At a conference recently, someone pulled me aside and told me they’d be down to offer me funding for my research. I was like, awesome, because for years people have been yelling at me for doing surveys on the internet.

If you’re new to my work, I’m a self-taught researcher, driven to learn through fortunate access to high sample sizes and sheer curiosity. As of this writing, over 700,000 people have responded to my Big Kink Survey on fetishes and sex — likely one of the largest surveys on the topic ever done. It grew out of my desire to answer two questions: How weird is my own sexuality, exactly? and Why are there way more women interested in submission than men interested in dominance?

I originally started doing surveys via porn (I also do porn), which is a super easy way to get a lot of horny social media followers willing to do anything you ask them to. I’d ask them to help answer random questions I had, and quickly realized the data was terrible because my questions were terrible.

This started years of rapidly iterated survey learning. I’d pump out tons of questions, look at the resulting data on Google Sheets, and then rock back and forth with agony that I hadn’t thoroughly thought through how people might interpret question five and now all my correlations were useless. I’d often practice with questions in polls on X (formerly known as Twitter), which got lots of people in the comments reliably misinterpreting my phrasing in the most ingenious ways. I became a calloused, precise, question-asking machine. My phrasing could split any hair, and I started failing less and less frequently. Now, my crying-in-corner-agony per question rate is probably less than 1%.

It turns out that, when you get data, you have to figure out how to interpret it. Since I’m a homeschooler without a college degree, I started wrangling smart people to tell me how the hell a correlation worked, what p-hacking was, and how to do factor analysis. I’m a slow but thorough learner, and eventually got to the point where I was pretty solid with simple statistics.

But my datasets were getting huge, because my survey design was getting better at seducing people into taking them. My surveys, originally getting 10,000 responses, recently started to explode into the hundreds of thousands. You can’t process huge data in Google Sheets, so starting in 2021 I begrudgingly learned how to program in Python.

My audience expanded past just porn users into other, usually nerd-adjacent, platforms. Often my surveys would go viral elsewhere, bringing in novel audiences — recently, mostly young, female social media users, primarily on TikTok. I started buying randomized samples from the study recruiting site Positly to compare against my own, to try to understand the degree to which the respondents I drew from social media might be unrepresentative.

I’ve been saturated in the world of rationalists — an intellectual community formed around the group blog LessWrong — for eight years or so, and from watching them viciously tear apart other studies, I’d gained a sense of mistakes not to make: Don’t control for too many or too few things, do not trust your intuitions on how stats should work, don’t avoid reporting things that contradict your results, don’t check for new ways to investigate data just because it didn’t match your expectations, and report the null findings. I got odds-ratio-pilled and used those over p-values when I could; I published raw data; I shared my code; and I tried to keep my analysis and interpretation separate.

Finally, I’d write up my findings and post them online. I’m not an expert — there’s a ton of stats stuff I don’t know — but I tried to keep it simple and just say what I found.

It turns out that a lot of people really hate this. People say I’m not trustworthy because I’m a sex worker and thus biased, because I don’t have a representative sample, because I don’t have credentials. My work isn’t peer reviewed, I haven’t weighted the data to ensure it’s representative, and I’m an “OnlyFans girl just polling her horny audience.”

So when the guy at the conference offered funding, I was excited to do official, real work. I’d dreamed that one day, if I had money, I’d do a formal, grown-up study, with verified participants. It’d be reliable and like the way professors do it and make all the angry internet people embarrassed for ever doubting me. I fantasized about repentant emails drizzling into my inbox. “You were right all along,” they would say. “The silly online correlations you found do in fact match the official serious business ones too! You’re a real scientist and we trust you."

But as I went about the rest of the conference, trying to plan how I would do my research respectably, it slowly dawned on me that … there actually wasn’t a clearly better option to recruit people to talk about kinks. Each potential study concept I worked through suffered from equal, if not greater, vulnerabilities than mine did.

For example, let’s say you want to study people who are really into vore, a fetish around swallowing someone whole or being swallowed whole, typically by a much larger creature. Maybe you want to find out if it correlates with things like childhood abuse or mental illness — common theories people have about the origin of fetishes. Let’s say you want around 300 people to check.

Firstly, how do you find these people? Vore, like most fetishes, is rare. If you wanted to do a survey where you randomly ask people about their childhood, and also if they’re into vore, you’d have to ask thousands to get the 300 people who mark “yes.” So how do you find your 300 body-swallowing enthusiasts? You probably need to look at some forums for vore lovers.

Already, we’re introducing a nonrandom sample — we’re picking from people active on a vore forum. Maybe these people are less ashamed about their fetish, maybe they’re lonelier in real life, maybe they’re much more into vore than the actual population of people into vore. These people are probably unusual in much more specific ways than the category of “people just browsing TikTok,” for example.

This issue plagues most existing sexual fetish research. Pedophilia is the most extreme example. In a society with such massive penalties for even suspected pedophiles, it’s obvious that any pedophile-heavy source you go to is going to be hugely affected by selection bias — e.g. sex offender registries. One study recruited pedophiles from an “outpatient facility for sexual offenders” and — shockingly — found that they had compulsive behaviors and anti-social personality traits.

But back to our comparatively innocent vore enthusiasts. However we find them, we need to incentivize them to participate. You can pay participants in money or gift cards or, if you’re a professor, with grades. This might work. On paid survey-taking sites, you can try to improve the quality of your sample by looking at things like how fast it takes them to do the survey, are their answers consistent, do they have high quality ratings by previous researchers, etc., But you can’t filter by “Do they like vore?” — that’s not a default demographic that most services provide. 

Let’s say you have a survey that takes a half-hour, which means you might be paying $4 a survey. If you want 300 people who are into vore, and let’s say ~4% of the population likes vore, this means you’re going to need to pay roughly 7,500 people, which makes this a $30,000 study. And keep in mind you don’t know the vore prevalence beforehand, because this data isn’t available in traditional journals, because nobody has done a high-powered general fetish prevalence study before. Except me, which is why I know that number.

So let’s limit ourselves to confirmed vore fetishists. In that case, what’s to stop people from filling out the survey dishonestly in order to get the money? Maybe they wouldn’t, but “studies done on people trying to get $20” is a commonly mourned problem in research, and it seems reasonable that niche fetish enthusiasts are even more inclined to be dishonest than usual if they believe researchers are going to be scrutinizing the results.

And we want this to be a real-life survey anyway. To do this, we need to get our vore friends into an office, verify their identity, and then have them fill out a survey in exchange for payment.

Uh, this is terrible. How could we possibly trust any results we get from this? 

Problems like “How do you find a not-super-selected source, verify nontrolls are taking the survey, and get them to be honest?” exist in a lot of fields, but they’re especially difficult in sexual fetish research. The entire thing is studying rare traits that people are incentivized to lie about!

I recently looked at the the ratio between the frequency of fetishes in my sample and the degree to which people were aware of those fetishes in others. The least visible item I measured was BDSM. And the more uncommon the fetish, the more taboo, and the greater the pressure against sharing it.

Genital response measuring is an option — but we’d really be selecting for people who are okay with an invasive measuring process, where your genitals are showing exactly how turned on you are by, e.g., an anime of a young boy screaming as he slides down the vast gullet of a half-lizard woman in a microkini. There are benefits to this — it might be good for checking correlations across lots of fetishes — but there’s no way this sample can be random.

So, given all this difficulty, it makes sense that most fetish research involves the scenario I described above — going to various fetish communities to ask them questions, usually in the form of visiting a forum or club dedicated to this fetish and posting a survey there. Consider the following sample of studies published between 1980 and 2020:

Sexual Fetishes: Sensations, Perceptions, and Correlates (316 people who took an anonymous internet survey posted in fetish communities)

Personality Characteristics of Fetishists, Transvestites and Sadomasochists (543 people who took this questionnaire distributed in person to fetish clubs)

Forbidden Games: The Construction of Sexuality and Sexual Pleasure by BDSM ‘Players’ (343 who responded to requests for interviews on Yahoo, sexuality, and BDSM communities)

BDSM: A Subcultural Analysis of Sacrifices and Delights (73 people, from a post to a BDSM email discussion group)

BDSM Role Fluidity: A Mixed-Methods Approach to Investigating Switches Within Dominant/Submissive Binaries (202 people, from posting to internet fetish communities and flyers in in-person fetish meetups)

The Psychology of Gay Men’s Cuckolding Fantasies (580 people who responded to an internet ad asking about gay men’s cuckolding fantasies, posted through — I presume David Ley’s — social media channels and Dan Savage’s podcast)

If the Shoe Fits: Exploring Male Homosexual Foot Fetishism (262 respondents from an organization for foot fetishists)

To be clear, I still think these studies are respectable! I don’t mean to say a study is bunk just because the researchers behind it have limited access to populations - but the method seems ripe for accidental correlations.

Lots of existing research (maybe due to the necessity of using active community groups to get your sample) also focuses on action as opposed to fantasy, which conceivably biases the sample. A lot of things unrelated to a fetish probably affect whether someone acts on it. Are they in a liberal culture that celebrates their love of sitting on cakes? Are active cake sitters more disagreeable than those who wistfully dream about one day icing their own buttocks? Less afraid? Then we’d expect any studies done on cake sitting to find the cake-sitter community to be more liberal, disagreeable, and bold than the true population of people aroused by cake sitting.

Given all this difficulty, the ideal form here is probably an anonymous survey where people are properly incentivized to be honest — and what does this better than giving them information back about themselves? You need to give people a reason, and that reason should be as closely aligned with reporting the truth as you can make it. Telling people what you’ve found out about them based on that survey seemed like an obvious answer.

Basically: an elevated take on the classic BuzzFeed quiz, “What Princess Are You?” Of course this isn’t perfect — one can easily imagine someone deliberately putting in odd answers to see what this does to the results — but I think this still beats out other methods of research for most purposes.

Plus, it’s easy to disincentivize people spamming funny answers into your form for fun — just make your survey really friggin’ long.

I’d gotten a year’s worth of funding from the wonderful Invisible College in 2021, and so I cracked my knuckles, quit OnlyFans, and settled down to work.

I wanted to make a survey about the origin of sexual fetishes because we know almost nothing about it, which is probably because this is really hard to do, for the reasons I’ve outlined above. Luckily I didn’t fully appreciate the difficulty when I started on this problem, or I might never have tried.

My goal was simple — I’d pick a fetish, ask if people were into it, and then ask them a bunch of questions about childhood. This would allow me to determine, for example, if you got spanked in childhood, did that correlate with interest in being spanked as an adult? This was pretty easy for a narrow fetish like this, but I wanted a big sample for the reasons I explained above — searching specifically for people interested in spanking seemed like it would do the whole “skewing results ’cause you’re drawing from the spank-specific community” thing, so I needed to get a bunch of people to answer for some other reason and then happen to ask them about spanking. This meant a really big sample.

And my access to big samples was limited. My audience — the most active of which was around 150,000 followers on X, though my total was around a million distributed across various platforms — would only take so many surveys, and I didn’t want to ask the same background questions over and over for each new fetish survey. It would be so much more efficient to combine all the fetishes into one big survey!

And while we were there, we should throw in more questions — what about porn use, sex rate, BMI, IQ, Big Five personality traits? You can check correlations for each additional question with everything else, so the more questions you add, the more valuable the data becomes. As my list of questions increased, the more compelling it became to add on yet another question, and this turned into a nightmare snowball.

In the end, I included 61 demographic, childhood, and other questions; 35 “general sex and porn” questions, and 206 fetish questions, for a total of 302 questions. But many of these were checkbox lists, where I asked people to “mark all that applied.” If you count those, the total amount of items I measured climbed to around 1,000.

I don’t know if you know how survey-taking norms work, but trying to get someone to answer 1,000 items is absolutely unhinged. It’s like asking someone to meet for coffee and then forcing them to stay for 12 hours of small talk. And the final cherry on the top of this sundae of horror was that the size of sample you need to make findings significant in the traditional sense increases with the number of questions you’re asking (or, more specifically, correlations you’re checking). So I needed a big sample size — many thousands, at least. But how do you get many thousands of people to sit down and answer a thousand questions?

It depends on the type of question, but on average I’ve found people take around two to six seconds to answer a survey question, which meant it would take roughly between 30 and 90 minutes to complete my survey.

I could pay people to take the survey, but that would get really expensive. If I were paying people $8 per hour to do this, and the survey optimistically took 30 minutes to do, then 10,000 responses would cost me $40,000.

I had to get creative.

First off was shortening the amount of time to take the survey, which sounds simple but was agonizing. I couldn’t let go of any of my precious questions. Each question I considered cutting meant I was releasing all of the other correlations I asked about into the wind. I felt like a hoarder on a TV show, wailing as I watched Marie Kondo slowly approach my front lawn.

But I could condense them. After lots of experimentation, building and testing out various survey versions, I settled upon a gated method. I’d ask people the first two categories — personal stuff and general sex stuff — like a normal survey, but the fetish part would be gated. I’d list a bunch of vaguely described fetish categories, like “Humiliation (defilement, impotence, cuckoldry, ridicule, etc.)” or “Transformations (growth/shrinking, bodyswapping, furries, etc.),” ask them to check off the ones they thought might apply, and then I’d feed them relevant questions from the categories that they checked.

But this meant I needed to define fetish categories, and there is no good research on this anywhere. I had to invent them. So I went and scraped the internet for all the fetish lists I could find, compiled them into a spreadsheet, and painstakingly tagged them all. There were nearly 900 fetishes, and I had to individually research most of them by wading into enthusiast boards in order to know how to tag them. I as a person really lack willpower (or discipline in general), and so here the only thing that kept me going was deliberately, frothily fantasizing about getting all the data in the end.

This still wasn’t ideal. My personal categorization of these fetishes was going to artificially cluster the data. If you clicked on, say, the “clothing” category, you would be exposed to more clothing-specific questions, which might make you more likely to mark “yes” to things you’re only mildly interested in, while someone who didn’t click that category wouldn’t see those questions at all.

And I wasn’t that confident in my fetish clustering. What if I put “cake sitting” into the “humiliation” category instead of the “sensory” one, but it turned out cake sitters didn’t care at all about humiliation? Despite my rigorous research, there’s no way I was going to have a 0% error rate.

To help compensate for this, I picked a few of the gated fetish questions and repeated them outside the gate, so I could compare the answers between the whole survey population and people who expressed an interest in a specific cluster to see just how much my survey design was affecting responses. This still wasn’t perfect, though!

The ideal question format was a Likert scale — “I find light bondage to be:” on a “not erotic” to “extremely erotic” zero-to-five scale. With this information, I could help differentiate between responses like “meh I guess” and “this is essential to my sexual function.” But scales take a while to complete, and so this ended up as a sacrifice on the altar of time sensitivity. I reserved Likert scale questions for general categories, like “clothes,” and then condensed many of the subcategories into increasingly nested checkboxes. I get to see how much you like the general category, and then binary responses to items within that category. If you like clothes, does that mean specific clothing textures, states, or items? If textures, is it latex? Leather?

The entire process was a series of tiny, painstaking tradeoffs. Each time I decided to make a question into a scale, this boosted the total time it took to complete the survey by a few seconds. And I had to care about those few seconds.

In the end, the average time it took users to complete the survey was 40 minutes. This is really long for a survey! You really have to get people to care about it in order to finish.

So: How do I get them to care?

I decided to give users a score at the end — How freaky are you compared to others? 

This turned into watery-eyed frustration all on its own, because how do you measure freakiness? Sure, you could do a survey to get people to rate how taboo various sexual things were — and this is precisely what I did — but this still leaves the problem of weighting. Necrophilia is really freaky, sure, but what if someone is only slightly into it? Is that freakier than someone who’s extremely into voyeurism? And then there’s the fact that freakiness is also a function of gender norms. If you’re a really submissive man, that’s certainly culturally freakier than being a dominant man, so I had to calculate different scores independently for men and women.

Scoring the survey took almost as much time as actually building it did.

The last part was easy. I assembled a list of fictional and well-known characters — like SpongeBob, Hitler, and Bambi — and asked people to rank them on how taboo they assumed their sexuality was. I then matched them up, based on these rankings, to the scores that people got in the Big Kink Survey. I personally got House, from that TV show some years back, or roughly 85th-percentile kinky. The character that people assigned the highest kinkiness score was Willy Wonka.

As of my writing this, the survey has around 550,000-700,000 respondents, depending on how aggressively I’m cleaning the data. Most respondents are in their early 20s, and around 70% of respondents are women — a reflection of the demographics that most love taking internet surveys (as verified by my friend who runs one of the oldest, biggest personality-testing websites on the internet).

One great thing about my sample is that I asked about hundreds of niche fetishes without having to go into a single niche-specific community and ask its members to take my survey. The survey results for the fetishes I’m asking about are as close to an unbiased sample as we can reasonably get given you’re trying to get a big sample with low cost.

I think information about your personality is one of the best possible incentives to complete a survey. Yes, your answers are probably warped a bit by your self-perception, but at least it’s an incentive directly related to you, as opposed to trying to get paid.

And the length of the survey probably indirectly helped accuracy here too — when I was younger and took personality tests for fun, I would retake them to see what changing little answers did to my score — but this is really annoying to do if the survey is 40 minutes long! This also made it resilient to intentional brigading or trolling — it’s simply difficult to sink that much time into warping the results, especially to the degree where it would show up on my final analyses, and especially with enough cleverness to get around all the survey checks I included to catch unserious responses. And when looking at survey results, I only include data from people who’d completed a majority of the survey.

The data I’ve gathered so far seems to match up quite well with existing data, at least for well-studied variables. For instance, my height and weight numbers correlate at around the standard 0.5, the rate of pedophilia in men is 3%, and the gap in interest between women and men in BDSM is clearly represented.

I don’t know why academia hasn’t tried to come up with more innovative solutions for gathering large amounts of survey data. Researchers are already doing internet surveys, so why not sink some effort into marketing? Hire some influencers? Get on podcasts? I don’t know. I treat my surveys not only as a vehicle to gather data, but also as products in themselves — to sell your survey, you have to make respondents enjoy it. You have to make them like giving you their time. It seems like current academia — or at least the sex part — views taking surveys as something like a chore, or like school. I’m not sure they’ve thought about making it fun.

I don’t mean sacrifice the quality of the survey itself. I was painstakingly dry with the wording of the questions, because I didn’t want a cheeky tone to prime respondents to give cheeky answers — but results at the end are a clean way to make it fun. I didn’t have to tell people they were the sexual equivalent of SpongeBob or whatever at the end of this survey, but I did — and this caused them to share it with friends and on TikTok.

I come from a background of marketing sex work, where I had to claw my way up the internet flesh by sheer “how do I make people want to send this thing to their friends.” And this is the eye with which I’m approaching research. I have done most of this without funding, yet have gotten samples that, if you paid test takers to do, would cost upwards of $4 million.

I also make my research a community effort — not only do I share my raw data and code, I regularly crowdsource questions from the public about what to study next. What hypotheses do people have that they want tested? I do drafts of survey questions in X polls, to see how commenters will inevitably misinterpret my wording and thus inform me on how to write the question more clearly in the future. I hope this process helps vanquish the sacredness of research.

And more specifically, I hope it normalizes people with fetishes. I feel such care and compassion for people walking around with these strange arousal patterns in their head that often cause such alienation. They’re shunned or ignored socially, but also by researchers — because of the logistical difficulty, because institutional review boards make approval hard, because sexuality is a subject rife with potential triggers, or because people simply don’t want to investigate things that aren’t trendy or socially sympathetic. I’ve got half a million data points, with an individual behind each one, and I’ll keep trying to understand all of them.