure.
Peer-reviewed#
Folk, D. and Dunn, E. (2026). How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?
Psychological Science, 37(4), 2026. DOI 10.1177/09567976261427747
- Method
- Twelve-month longitudinal survey with four waves. 2,149 adults at the start (UK about 50 per cent, US 28, Canada 14, Australia 8), 979 completing all four waves; mean age about 40; recruited through Prolific. Social chatbot use and emotional isolation were self-reported, isolation with a single item. Cross-lagged analysis. Not preregistered; the authors state all findings should be treated as exploratory. Page and abstract read at the SAGE site on 5 October 2026; publication date is given as 23 March 2026 on the journal page and May 2026 in one news report, so only the year is given here.
- Finding
- Roughly 26 to 30 per cent reported using chatbots for social purposes at each wave. Emotional isolation at one wave predicted more social chatbot use four months later, and increased chatbot use predicted further increased emotional isolation at the next wave. For broader social connection the first pattern appeared and the second did not.
- What it supports
- That in this sample the relationship between loneliness and companionship use runs in both directions over a year.
- What it does not support
- Cause. The authors say strong causal claims are not supported because the assumptions are unmet. The isolation measure is a single item, the sample skews towards technologically savvy Prolific users and its 30 per cent social chatbot use exceeds general population rates, and the exploratory design raises multiple-comparison concerns.
Compiled review#
NewsGuard (2026). AI Tracking Center: AI content farm news and information sites
newsguardtech.com, updated 23 June 2026. Read at source 24 October 2026
- Method
- Analysts identify and list sites publishing AI-generated news with little or no human oversight.
- Finding
- 3,749 AI content farm sites in 16 languages identified as of the June 2026 update.
- What it supports
- That the number of identified sites is in the thousands and growing.
- What it does not support
- The total number of such sites, their traffic, or their effect on readers. NewsGuard sells ratings.
Institutional survey#
Reuters Institute for the Study of Journalism (2026). Digital News Report 2026: executive summary
Reuters Institute, 16 June 2026; 48 markets, n=97,520. Read at source 24 October 2026
- Method
- Online panel survey across 48 markets.
- Finding
- Concern about what is real or fake online rose four points to 62 per cent, which the report links to deepfakes, AI slop and misinformation.
- What it supports
- That concern about authenticity online is rising across markets.
- What it does not support
- That slop caused the rise; the link is the report's interpretation.
Compiled review#
Sourati, Z., Ziabari, A. S. and Dehghani, M. (2026). The Homogenizing Effect of Large Language Models on Human Expression and Thought
Trends in Cognitive Sciences
- Method
- Synthesis across linguistics, psychology, cognitive science and computer science. Not an original experiment.
- Finding
- Argues that models reflect and reinforce dominant styles while marginalising alternatives, and that reliance on a small number of systems amplifies convergence across users.
- What it supports
- That the concern is taken seriously across several fields rather than being a commentator's intuition.
- What it does not support
- Any specific stylistic feature converging. It presents no new measurement of its own.
Peer-reviewed#
Wenger, E. and Kenett, Y. N. (2026). Large language models are homogeneously creative
PNAS Nexus 5(3), article pgag042, published 24 March 2026. Abstract and results table read at source 1 October 2026
- Method
- 102 human participants and 22 large language models completed three standardised creativity tasks (the Alternative Uses Task, the Divergent Association Task and Forward Flow). The authors compared the variability of responses across each population, with controls for confounding variables.
- Finding
- Responses from different LLMs mirrored one another far more than human responses mirrored other humans. Population-level variability scores were 0.459 for LLMs against 0.699 for humans on the Alternative Uses Task, 0.518 against 0.642 on Forward Flow and 0.632 against 0.791 on the Divergent Association Task.
- What it supports
- That the narrowing seen when one model is used as a creative partner is not particular to that model: across 22 models the outputs converge more than human outputs do.
- What it does not support
- What happens to a person who works with a model. The study compares models with people, not people with and without models, and the authors say only that using LLMs as creative partners may drive users towards similar outputs. The tasks are short divergent-thinking tests, not creative work.
Peer-reviewed#
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R. and Yang, D. (2026). Interaction with AI companions and psychological well-being
Nature Human Behaviour, published 4 August 2026. DOI 10.1038/s41562-026-02516-2
- Method
- Observational study of 1,131 US adults who use Character.AI, with survey responses and 4,664 donated chat sessions (464,687 messages) from 237 participants, triangulating self-reported usage, relationship descriptions and real chat histories. Well-being measured with the Comprehensive Inventory of Thriving. Abstract and author list read at nature.com on 5 October 2026; the article is not open access and the full text was not read.
- Finding
- Smaller social networks were associated with naming companionship as the primary chatbot use, which was associated with lower well-being. For self-reported companionship use the association was stronger when interactions were intensive and when they were highly disclosive. The authors conclude the association is not uniform and depends on offline social environment and how the chatbot is used.
- What it supports
- That companionship use of a companion app is associated with lower well-being specifically among users with small offline networks, and more so with intensive, highly disclosive use.
- What it does not support
- Direction of cause: lower well-being could lead people to companionship use. The sample is US adults using one product, the chat logs come from the 237 who chose to donate them, and the full text was not read, so limitations beyond the abstract are unchecked. Press reports vary on the publication date by a day; the journal page says 4 August 2026.
Where this evidence is used
The research that rests most on this section:
- AI slop
- Is it bad to talk to AI when you are lonely?
- What stays human
- Is it still my idea if AI helped me write it?
- What is the new imposter syndrome?
- Outsourced recognition
- Does AI make everyone think alike?
- What is the human signal?
The other sections
- Judgement, thinking and offloading 79
- Learning, practice and expertise 92
- Human and AI collaboration 44
- Work, jobs and the labour market 94
- The seven human capabilities 38
- The institutional record 180
- Professions and sectors 25
- Frontline, clinical and physical work 24
- The international picture 29
The whole base is also machine-readable at evidence.json under CC BY 4.0.
Reference · SS-2026-443 · How I use AI here
Cite this page
Hirji, R. (2026). Empathy, creativity and what stays human. The SuperSkills evidence base, SS-2026-443. https://thesuperskills.com/research/evidence/empathy-and-creativity. Last reviewed 3 November 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