AI in Journalism: A Guide to Verifying, Researching, and Writing

Where can AI concretely help a journalist?

AI can intervene across the entire information chain, from formatting, verification, and distribution of content. A journalist can use it as a tool in daily practice, and this is what the Revue des médias mapping by the INA describes [22].

It is mainly useful for time-consuming tasks:

• Document research: it helps explore a topic and identify sources, for example with Perplexity and AI search engines, cited in a CFPJ training course [20].
• Transcription: speeding it up frees time for analysis [13].
• Synthesis and data analysis: a practical guide lists these among possible uses [25].
• Verification: AI is used for fact-checking, along with translation, according to a study of Chadian media [1]. For cross-checking sources and verifying figures, it can speed up the work without fully replacing the journalist [27].

The common principle is assistance: AI relieves the journalist of repetitive steps to free up time for analysis [13]. It does not substitute for them.

What the sources allow us to claim

Several of these sources are commercial: a tool publisher [25], a training school [13], and a professional training provider [20]. They have an interest in promoting the uses they describe. Others are partial extracts, such as the study of Chadian media [1]. The BnF summarizes the stakes by saying that AI "promises" to transform work organization in media [12], which remains a promise. The file therefore does not make it possible to quantify time savings or error rates.

How can AI be used for document research and transcription?

For document research, AI serves journalists in four tasks: exploring a topic, identifying sources, transcribing recordings, and summarizing documents [25]. It speeds up research, interview transcription, and cross-verification, freeing up time for analysis [13].

The uses

• Exploring a topic and identifying sources: CFPJ trains people to use Perplexity and AI search engines for this purpose [20].
• Transcription: Trint is cited for audio and video transcription with editing features [11].
• Synthesis and analysis: the publisher Delos describes a feature that analyzes websites, verifies sources, and produces sourced reports with key quotations and summaries [32].

The tools cited. A blog ranking names Perplexity, NotebookLM, Pinpoint, and Happy Scribe for research, transcription, and monitoring [23]. These lists come from a blog, a transcription vendor [11], and training organizations [13][20]. These are commercial sources: they catalog tools but do not evaluate them independently. Test them on your own cases before adopting them.

Why require cited sources. A synthesis without references forces you to re-verify everything. A tool that cites its sources lets you trace back to the original documents, which you cross-check with primary documents anyway, because AI can hallucinate [25]. Oupi conducts sourced research and cites its sources: each claim can be linked to the document it comes from before being reused in an article.

A practical rule: treat AI output as a lead. Only the original document, read by you, can support a published piece of information.

Can AI do fact-checking, and how far can it be trusted?

Yes, AI can speed up fact-checking, but it prepares the ground and the journalist makes the final call. Cross-checking sources, verifying figures, contextualizing statements, and identifying precedents require hours of painstaking work. According to a specialized website, AI can accelerate this phase "without fully replacing it" [27].

What AI does usefully

• Verifying images and locations: reverse image search on the web, exploring satellite images to locate a building, geolocating content [28].
• Checking against databases: information can be compared with large databases [28].
• Identifying claims: Factiverse is a patented tool that identifies claims made by public figures in articles and triggers verifications [37]. It is presented by Orange's innovation site, a commercial player.
• Consumer-facing verification: Vera, a French AI, presents itself as "the trusted number for checking facts." It is described on an educational site of the Toulouse academy [36].

Where trust stops

Final verification remains human. In the case study of the Tunisian television archives, verification and review of AI-generated analyses are entrusted to humans, in line with the ethical charters of the journalism profession [8]. A study of Cameroonian journalists places the imperative of verification among the pillars of journalism to be preserved in the face of AI [6].

In practice, a tool's output is treated as a lead, never as proof: you trace it back to the primary source before publishing. A tool that conducts sourced research and cites its sources, such as Oupi, facilitates this checking, since each claim can be compared against its origin.

Can AI write articles, and what should be reserved for the journalist?

Yes, AI can write, provided it is confined to standardized factual content and analysis and angle are reserved for the journalist.

It significantly speeds up the production of factual articles such as sports results or financial reports. Best practices recommend that journalists keep analysis and editorial angle [40]. Writing tools save considerable time and ease the blank-page syndrome, but they require human review and risk producing generic content [29]. These two sources are specialized websites, not studies: they describe recommended practices, not measurements.

Automation is not new. For the 2015 departmental elections, Le Monde used a robot to write 36,000 articles [38].

As for the current scale, Slate reports that "one article in ten would be written by AI" in the American local press [39]. This figure is stated conditionally, and the available excerpts do not give its methodology. It should be read as an order of magnitude, not as an established statistic. The same article indicates that the Syndicat national des journalistes recommended in 2025 banning any publication entirely produced by AI without explicit mention [39].

In practice:

• Delegate texts with repetitive structure and already established data.
• Keep the analysis, the angle, and responsibility for what is published [40].
• Check every generated text. Some newsrooms use a grid covering factual accuracy, editorial quality, and respect for journalistic values, with several levels of validation [21]. In the example cited by IBM, human editors verify sources and cite original information through links [31].

A transparency issue remains. Studies evaluate tools such as JENNI IA for writing web articles and question how content is validated [2]. A journalism festival has also raised the question of distinguishing what is written by journalists from what is produced by AI [3].

What ethical and professional rules govern the use of AI?

The ethical framework for AI currently rests on professional recommendations and newsroom charters, not on a single rule. It establishes a simple principle: AI does not exempt the journalist from any obligation.

The CDJM: mastering the tool. The Conseil de déontologie journalistique et de médiation encourages journalists and newsrooms to train on AI tools, to understand what they produce and identify their risks, biases, and limits [24]. It also reminds that the journalist, whether a simple user or someone participating in configuring these services, remains bound by the rules that underpin journalistic ethics [24].

The SNJ: transparency toward the public. In 2025, the Syndicat national des journalistes published a recommendation calling for a ban on any publication entirely produced by AI without explicit mention [39]. Several newsrooms have adopted usage charters [39]. The SNJ has also examined the arrival of generative AI in newsrooms [16].

The INA: an overview of charters. The INA's Revue des médias reviewed 20 media charters devoted to AI [38]. This corpus makes it possible to compare existing commitments before drafting one's own.

Research: a question of power relations. A research article analyzes the ethical issues of AI in the press as power relations that reconfigure the mediations of knowledge [4]. The Groupe d'études et de recherche Journalismes also places the issues surrounding AI and platforms within its research on journalism, particularly from the angle of journalists' professional identity [9].

In practice, these texts converge on three points:

• train on the tools and their biases [24];
• uphold the ethics of the profession, whatever the tool [24];
• explicitly flag any content produced entirely by AI [39].

How can AI use be organized in a newsroom?

A newsroom can frame AI use with five elements: a usage charter, permanent editorial supervision, an evaluation grid, multi-level validation, and training.

• Usage charter. It sets out what is permitted, what is prohibited, and what must be disclosed. Training courses place it at the heart of their programs: CFPJ offers to draft the first rules of a usage charter [20], and the EMI to develop writing guidelines [33]. These offerings are commercial or institutional.
• Supervision and evaluation grid. According to Narratiiv, a training organization, permanent editorial supervision is a golden rule. A dedicated team examines each generated piece of content against a grid: factual accuracy, editorial quality, and respect for journalistic values [21].
• Multi-level validation. The same source describes validation circuits with several levels before publication [21]. If the tool cites its sources, as Oupi does in its research, each claim can be checked against its source. Control remains human.
• Training. The CDJM encourages journalists and newsrooms to train on AI tools to understand what they produce and identify risks, biases, and limits [24]. ESJ Paris aims to integrate AI without degrading editorial quality [30], and CFPJ to gain efficiency without sacrificing rigor [34].

Initial training is also concerned. Work on Chadian media suggests introducing AI into journalism schools through dedicated modules [1].