AI tools are used by journalists for document synthesis, beat orientation research, interview question preparation, and background research. They offer genuine efficiency gains for specific, well-bounded tasks. They also carry specific risks in a profession where factual accuracy is the core professional standard and source protection is a professional and ethical obligation.
The AI research protocol for journalists distinguishes where AI saves meaningful time from where it creates risks that undermine journalism's fundamental obligations: to report accurately, to protect sources, and to distinguish between verified fact and allegation.
Where AI Genuinely Helps Journalists
Strong use 1: Document synthesis for large document sets
Investigations and FOIA responses generate hundreds or thousands of documents that need systematic review. AI can help identify the most significant documents within a large set, summarize long documents for context, and extract specific types of information across a document corpus.
The verification discipline: AI document synthesis is a navigation tool, not an evidentiary source. Every specific fact, quote, or data point used in a published article comes from the primary document, not from the AI summary. AI may miss significant documents; it may mischaracterize documents; it may summarize inaccurately.
Strong use 2: Beat orientation for new coverage areas
When a journalist begins covering a new beat or a new issue area, AI can provide an initial orientation to the institutional landscape, key actors, regulatory framework, and major recent developments. This orientation is faster to generate than reading through a large volume of background material from scratch.
The verification discipline: AI beat orientation is a starting point, not a research conclusion. The specific institutional histories, regulatory frameworks, and key personnel described by AI require verification against authoritative sources before use in reporting.
Strong use 3: Public records research framing
For FOIA requests and public records requests, AI can help frame the request: what types of documents to ask for, what agencies or offices are most likely to have relevant records, what exemptions to anticipate and how to argue against them. FOIA strategy is an area where AI can accelerate the thinking without being the primary source of factual claims.
Strong use 4: Interview question preparation
Given a topic area and context, AI can generate interview question frameworks — the types of questions to ask, the angles to explore, the follow-up questions that probe for specific information. The journalist then develops these based on specific story context and beat knowledge; AI provides the structural scaffolding.
For public figures — politicians, executives, public officials — AI can provide background orientation: their public career history, their major public statements and actions, their institutional affiliations and reported positions. This background is the starting point; any specific claim used in a story is verified against primary sources.
Strong use 6: Data and statistics context
When a story involves quantitative claims — economic statistics, health data, population figures — AI can help contextualize those numbers (what they compare to historically, what they compare to across jurisdictions, what they suggest about trends). The specific figures used in the story come from authoritative primary sources; AI provides the contextualizing frame.
Where AI Creates Genuine Risk for Journalists
Risk 1: Fabricated sources and citations
AI tools fabricate sources. When asked to identify sources to interview for a specific topic, AI may generate plausible-sounding names, affiliations, and contact details that don't correspond to real people or to real people with the described expertise. When asked to cite research, AI may generate citations to studies that don't exist.
The specific risk in journalism: a reporter who uses AI-generated source suggestions without independently verifying that the person exists, has the described expertise, and is reachable has conducted no actual source development. A reporter who cites an AI-generated research reference without verifying the study exists has fabricated a citation.
The rule: all sources used in published journalism are verified to exist and to have the attributed expertise or position through independent verification. All cited research is verified to exist and to say what it's attributed as saying. AI is not a source for either.
If a journalist describes a confidential source's situation, location, or information to an AI tool to get help drafting a question or framing a story, the input to the AI tool may expose information that could help identify the source. Public AI tools process inputs through external servers; the confidentiality of that information cannot be guaranteed.
The rule: no information that could identify a confidential source — their institution, location, specific knowledge, or anything that narrows identification — is entered into an AI tool. Story-specific inputs to AI tools are stripped of any source-identifying information before use.
Risk 3: Verification cannot be outsourced to AI
AI cannot verify claims. When an AI tool says "According to [source], [claim is true]," the AI is asserting plausibility, not establishing fact. AI cannot interview sources, review documents independently, or apply the professional judgment that determines whether a source's account is credible.
Journalists who use AI-generated research as if it were independently verified reporting are not using AI as a research aid — they are outsourcing the verification work that makes journalism journalism.
AI training data has a cutoff date. For beat reporting involving recent developments — current regulatory positions, recent corporate changes, new legislative developments — AI information may be months or years out of date.
Risk 5: Source credibility evaluation
AI cannot assess source credibility. When AI identifies a source as "an expert in [area]," it is identifying someone with public credentials in that area — not evaluating whether that person's specific claims are reliable, whether their institutional affiliation creates interests that affect their account, or how their statements have historically proved out in practice. Source credibility assessment is the reporter's job, developed through beat experience.
The Journalist AI Research Protocol
| Research need | AI role | Verification required |
|---|
| Large document synthesis | Navigation: identify significant documents, summarize for context | All specific facts verified against the primary document |
| New beat orientation | Initial institutional landscape and regulatory framework | All specific claims verified against authoritative sources |
| FOIA strategy | Frame requests; identify likely document custodians; anticipate exemptions | Specific request language reviewed against FOIA statute and agency rules |
| Interview question preparation | Question frameworks and angles to explore | Journalist develops questions from beat knowledge; AI provides structure |
| Public figure background | Career and public position orientation | Any specific claim used in a story verified against primary sources |
| Data contextualization | What statistics compare to; trend framing | Actual figures from authoritative primary sources only |
| Source identification | Not appropriate — AI fabricates sources | All sources independently verified to exist and to have attributed expertise |
| Research citations | Not appropriate — AI fabricates citations | All cited research verified to exist and accurately characterized |
| Source-identifying story inputs | Not appropriate — source exposure risk | No source-identifying information in AI tool inputs |
| Verification of factual claims | Not appropriate — AI cannot verify | Verification through independent sources and documents |
Document analysis tools (with journalism-specific features):
DocumentCloud, Trint (transcription), and similar tools designed specifically for journalism document work integrate text analysis features that work with the actual documents the journalist uploads — unlike general AI tools that generate from training data. These are generally more appropriate for journalism document work than general-purpose AI.
General AI assistants:
ChatGPT, Claude, Gemini — useful for orientation, question frameworks, document context, and background research. Not appropriate for source identification, citation generation, or any input that includes source-identifying information.
WebSnips for capturing and organizing web-based research with date stamps — useful alongside AI orientation because web captures are dated, citable, and verifiable in ways that AI output is not.
Before entering any information about a story or source into an AI tool:
1. Does this input contain information that could help identify a confidential source?
Any description of a source's institutional position (specific enough to narrow identification), location, knowledge domain (specific enough to narrow identification), or relationship to the story → Remove before input.
2. Does this input contain information provided by a source under confidentiality?
Information shared off-record or on background that reveals the source's knowledge → Do not input. The story context may be described in general terms; the source's specific knowledge is confidential.
3. Would a sophisticated adversary, reading this AI input, be able to narrow the identity of a source?
If the combination of information — institutional context, timing, subject matter, geographic area — narrows source identification significantly → Strip before input.
When uncertain: do not use AI for this input. Use the research tools that don't require sharing story-specific context with an external service.
A Worked Example: Beat Orientation for New Coverage
Situation: a reporter with five years on a state-level politics beat is assigned to add federal healthcare policy to her coverage.
AI use (appropriate):
Prompt: "Provide an orientation to federal healthcare policy coverage: the key regulatory agencies and their roles, the major ongoing policy debates as of [AI's knowledge cutoff], the types of documents most useful for healthcare policy reporting, and the categories of sources most relevant to understanding federal healthcare policy decisions."
This produces: CMS, FDA, HHS organizational context; major current policy areas (coverage markets, drug pricing, Medicare/Medicaid financing); key document types (CMS proposed rules, FDA guidance documents, Medicare payment rate announcements, Congressional Budget Office analyses); and source categories (academic health economists, former CMS officials, insurer trade association staff, patient advocacy groups, congressional health committee staff).
What the reporter then does:
- Verifies the regulatory landscape against current CMS and HHS website
- Identifies specific officials (by public title and role) from the agencies' current public directories
- Searches for actual people with expertise in each source category through journalism networks (IRE, health journalism organizations), academic publications, and public testimony records
- Identifies specific watchdog organizations, journalists who've covered this beat, and annual reports from relevant agencies
- Begins building actual source relationships
What AI did not do:
AI did not identify specific sources to call. AI did not provide verified current regulatory positions. AI did not produce citations to use in stories. It provided the map of the territory; the reporter verified every element of the map and built the actual source network through independent work.
Key Takeaways
-
AI tools in journalism have a clear use boundary. Strong for document navigation, beat orientation, FOIA framing, interview question structure, and public figure background. Not appropriate for source identification, research citations, source-identifying inputs, or verification — the foundational activities that distinguish journalism from speculation.
-
AI fabricates sources and citations. Any AI-suggested source is verified independently to exist and to have the described expertise. Any AI-suggested citation is verified to exist and to accurately reflect what's attributed to it. These are not optional verification steps.
-
Source protection applies to AI inputs. No information that could narrow the identification of a confidential source is entered into an AI tool. Story-specific inputs are stripped of source-identifying information. The standard: would a sophisticated adversary be able to narrow source identification from this input?
-
AI cannot evaluate source credibility. Source credibility — whether a specific source's account is reliable, whether their interests affect their account, how their statements have historically proved out — is the reporter's professional judgment, developed through beat experience. AI identifies public credentials; it cannot assess credibility.
-
AI-generated story context is a navigation tool, not an evidentiary source. Every specific fact, quote, and data point used in a published article comes from a verified primary source — document, interview, authoritative data — not from AI synthesis.
Conclusion
AI-assisted research for journalists creates genuine efficiency gains for specific tasks — document navigation, beat orientation, FOIA strategy, and interview question framing — while maintaining strict discipline on source identification, citation verification, source protection, and the fundamental verification work that makes journalism journalism. The reporter who uses AI within these boundaries saves research time without trading away the accuracy and source protection obligations that define the profession. The one who outsources verification to AI is not using AI as a research aid — they are replacing the professional judgment that distinguishes journalism from aggregation.
Related reading: Clip Articles for Later Reading.