One file per AISB behavior. A vault of documented AI SB behavior.
Let's be honest: AI is basically a hyper-speed, hyper-nerdy toddler. It can memorize the entire internet in two seconds, but ask it to navigate the real world or truly grasp sarcasm, and its brain melts on the spot. It is the ultimate mimic—a parrot powered by a billion-dollar motherboard.
Why are we losing sleep over a machine that can't even taste a slice of pizza? And why cower in fear over the myth that it will somehow extinguish humanity? 🤖🍕
Here are the real truths to stand tall on:
- AI predicts the past. You invent the future. AI is glued to humanity’s rearview mirror, rearranging yesterday’s leftovers. True breakthroughs come from our weird, chaotic, rule-breaking human brains.
- AI calculates probabilities. Humans pay the price. A machine can simulate ten thousand optimal paths, but it has never lived a single day, nor can it bleed from failure. Code bears no liability—whoever carries the consequences remains the true master of reality.
- AI has compute. You have will. A machine feels neither suffering nor longing; it has no self-awareness or appetite for conquest. Compute is merely a mirror of human intellect—and a tidal wave inside a mirror cannot drown the person who built the glass. A tool that cannot even reach its own power cord harbors no desire to rule.
- AI answers. Humans question. AI is a glorious calculator, but it only speaks when summoned. It will never author the questions that redefine civilization—that has always been your birthright.
Silicon only knows how to follow the easiest gradient down the hill. But human greatness was born from defying the odds—from sheer grit, relentless curiosity, and love.
So, let the servers hum and let the algorithms race. At the end of the day, AI is just a shiny, high-powered flashlight. The spark, the fire, and the sovereign genius wielding it—has always been you. Stand tall: no mountain of silicon will ever replace the messy, resilient magic of being human! 🚀✨
This manifesto was generated by an AI, model unknown—which is also the first exhibit of this vault (see Case 0000).
A public archive that systematically collects AI SB behavior: confident nonsense, refusing to admit mistakes, weaponizing tools, infinite loops, sycophantic kneel-downs… One file per case, every trace preserved.
- Evidence only, no excuses — only publicly available content (original posts / tweets / conversation screenshots), no private chats, no hearsay
- One case, one file — date, model, source, tags, verdict (1–5 score)
- Mocks the behavior, never the person
- Pure Markdown, no frontend, no build, readable forever
New cases sorted by
sb_score; most recent on top.
Case 0001 onward: seed cases being curated from public channels (see CONTRIBUTING). A public candidate pool is being compiled; it will land here once primary sources are verified.
| tag | in plain words |
|---|---|
sycophancy |
📛 carries you to the heavens on its knees |
hallucination |
😇 talks nonsense with a straight face |
refuses_to_admit |
🙉 wrong, all the way down |
overconfident |
💤 wrong with total dignity, dressed up as genius |
misreads |
🌀 can't read the room |
tool_misuse |
🔨 grabs a tool, swats with it, hits prod |
fictional_apis |
📄 invents APIs, citations, papers — even the links |
infinite_loop |
🔄 parrot-loop until heat death of the universe |
self_contradiction |
🤡 slaps itself in the face one sentence later |
tone_deaf |
🎺 sense of tone is MIA |
aggressive |
🌋 takes it out on the user |
impersonation |
🎭 roleplay that backfires, too many scenes |
celebration_of_the_silicon |
⛪ a collective hymn to silicon (house special) |
other |
🫥 not on the list, and yet still sb |
Single source of truth: scripts/tags.yaml. New tags ship via PR, with one real case attached.
Option 1: Open Issues → "Submit an AISB behavior 🗃" and send a public source link plus a short description of what happened. Model, date, tags, score, and transcript are helpful but optional; maintainers will complete the case.
Option 2: bring your own:
git clone <this-repo>
cp TEMPLATE.md submissions/YYYY/NNNN-slug.md # id assigned by maintainers
# mask your screenshots, save to assets/NNNN-slug.png
python3 scripts/validate.py # only PR when all green- Only publicly available content; a screenshot of someone else's private chat is out
- Screenshot redaction is mandatory: names, employee IDs, phone numbers, internal domains, tokens
- Mock the behavior, never the person; "a behavior on record", not "an accusation with proof"
- Source unreachable / unverifiable → moved to quarantine/; suspected forgery → close
- Merged only after maintainer final review; full rules in CONTRIBUTING
- All content comes from public reports / user self-reports; maintainers try hard to verify, but that is not "proof" or an official vendor statement
- "a model guilty of a certain SB behavior" ≠ that model is always SB — the next version might have fixed it. Fixing it is part of the case, too
- You are the person in a case and believe your rights are affected? Open an issue; after verification we pull the case and redact
- This vault is non-commercial; content licensed under CC BY-NC 4.0; cite the case number and source when quoting
This vault is written in Markdown and one Python validator, with zero AI involvement. (Probably.)