Shared risk worksheet: a fictional worked example

Stochastic Parrots and Beyond · fictional example

A completed shared risk worksheet. The person, the project, and the vendor are invented for teaching.

ImportantFictional example

The person, the project, the vendor, and the contract are invented for teaching. No real submission, person, or product is described.

A doctoral candidate in political science plans to use a hosted chatbot. The chatbot is licensed by the university. The candidate wants to translate and summarize 40 French policy documents for a literature review.

A completed worksheet for a fictional research use

Field	Notes
Research use	Translate and summarize 40 public French policy documents for a literature review.
System or vendor	A hosted chatbot that the university licenses from “Vendor X”. Record the version and the date of use.
Evidence source	The vendor’s terms of use and the university’s license page. No published documentation of the training data.
Lifecycle stage	Deployment (use of a hosted service) and evaluation (checking the translations).
People affected	The candidate; the readers of the review; the authors of the documents, who can be misrepresented; data workers behind the model (unknown).
Risk or benefit	Benefit: a faster first reading. Risks: mistranslated legal terms, invented content in summaries, an unknown energy cost per request.
Severity and uncertainty	Medium severity for the conclusions of the review. High uncertainty about the training data and the error rate for legal French.
Existing mitigation	The license contract keeps uploads out of model training. Fictional contract term.
Possible alternative	Read the originals with a dictionary; test a smaller local translation model; ask a bilingual colleague to check a sample.
Disclosure requirement	State the tool, the version, the date, and the purpose in the methods section. Follow the AI policy of the target journal.
Institutional owner	The supervisor and the research data office.
Next action	Within two weeks, check ten passages with legal terms from different documents. One wrong legal term or invented claim ends the use of summaries. This is a first check, not proof of reliability.

Decision in this example. The candidate uses the chatbot for a first reading only. Quotations and arguments come from the originals. The methods section discloses the use.

• Constraints: The documents are public, and the license keeps uploads out of model training. Both conditions hold.
• Comparison: A faster first reading outweighs the risk only because each claim is checked against the original.
• Accountability: The supervisor agrees. The candidate reviews the decision after the ten-passage check.

A non-use variant. With confidential interview transcripts, the constraint fails: the data would leave the university. The decision is non-use, whatever the time saved.

PDF ·Original handout, CC BY 4.0 · Stochastic Parrots and Beyond, Laura Wagner and Moritz Mähr.