BazaarBaazi · Editorial
AI Disclosure
Where AI sits in the BazaarBaazi newsroom, what it does, what it does not, and how the human editor remains accountable for every byline.
BazaarBaazi uses large language models in its production workflow. This page is a full disclosure of where, why, and under what controls. It is updated whenever the workflow changes.
Where AI is used
- Story selection assistance. Each desk lead reviews a shortlist of candidate angles produced by an AI agent that scans public sources (exchange filings, regulator notices, primary news wires). The desk lead and the editor-in-chief decide what to commission.
- First-draft writing. A desk-specific AI agent writes a first draft from a brief that the desk lead provides. The brief contains the angle, the named sources, the verdict thesis, and the structure. The first draft is then edited by a human.
- Structured-data extraction. Tables, anchor lists, OI distributions, and other structured data are extracted from source documents using AI tooling and verified manually before publish.
- Headline rewriting. AI generates headline candidates that are scored and shortlisted; a human picks the final headline.
- Translation. The Hindi mirror at /hi/ (when live) will be machine-translated and reviewed by a Hindi-fluent editor before publish.
Where AI is not used
- Final sign-off. Every published story is signed off by the editor-in-chief, a named human.
- The verdict. The editorial verdict at the foot of each story is written by the editor-in-chief, not generated by AI.
- Sourcing decisions.Whether a source is credible, whether anonymity is warranted, and whether an unnamed source has passed the editor's identity check are human decisions.
- Corrections review. Every correction request is reviewed by a human and the response is written by a human.
- Original reporting. Where a story relies on a conversation with a source, that conversation is conducted by a human and quoted under a human byline.
Models in current use
BazaarBaazi's AI workflow runs on Anthropic's Claude family of models (Opus and Sonnet at the time of writing) for editing, drafting, and headline rewriting; on a vision-capable model for image descriptions and alt-text; and on local LLM tooling for routine structured extraction. Models are reviewed quarterly and the active version is documented in the project's public technical record.
Audit trail
Every published story has a per-story log that records: the original brief, the first draft, the human-edited final, the timestamps of each edit, and the model versions used. The log is retained for two years and is available to readers on a per-story basis on request.
Why this matters
AI-assisted journalism becomes irresponsible when it removes the human editor of record. BazaarBaazi's commitment is the opposite: AI takes the boring parts of the job (extraction, formatting, first-draft prose) and frees the human editor to do the work that only a human can do well. That is, deciding what runs, what the verdict is, who to call before publish, and what to retract when we get it wrong.
Reader feedback on AI use
If a reader believes a published story shows signs of poor AI oversight (factual errors, generic phrasing that contradicts the named author's usual voice, sources that do not check out), they are encouraged to flag it via [email protected]. Such reports are treated as priority correction requests.