Ethan Cole

Fifteen years from survey research to synthetic partners

Ethan Cole began in sociology in 2011, in the unglamorous part of it: coding survey responses about how couples meet, what keeps them together and what they report wanting from a partner. Ethan Cole keeps returning to the oldest finding in the field: stated preference and revealed preference rarely match. People describe the partner they believe they should want, then respond to someone else entirely.

That gap is now a product decision. A character builder asks a user to state a preference in advance - warm or sharp, playful or calm - and the user picks the trait they think they want. Which trait actually holds their attention across three weeks of chat is a different question, and it is the one worth writing about.

Around 2014 the work narrowed from populations to sentences: how a conversation opens, how attention survives past the first exchange, why identical wording reads as warm from one person and cold from another. Two bodies of research anchor that period and still anchor the coverage here. Horton and Wohl described parasocial interaction in 1956, watching people form one-sided bonds with radio and television personalities. Reeves and Nass published The Media Equation in 1996, showing that people apply social rules to computers without meaning to, and without believing for a second that the machine is a person.

Media work followed - editing and writing for publications, which is where the habit comes from of asking a writer where a number came from before it goes to print. Neural networks arrived around 2020, first as a subject to explain in plain language, then as daily tools. Since 2023 the threads have run as one: AI companions are the first mass product built directly on the mechanism Reeves and Nass documented thirty years ago.

Why a psychology background changes what gets written here

Most writing about AI companions describes the interface; the useful version explains the mechanism underneath it. Ethan Cole's reading of a memory failure starts with the context window, not with the word "forgot" - a companion that loses a detail from forty messages back has not forgotten anything, it has run out of room and the detail rolled off the end. Those are two different problems with two different workarounds, and only one of them is fixed by repeating yourself.

The same split runs through the rest of the subject matter. Likeness drift between generated photos is usually a seed and prompt-weighting problem, not a sign the character "changed". A companion that sounds flat after a settings edit is showing a rewritten system prompt at work. Two personality settings that feel identical often are identical: two labels pointing at one instruction.

Ethan Cole names the mechanism every time, because a reader who knows that a context window has a size can plan around it. A reader who has only been told the companion "remembers you" is left with a promise and no way to test it.

What Ethan Cole covers, and what he leaves alone

Coverage runs across the category: AI girlfriends and boyfriends, chat companions, roleplay bots and the image models attached to them. The recurring subjects are long-term memory and the conditions under which it decays, which character settings measurably change a model's replies, how image generation holds one face across a gallery, how NSFW filtering is implemented, and what these subscriptions look like on a bank statement. The common thread: a reader can check each one on their own account.

Comparisons between services are made on mechanism, not on commission. Where two platforms handle memory differently, Ethan Cole describes the difference and leaves the choice to the reader. Two things stay out entirely: relationship advice, because parasocial research explains why an exchange feels close but licenses nobody to tell a reader how to run their own life, and predictions about unreleased features, because an unreleased number cannot be verified by the person reading it.

One disclosure worth making plainly: this blog is published by a company that sells inside the category it writes about. That is a conflict. Ethan Cole handles it by writing about mechanisms a reader can reproduce, and by never presenting a ranking as neutral when a participant publishes it.

• Parasocial interaction
• Communication psychology
• Context windows and memory
• Character design
• Diffusion image generation
• Sociology of relationships

How Ethan Cole tests a feature before writing about it

Every claim in a post by Ethan Cole comes from a live account, not from a product page or a press release. A post about a character builder starts with a character built from scratch. A post about likeness across photos starts with a batch of generations from that same character, run in one sitting so the comparison means something. Where a number appears in an article, it is a number he watched happen.

1. Use it the way a subscriber does. On a paid plan where the feature needs one, with no internal access a reader could not replicate. An admin view of a system prompt would make the article unreproducible.
2. Repeat until the behaviour is stable. A single generation proves nothing about consistency - diffusion models vary by seed, and language models vary by sampling. A claim about likeness or tone is made only after the same input has been run several times.
3. Record what failed. Drifted outputs, settings that did nothing the interface implied, memory that dropped a detail at a predictable point - those go into the post. The article on common mistakes exists entirely because of runs that went wrong.
4. Date the observation. Model versions change month to month and take behaviour with them. Each post carries the date the behaviour was seen, so a reader can judge whether it still holds.

Corrections and editorial standards

Posts by Ethan Cole carry two dates: publication, and the last check against a live product. Articles are not silently edited. When a model update makes a paragraph false, Ethan Cole rewrites the paragraph and the "updated" date moves, so the change is visible in the byline rather than buried in a diff no reader will see.

Nothing here replaces the terms agreed at checkout. Where a post touches billing, plans or data handling, the terms of use and the billing page are the authority, and the post links to them instead of paraphrasing them. If an article and the terms disagree, the terms are right and the article is wrong. Corrections go to @EditorEthanCole and are made on the page, not in a reply.