The study concerned identifying news articles from excerpts. It is not a measure of every AI answer or present-day versions. Incorrect identification, fabricated links, and refusals are different outcomes.
What the research asks buyers to notice
The Tow Center’s March 2025 report examined AI search tools on a specific source-identification task involving news excerpts. The report and its coverage described problems with correct identification and citations. The findings are a reason to inspect evidence, not a universal error rate for all AI work.
A link can exist without supporting the sentence beside it. A recognizable publisher can be cited for the wrong article. A refusal to answer is different from confidently inventing a reference. A meaningful review preserves those distinctions.
Evidence should be usable, not decorative
For an important business claim, ask for the supporting passage, source date, and a clear explanation of what the passage does and does not establish. Treat unavailable evidence as a gap to resolve, not permission to fill the space with confident language.
- Open the source for a material factual claim.
- Separate quotations, supported facts, inference, and uncertainty.
- Check current information before using an older source for a current decision.
A PRACTICAL THADDEUS WORKFLOW
Ask for a research trail you can challenge.
Deep Research combines source inspection, counter-evidence, and claim classifications. Use those capabilities to build a brief that lets a reviewer see both the support and the gaps, instead of mistaking length or confidence for depth.
Try this assignment
Compare these suppliers against our approved requirements. For each material claim, include the source, relevant excerpt, date, and evidence classification. Look for counter-evidence and mark inaccessible or unsupported claims. Do not invent missing references.
Claim labels and citations do not guarantee truth or freedom from bias. Important conclusions still need review. No matched test here establishes that Thaddeus outperforms the products in the historical study.
Sources & further reading
- We compared eight AI search engines. They’re all bad at citing news. ↗
Tow Center for Digital Journalism / Columbia Journalism Review. March 6, 2025. - AI search engines fail to produce accurate citations in over 60% of tests, according to new Tow Center study ↗
Nieman Lab. March 2025 coverage of the bounded study.
Educational information, not a security audit, professional advice, or guarantee. The assignment is an example to evaluate, not a completed demonstration. Historical incidents and product capabilities do not establish comparative safety. Confirm current policies and agreements for your intended workflow.