A New Tool Measures Whether AI Models Remember You

Searching your own name on Google no longer tells you much. A new site argues the real 2026 question is whether the major AI models can recall you at all.

What In the Weights Actually Does

The premise is simple and a little unsettling. TechCrunch reported that In the Weights measures how well a model can recall a person without leaning on tools like web search.

The mechanics are straightforward. The site asks several models a standard prompt of the form who is this person, requests up to ten results each with a confidence rating, clusters the similar answers, and assigns a strength score.

The roster is broad. It queries Grok, Gemini, multiple versions of GPT, Claude, and Llama, then shows which model returned which answer and flags responses that look like hallucinations.

There is a leaderboard, naturally. Bitcoin World explained that higher scores reflect stronger and more consistent recall across the models, and the standings shift in real time as people run new searches.

Who Built It and Why

The creators bring relevant pedigree. Thomas Dimson and Joey Flynn are former OpenAI designers who joined the company when it acquired their startup Global Illumination, and Dimson previously helped build Instagram's ranking and discovery systems.

The motivation was partly mischief. Dimson told TechCrunch he and Flynn wanted to get the creative juices flowing again after leaving OpenAI, and a tongue in cheek blog post sealed the direction.

But the underlying thesis is serious. Dimson framed it around the idea that Google vanity searches are the wrong objective in 2026 as more traffic moves to language models.

Why This Is a GEO Story, Not a Gimmick

It is easy to file this under novelty, and skeptics do. Newsgab noted that critic Anthony Moser dismissed the project as essentially asking several chatbots the same question and aggregating the replies.

That critique misses what the leaderboard exposes. The site makes visible a surface that brands and individuals have been told to care about but could never quite see, which is what models actually hold about them.

The reputational stakes are real. Startup Fortune argued that reputation used to be inspectable, since you could search a name, scan the first page, and request a correction, while an AI answer is generated rather than listed.

For founders, that shift is not vanity at all. You can run a clean site and still be thin in a model's memory, or carry years of messy public material that a chatbot now quietly edits into three sentences.

The Harder Questions Underneath

The tool also surfaces uncomfortable issues about consent and bias. European Purpose noted that most people encoded into a model never agreed to it, raising live questions under European privacy law.

Dimson himself flagged the bias angle. He said he wants to study why different versions in the same model family diverge and which people should have a reference entry but do not, a gap that mirrors the skew in training data.

That gap matters for discovery work. Model recall is not a neutral mirror of real importance but a reflection of which sources got scraped and trusted, which is exactly where structured, citable content earns its keep.

The same dynamic shows up in research on how models reason, with recent AI reasoning research probing how internal representations shape outputs. Recall and reasoning both come down to what a model encoded and how confidently it can use it.

What Operators Should Take From It

The practical move is to treat model recall as a channel you monitor. Run your brand and your founders through the major engines and read how each one describes you, including the confident errors.

The fix is familiar but newly urgent. Clear source of truth pages and references on sites that models trust, including assistants like ChatGPT, are what shape whether a model can place you accurately.

In the Weights will probably be remembered as a fun afternoon for most people who try it. The idea behind it is going to outlast the novelty, because being legible to a model is becoming its own form of being known.