How common it already is
In a national online sample of 2,658 German adults, 31.6% reported at least one AI-supported sexual activity in the previous twelve months, and consuming AI-generated pornography was among the three most common. Men, younger adults and non-heterosexual participants reported more of it. It was an online sample rather than a probability sample, so read 31.6% as scale, not a population rate.
A content analysis of 36 generation sites describes the product: 80.6% offered image generation, 41.7% video, 44.4% some form of interaction with an artificial agent. Customisation was close to universal — 97.2% by feature selection, 72.2% by prompting.
What nobody knows
No study has compared AI-generated with conventional pornography on use, on arousal over time, on wellbeing, or on any clinical outcome. The field's 2026 flagship review of compulsive sexual behaviour and problematic use — 31 authors, interdisciplinary — does not discuss AI at all.
There is exactly one head-to-head comparison. 649 gynephilic adults rated real photographs against computer-generated, AI-generated, doll and hentai nudes, and the AI-generated images scored highest on aesthetic appeal, sexual attractiveness and pleasantness. That is a one-shot rating of static pictures in a survey — no measure of arousal, no measure of use, no repetition, no problematic-use sample. It is the only comparison there is.
So this page cannot tell you AI porn is more addictive, or that it escalates you faster. Nobody has measured either. If you have read otherwise, the confidence came from somewhere other than the evidence.
The mechanisms that do apply
Sexual arousal habituates to a repeated stimulus and recovers when a novel one arrives. That has been shown across three decades, on genital response and on self-reported arousal, in men and in women. It is the mechanism the endless-novelty argument runs on. The scope matters: those are short single laboratory sessions with film and fantasy stimuli, which do not establish that novelty-seeking compounds over months and say nothing whatever about AI-generated material. The one preregistered experimental test of the popular endless-novelty claim about internet pornography found only partial support, and its authors concluded that far greater nuance was required.
The other relevant piece is what dopamine tracks: anticipation and pursuit — wanting — more than the pleasure of consuming. A product whose whole design is that the next thing can be generated on demand is working on the anticipation side. That is two established findings put next to each other, not a measured result about AI, and it is worth holding loosely.
Who is telling you this
Nightshore is a paid subscription, and every company in this market — apps, blockers, coaching, residential programmes, supplements — has a financial incentive that favours the addiction framing. That applies with extra force to AI sexual products, whose commercial model is open-ended generation: a generator has no last video the way a site has a last page. Both are observations about who profits from what, not findings.
How you quit it
The same way, because nothing in the evidence identifies an AI-specific mechanism that would need an AI-specific method. If that reads as a let-down, it is what is supported today.
- Pushing the thought away backfires. Suppressing sexual thoughts produces a rebound: the thoughts increase. This is one of the few things on this page that is settled.
- Friction is the sourced move. Increasing the friction of a behaviour — moving the cue further away, reducing availability, breaking the sequence that leads to it — reduces how often it occurs. The direction is consistent and the effects moderate, though the evidence is graded low-certainty and comes mostly from short laboratory studies of food. What is separately well established is that intentions on their own translate weakly into behaviour.
- Friction is harder to arrange here, and that is an inference rather than a finding. Generation often lives inside a general-purpose app or account rather than on a set of blockable domains. Nobody has tested blocking strategies against AI generators.
This is a new surface on an old problem, and the reason nobody can tell you more is that nobody has looked yet.
Where this comes from
Read the full lesson: Where things stand now →
Sources and how strong they are (6)
- AnecdotalAI-generated pornography is already a mass consumer product and almost nothing is known about what it does to the people who use it. In a national online sample of 2,658 German adults, 31.6% reported at least one AI-supported sexual activity in the previous 12 months, and consuming AI-generated pornography was among the three most common; men, younger adults and non-heterosexual participants reported more of it. A content analysis of 36 generation sites found 80.6% offering image generation, 41.7% video, 44.4% interaction with an artificial agent, and customisation by prompting (72.2%) or feature selection (97.2%) across body features, clothing, sociodemographic characteristics and setting. What does not exist is a comparison. No study has compared AI-generated with conventional pornography on any use, arousal-over-time, wellbeing or clinical outcome. The single head-to-head comparison in the literature is a survey of subjective ratings of still images, in which 649 gynephilic adults rated AI-generated nudes as more aesthetically appealing, sexually attractive and pleasant than real photographs - a rating of pictures, not a finding about how anyone uses them or what follows.
- Döring N, Mikhailova V, Mohseni MR 2026, Archives of Sexual Behavior, doi:10.1007/s10508-025-03382-1, PMID 42265505 (correction, 5 August 2026: doi:10.1007/s10508-026-03612-0)
- Lapointe VA, Dubé S, Rukhlyadyev S, Kessai T, Lafortune D 2025, Archives of Sexual Behavior, doi:10.1007/s10508-025-03099-1, PMID 40032709
- Zakreski E, Marečková A, Vaníček O, Hůla M, Klapilová K, Lindová J, Pfaus JG 2026, Archives of Sexual Behavior 55(1):329-343, doi:10.1007/s10508-025-03357-2, PMID 41652242
- Döring N, Le TD, Miller DJ 2025, Archives of Sexual Behavior, doi:10.1007/s10508-025-03227-x, PMID 40968214 (correction: doi:10.1007/s10508-025-03385-y, PMID 41369810)
- Negative evidence, recorded deliberately — Ince et al. 2026, J Behav Addict 15(1):32-98, doi:10.1556/2006.2025.00337, PMID 41879870, PMC13132360
- Well establishedSexual arousal habituates to a repeated erotic stimulus and recovers when a novel one is introduced. This has been shown in the laboratory across three decades, on genital response and on self-reported arousal, in men and in women, with the recovery appearing on attentional and affective measures at the same time. It is the mechanism that the endless-novelty argument about pornography rests on, and it is the reason a product that can generate the next thing on demand is acting on anticipation rather than on satisfaction - see C47, where dopamine tracks wanting rather than liking. What the mechanism does not establish is any effect specific to AI-generated content: no study has measured habituation, novelty recovery or craving for AI-generated pornography, and the one direct experimental test of the popular endless-novelty claim about internet pornography found only partial support and called for far greater nuance.
- Koukounas E, Over R 1993, Behaviour Research and Therapy 31(6):575-585, doi:10.1016/0005-7967(93)90109-8, PMID 8347116
- Koukounas E, Over R 1999, Archives of Sexual Behavior 28(6):539-552, doi:10.1023/a:1018769200845, PMID 10650440
- Koukounas E, Over R 2000, Behaviour Research and Therapy 38(6):573-584, doi:10.1016/s0005-7967(99)00075-3, PMID 10846806
- Both S, Laan E, Everaerd W 2011, Journal of Sexual Medicine 8(1):167-179, doi:10.1111/j.1743-6109.2010.02051.x, PMID 20946171
- The sceptical counterweight, cited deliberately — Sculley J, Watkins CD 2022, Archives of Sexual Behavior 51(6):3071-3082, doi:10.1007/s10508-022-02317-4, PMID 35790609
- Well establishedDopamine is more closely associated with the anticipation and pursuit of a reward - "wanting" - than with the pleasure of consuming it - "liking".
- Berridge & Robinson 1998, Brain Research Reviews 28(3):309-369, doi:10.1016/S0165-0173(98)00019-8, PMID 9858756
- Berridge 2007, Psychopharmacology 191(3):391-431, doi:10.1007/s00213-006-0578-x, PMID 17072591
- Well establishedSuppressing a thought tends to make it more frequent once you stop suppressing — the rebound effect, found consistently across 31 experiments. Whether the thought increases while you are still suppressing it shows up only under additional cognitive load. On sexual thoughts specifically the evidence is indirect: the one study in this area is a cross-sectional survey of religious adolescents that models suppression as a mediator, not as a cause.
- Wang D, Hagger MS & Chatzisarantis NLD 2020, "Ironic Effects of Thought Suppression: A Meta-Analysis", Perspectives on Psychological Science 15(3):778-793, doi:10.1177/1745691619898795
- Does not report a rebound effect — Efrati Y 2019, "God, I Can't Stop Thinking About Sex! The Rebound Effect in Unsuccessful Suppression of Sexual Thoughts Among Religious Adolescents", J Sex Res 56(2):146-155, doi:10.1080/00224499.2018.1461796, PMID 29702013
- Mixed evidenceIncreasing the friction of a behaviour - moving the cue further away, reducing availability, breaking the sequence that leads to it - reduces how often it occurs. The direction is consistently found and the effects are moderate, but the evidence is graded low-certainty, comes mostly from short laboratory studies of food, and has never been compared head-to-head with intention-based approaches. What is separately well established is that intentions on their own translate weakly into behaviour.
- Hollands et al. 2019, Cochrane Database of Systematic Reviews CD012573.pub3, doi:10.1002/14651858.CD012573.pub3, PMID 31482606
- Webb & Sheeran 2006, Psychological Bulletin 132(2):249-268, doi:10.1037/0033-2909.132.2.249
- Wood & Neal 2007, Psychological Review 114(4):843-863, doi:10.1037/0033-295X.114.4.843
- Well establishedNightshore is a paid subscription product, and companies in this market - apps, blockers, coaching, residential programmes, supplements - have a financial incentive favouring the addiction framing. This is a disclosure about the app and the market it sits in, not an empirical finding. This applies with extra force to AI sexual products, whose commercial model is open-ended generation: a generator has no last video the way a tube site has a last page, and the same incentive that favours the addiction framing here favours unlimited engagement there. Both statements are observations about who profits from what, not measured findings.
- docs/premium.md — this app's own commercial position, which the first half of the entry is a fact about
- Lapointe et al. 2025, doi:10.1007/s10508-025-03099-1
- Not cited as evidence of an effect, and not peer reviewed — Knox et al. 2025, arXiv:2511.14972
- Not cited as evidence of an effect, and not peer reviewed — Chen & Hiniker 2026, arXiv:2609.14696
More on urges in the app.
The full 41-lesson course is in the app, every claim graded the same way, and a Panic button for the moment an urge hits. Free.
Coming to iOS and Android.