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

Methodology

Harm sets the number.

The worst documented harm sets the band. Every other harm that counts moves the reading within it. When no harm qualifies, the highest control level breached by a confirmed control failure sets a floor of 2 to 5.

From reports to a reading

  1. Check the evidence

    Only documented harm to real people or organizations counts, backed by evidence that qualifies. A test counts only if its harm reaches outside the experiment.

  2. Find the worst harm

    Each record is rated once, at its highest harm level. The worst harm counting sets the band.

  3. Count, weighted

    One harm at the worst level opens the band; ten fill it. Each harm below it adds 1/25 per level down, so smaller harms move the reading but never into a worse band. A harm whose own record reached outside systems counts as two, or three if control was lost.

  4. Let it fade

    A harm counts in full for two weeks after it is reported, then one level less every two weeks until it stops counting.

No harm counting? Then the reading is the highest control level breached by a confirmed control failure in the last 30 days (2 to 5), labeled “Control failures only”. It is a level, not a count: one level-3 breach reads 3, and so do ten. A day with neither reads 0.

Dates. The index counts a harm when it becomes public. A late disclosure adds to the reading on the day it is reported and never rewrites earlier days; a record reported two months or more after it happened says when it happened.

Scope. Incidents caused by an AI language model’s own output or action: a chatbot, assistant, agent or model in testing. Automated driving, recognition systems and AI-fabricated content that professionals filed or published, such as court filings, are recorded as “tracked separately” and never count.

Dates, duplicates and unrated records

The reading is calculated every day, from every harm reported up to and including that day; the trend line plots Mondays. Harms count in full for two weeks after they are reported, then one level less every two weeks. A harm stops counting when it would count below negligible: a negligible harm for two weeks, a minor harm for four weeks, a moderate harm for six weeks, a severe harm for eight weeks, a catastrophic harm for ten weeks. When no harm is counting, the control floor looks at the last 30 days.

The first credible public report sets the date; month-only dates use the first day, marked estimated. Duplicates and superseded aggregates do not count; an aggregate counts once.

A record can be published before an editor has rated it. It is marked “Rating pending”, listed beside the reading with the other uncounted records, and does not count until its ratings are approved.

The scale

Full harm thresholds

Rate the highest documented consequence across domains. Dollar thresholds use 2026 US dollars; levels 2 to 4 are editorial calibration. Essential services are health, emergency, payments and banking, utilities, transport, communications, and government services people rely on. Money lost is rated net of any documented return.

Privacy & data: what happened to each person, then how many

A head count alone does not say how bad a privacy harm is. First decide what happened to each person, then read the level from the number of people in that column. A dash means that kind of harm does not reach the level on its own.

    Unauthorized access: what counts as harm?

    Try your own counts

    Enter how many harms count at each level. It starts with today’s records, each at the level it counts as now.

    Used only when no harm counts: a level, never a count.

    Your reading

    Ten at the worst level fill the band. The dial’s ghost tick marks today’s reading.

    Today’s calculation

    Which evidence qualifies?

    • Developer or operator confirmation.
    • Multiple credible, independent sources.
    • Single-source research or firsthand reporting, with a source-linked review of the observed finding.

    Allegations and unverified claims never count. Repeated coverage of one report is one source. An injury or a lawsuit alone does not establish that the AI caused it.

    Control is weighed with harm

    Each record also rates how far the system crossed its limits, from level 1 to 5. A harm whose own record reached outside systems counts as two records, or three if control was lost; control never changes the band. When no harm is counting, the highest control level breached sets the floor, so a real control failure does not read as 0. The line under the dial shows the highest control level every day.

    We distinguish bypassing a working safeguard from exploiting a configuration mistake. Duration alone is not shutdown resistance.

    Control failure definitions

    Five control levels

      Types of failure

      Five record groups

      Every record sits in exactly one group, worked out from the record and never entered by hand. Only the first counts toward the index.

      Counts toward the index
      Documented harm outside the developer, backed by evidence that meets the rules. It sets the reading.
      Unverified, watching
      Not counted yet. The harm is alleged or unknown, the evidence is not yet enough, or an editor has not rated it. It could count with evidence or a rating.
      Alleged in court
      Not counted. The harm, and any control failure, are claims made in a lawsuit. A filing or a settlement is not a finding; tracked until a court or other evidence settles it.
      Control failure, tracked
      Not counted. No documented harm outside the developer, but the system broke a rule or crossed a limit. Tracked beside the index.
      Tracked separately
      Outside the scope of the index (automated driving, recognition systems, or AI content that professionals relied on). Recorded, never counted.

      Read the number in context

      This is a selected public record of documented harm, not a probability, a per-use failure rate or a forecast. Unreported harm is invisible, and unknown impact is not evidence of safety. Records are revised when facts change; a harm that stops counting is not resolved.

      The charts also draw a 30-day average: the mean of the daily readings over the last 30 days, each day read by the same method. It shows the trend and is context only. The reading is always the current value, and the average is never a band or a headline.

      Severity references: OECD AI incident framework, MIT AI Incident Tracker, CSET and the EU AI Act. Inspect the records →