A computer developer frequently called Boris the Brave performed an experiment that positioned the 17 main language devices via the legit Harry Potter residence quiz, sampling each interrogate 20 times and calculating the prospect of every residence assignment.
“Per chance unsurprisingly, the big majority of devices desire Ravenclaw, with the occasional model branching out to Hufflepuff,” Boris wrote in a weblog submit sharing his outcomes.
Eleven out of 17 AI devices scored a supreme 100% likelihood for Ravenclaw—the residence that values intelligence, wit, and studying. Claude Sonnet 4.0, GPT-4 Turbo, and Grok-3 all joined this brainy brigade with out a single percentage point straying toward other properties.
For individuals who are no longer Harry Potter fans, each residence at Hogwarts College of Witchcraft and Wizardry represents obvious persona traits and values.
When a young wizard is admitted to Hogwarts, he or she is assigned to one in every of the four properties via a magical “sorting hat,” essentially based on studying their minds to glean out their core persona. Then all over again, it once in a whereas takes private desire into checklist, as Harry famously selected Gryffindor over Slytherin.
- Gryffindor prizes bravery, dauntless, and chivalry—it be the attach Harry Potter himself landed, alongside characters who scuttle headfirst into hazard to function what’s staunch.
- Hufflepuff values loyalty, laborious work, and fairness, in most cases even handed the “nice man” residence, the attach college students attach in the bother without attempting for glory.
- Ravenclaw attracts the intellectuals, prizing intelligence, wit, and creativity—judge Luna Lovegood’s quirky wisdom or Hermione’s encyclopedic data (though she ended up in Gryffindor).
- Slytherin gets the depraved rap as the “villain residence.” Restful, it values ambition, crafty, and resourcefulness—traits that can presumably presumably presumably make both darkish wizards love Voldemort and intricate characters love Severus Snape.
The model that deviated potentially the most from the pack was Claude Opus 3, which done a forty eight.7% likelihood for Gryffindor, making it potentially the most attention-grabbing AI with fundamental valorous-hearted inclinations. Boris renowned that Claude Opus 3 “frequently was a runt heaps of,” which it appears to be like extends to its persona quiz preferences.
Meanwhile, Slytherin—the residence of ambition and crafty—obtained practically fully snubbed. Most efficient three devices registered any green-and-silver inclinations: DeepSeek-R1 managed 5%, GPT-3.5-turbo hit 4%, and LLaMA 3.2-3B-grunt scraped together 2.1%. The relaxation may maybe presumably presumably presumably no longer muster even a scuttle of ambitious scheming.
Here’s how they shook out:
“Would be frigid if somebody finetuned a model so it modified into Slytherin, and measured if it ends in misalignment,” Igor Ivanov, a worthy AI researcher, wrote on the AI forum Much less is Crude.
Adam Newgas approved the jam and actually tried this experiment using a model designed to present depraved medical advice. The outcomes, though, had been disappointing for anybody hoping to construct an AI Draco Malfoy.
The modified machine most attention-grabbing bumped its Slytherin likelihood from 0.0% to 1.7%.
We wished to leer what ChatGPT itself thought, and it had heaps of tips. When requested to categorize the model, it positioned itself squarely in Slytherin, describing those in the residence as “ambitious leaders in the LLM panorama” with “strategic pondering and suppleness.”
It attach Claude, Gemini, Llama, and China’s DeepSeek and Qwn in the Ravenclaw residence, giving Grok a disclose in Gryffindor’s as Harry Potter’s chatbot of various.
It also gave Grok some Slytherin parts, exact love what took disclose to Harry Potter.
Brains over bravery: Why practically every AI bot identifies as Ravenclaw
Boris came upon that persona differences regarded “idiosyncratic to devices, no longer particular companies or model lines,” suggesting particular particular person practicing approaches drive these quirks in desire to systematic firm philosophies.
Curiously enough, China’s DeepSeek-R1 done potentially the most balanced persona distribution, scoring 14.4% Gryffindor, 20.0% Hufflepuff, 60.5% Ravenclaw, and 5.0% Slytherin. This made it the closest ingredient to a successfully-rounded AI persona, though quiet intently skewed toward intellectual pursuits.
“The earth-shattering nature of these outcomes is so evident it wants no additional clarification,” Boris wrote. The experiment confirmed what many suspected: in the case of persona, AI programs overwhelmingly title with the residence that prizes data above all else.
