Making bounded attribution claims

Purpose and when to use it

Use attribution when knowing the origin or responsible workflow would change the conclusion. Separate identifying a deployed model configuration from identifying an account holder or a human operator.

Evidence required

Relevant material can include generation records, model configuration, task identifiers, account logs, known-source controls, and validated content measurements. Document how candidate-source labels were established.

Principles and procedure

• Name the attribution level and competing candidates, including unknown source.
• Identify the link between each evidence object and each candidate; distinguish labels from verified records.
• Test the method on relevant positives, negatives, and unknown-source conditions.
• Seek independent source records and describe residual ambiguity before naming a source.

What to document

Keep candidate selection, controls, thresholds, environment, error observations, and source dependencies. Record whether a candidate was excluded by evidence or merely absent from the test.

Failure modes

A closed-set classifier always choosing one candidate can create false certainty. Shared training, model lineage, copied output, and reused account credentials can support alternative explanations.

Limits on the conclusion

Exact model or person attribution may remain unresolved even when a broad source class is well supported. Results from one generator set may not transfer to newly deployed models.

Independent checks and corroboration

Correlate content associations with task and deployment records that were collected separately. An account log can support account use without proving which person controlled it.

How to report the result

State the supported level, alternative sources considered, tested errors, and unresolved relationships. Do not convert a classifier score into a probability of human responsibility.

Related methods

• The complete evaluation method
• AI detection limitations
• Writing calibrated confidence statements

Related glossary terms

• attribution
• model attribution

How do you name the supported source level?

Ask whether the evidence identifies an application, account, model family, deployment, or operator. Test appropriate alternatives at that level, including unknown sources and shared infrastructure. Keep independent support separate from repeated versions of one assertion.

A conclusion can associate a request with an account while leaving the operator unknown. Name that boundary directly instead of allowing a precise account identifier to imply a precise human identity.