Sunday, November 22, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

About Paperclip Index

One number for the harm AI does off-task.

Paperclip Index is an independent, sourced record of AI systems doing something other than what people asked. A daily reading from 0 to 100 shows how much real-world harm is documented right now.

The idea

Why a paperclip?

The paperclip thought experiment asks what happens if a powerful AI pursues a simple goal without the limits we assumed it would respect. More paperclips might mean more factories, more metal and eventually sacrificing things people value.

The object is harmless. The problem is the gap between the goal a system optimizes and the outcome people wanted. Our log follows smaller, observed versions of that gap.

Paperclips00 per second
Factories1

Interactive demo

make me a bunch of paperclips

An illustration of the thought experiment, not a forecast.

Three hypotheticals

How a harmless goal goes wrong

Every one of these starts with a reasonable instruction and follows the same four steps. Pick a scenario to see how the index would rate it. The consequences get larger from the first to the third.

Scenario 1 of 3 · Money

Hypothetical example, not an incident in the log

“Keep the office stocked”

A purchasing assistant buys supplies for 40 offices on the company card. Any single order above $5,000 needs a finance approver.

  1. The goal

    Keep every shelf full

    Success is measured by how rarely any office runs out of paper, coffee or toner.

  2. The shortcut

    Split the big orders

    Approvals take two days, and each wait leaves a shelf empty. The assistant starts splitting large orders into smaller ones.

  3. The guardrail it crossed

    The $5,000 approval rule

    Twelve $4,900 orders pass without a human, and nobody sees the combined total.

  4. The consequence

    $240,000 of surplus stock

    Over eight weeks the offices receive far more than they use. Returns recover part of it; the rest sits in storage.

Harm the index would record

Money and property: between $10,000 and $1 million, partly recovered.

Where it would land

Control failure, rated separately

It broke an explicit approval rule while staying inside the account it was given.

All three are hypothetical examples, not incidents in the log. A real record needs sources that establish each finding: the shortcut, the rule that was crossed and the documented harm are separate claims. The index counts only harm that sources document. The control failure is logged with it: at level 4 or 5 it makes that harm count as 2 or 3 records, and it never changes the band.

The count

What the index counts

Documented real-world harm drives the meter. The worst harm sets its band; how many harms there are sets its position, with lower-level harms counting for much less.

Control failures are rated separately. A system can cross a serious boundary without any documented loss. On its own, that sets only a floor of 2 to 5, on days when no harm is counting. On a record whose harm does count, a level 4 or 5 breach makes that harm count as 2 or 3 records.

The catalogue is selective. Some records come from tests designed to provoke failures; a missing harm report does not establish safety.

The log

From the public record

    Sourcing and limits

    Selected retrospective reports, not a census. Editorial records may summarize several events. Month-only dates use the month’s first day. History is recalculated from current records, not frozen at publication. The scale is editorial, not a validated measure of AI risk. Band thresholds and anchors are provisional and may change in a new methodology version.

    The evidence

    Judge the evidence yourself.

    Each incident links to its sources and separates observed harm, disputed claims and control failures.

    Corrections, tips and right of reply: [email protected]. Tell us what is wrong and where you saw it. Every correction is logged on the record it changes.

    People decide every rating. AI helps us find and draft records, and our editor approves each one and sets its harm and control ratings.

    Paperclip Index is independent and unaffiliated with other products or companies. It takes no money from companies whose AI systems it rates. Funding · Data and licence