About an hour before her report is due, journalist Maha Hussein*, 50, who has more than 15 years’ experience, opens Gemini on her computer. She types a short question and receives, within seconds, an answer that appears coherent and persuasive. In a crowded newsroom, where journalists race to submit their work before 3 p.m., Maha struggles to catch her breath under constant pressure from shift supervisors to finish on time.
In this race, artificial-intelligence tools appear to offer speed. Behind that speed lie more urgent questions: have journalists received adequate training? Do news organizations have written editorial policies governing these tools? Who is responsible for professional mistakes arising from their use?
These questions are borne out by the investigation. A survey conducted by the reporter, with 44 journalists participating at the time of publication, found that 86.4% used AI tools in Egyptian news organizations, despite the absence of institutional training and editorial policies regulating their use at 70.5% of the organizations represented. Meanwhile, 56.8% said they had received training outside their workplaces through professional, academic and union bodies, including Egypt’s Journalists Syndicate, ARIJ, Google, Arabi Facts Hub, Microsoft, CCIJ, the International Center for Journalists, Birzeit University and WAN-IFRA.
Public rules are also lacking. Only 40.9% of the Egyptian news organizations represented required journalists to disclose AI use in published work, while 59.1% imposed no disclosure obligation. The journalists’ survey also found that 29.5% did not require review of AI-produced or AI-edited content before publication, compared with 40.9% that did. Another 29.5% said review happened only sometimes. This increases the risk of publication errors and limits organizations’ ability to hold users accountable.
These figures reveal poorly regulated AI use in newsrooms. A Thomson Reuters Foundation study of journalists in the Global South and emerging economies found that 81% already used AI in their work, while only 13% had established AI policies.
Risks grow with language-model “hallucinations”—fabricated or inaccurate information produced at varying rates, as measured by Vectara’s leaderboard—and inadequate human oversight. Sports Illustrated’s publication of articles under invented author identities in 2023 illustrates the latter concern. Egypt’s constrained media environment compounds the challenge: the country ranked 170th of 180 in the 2024 World Press Freedom Index. Editorial policies and training are necessary to protect content quality and public trust.
Vectara’s leaderboard is an evaluation tool developed and maintained by the company to measure generative-AI hallucination rates using its HHEM model. The 1.8%–24.2% range cited in the verification table refers to rates recorded for models on the leaderboard in Vectara’s text-summarization test, rather than a general hallucination rate across all AI applications.
Shehata El-Sayed, CEO of OSH AI, an investigative journalist and AI engineer, says the survey demonstrates the absence of editorial policies regulating newsroom use. He maintains that no Egyptian news organization yet has a written policy specifying when and how AI should be used. The leading professional failures, he says, begin with nondisclosure, extend to publishing outputs without human review, and include relying on general-purpose models such as ChatGPT and Gemini without safeguards for sensitive sources and data.
He adds that even statements from government bodies have been written with ChatGPT and contained factual and narrative errors, deepening the trust gap between audiences and the media. He warns against ignoring algorithmic bias, particularly on political, social, gender and sexual issues.
Trust at stake
A separate survey conducted by the reporter among 61 members of the public found that 80.3% relied on Facebook as their leading news source, followed by news websites at 52.5%, television at 37.7%, Instagram at 24.6%, WhatsApp and YouTube at 14.8% each, X at 13.1%, and printed newspapers at 11.5%.
Only 26.2% reported trusting journalistic content, while 3.3% expressed complete trust. At the other end, 1.6% reported complete distrust and 13.1% said they did not trust it; 55.7% were neutral. Another 55.7% said their trust in the media had declined over the preceding five years. They attributed this primarily to inaccurate news (59.8%), misleading headlines and biased coverage (55.4%), misinformation (50%) and weak fact-checking (46.4%). These percentages do not add up to 100% because respondents could select several reasons, although the survey had 61 participants.
El-Sayed says declining public trust predates AI and reflects structural crises: the collapse of journalism’s economic model, reduced investment in journalists, falling advertising and subscription income, ongoing difficulties facing digital journalists, and organizations relying on “one-person crews” to perform multiple roles for low pay without sufficient training.
Dr. Mona Magdy Abdelmaksoud, an assistant professor in Cairo University’s Faculty of Mass Communication’s Radio and Television Department, agrees that mistrust predates AI and is linked to misinformation and declining confidence in the media. AI may add a new variable, positively or negatively, requiring deeper research. Journalist, fact-checker and trainer Ahmed Gamal also supports this view, calling for responsible AI use as assistance rather than a replacement for journalists.
Why are organizations falling behind?
Gamal says many Egyptian news organizations are still getting to know AI tools and have no written policies. The problem extends to limited use of more effective paid tools, either because of cost or managers’ doubts about their value. AI continues to be treated as a shortcut rather than a way to improve work.
He cites an unnamed news organization that used AI to reduce reliance on staff instead of developing journalists’ skills. Journalists are split between those who improve independently and those whose employers invest in their development, he says.
Many organizations he contacted about training showed no interest in equipping all journalists with verification skills, prompt engineering and AI ethics, sending only one or two editors instead. Efforts to improve remain dependent on management’s willingness to invest. Demand for training has increased in recent years with support from the Journalists Syndicate, but beneficiaries tend to be those already committed to developing their skills; most organizations still invest too little.
The most common failures involve weak verification because tools are not mastered, he says. Egyptian newsrooms lack specialist units to verify news, photographs and videos before publication, apart from a limited experience at Al-Masry Al-Youm. Verification often occurs only afterward through fact-checking platforms. Problems begin when trending social-media claims become news without sufficient scrutiny. He points to recurring decontextualization or inaccurate quotation, especially on talk shows, and misleading headlines written to chase trends at the expense of accuracy.
Gamal urges organizations to direct budgets toward training, professional tools or better salaries that help journalists build skills, rather than relying on free tools unsuitable for journalism. Some employers banned AI after careless use to meet production targets led to inaccuracy, hallucinations and misinformation. The issue is inadequate human oversight, he argues. He criticizes using AI to write news instead of supporting research and editing, manipulating pictures or video, and publishing generated material without disclosure. He cites viral advertisements that simulated accidents with AI despite breaching ethical standards. Training is also a challenge when nonspecialists offer courses using basic tools while other employers prohibit content production with AI.
Accounts from inside newsrooms
In some Arab news organizations, AI use has become a ready-made accusation capable of undermining journalists’ work in the absence of declared editorial rules. Rather than explaining assessment criteria and acceptable uses, organizations reject work on the grounds of AI use without evidence or transparent verification, opening the door to arbitrary decisions affecting professional and financial rights.
The experience of Egyptian editor and journalist Al-Shaimaa Farouk, 30, illustrates a problem faced by other journalists accused of using AI without being told institutional policies or given a chance to defend their work. Some ended their collaboration as a result.
Farouk has worked at major newspaper Al-Shorouk for seven years and previously freelanced for an Arab platform. She uses ChatGPT, Gemini and Google AI Studio, relying on self-teaching without employer-funded or personally funded training. She says her organization has no AI editorial policy and requires neither prepublication review of AI-produced or edited material nor disclosure of its use.
She believes AI has accelerated her work and has not identified professional errors resulting from her own use. She attributes newsroom mistakes to missing policies: the issue is how AI is used, requiring declared safeguards and disclosure mechanisms.
She describes a painful experience with an Arab journalism platform that refused a commissioned report, alleging AI use despite her insistence that she had used none. After discussions, the platform maintained its decision, refusing both publication and payment, although it had approved the report in advance.
“The lead editor told me: ‘We are not accusing you of using it to write the report, but we reject its use even for proofreading or audio transcription,’” Farouk says. “I explained that these supporting tasks do not diminish a journalist’s professionalism, and that I had not used them in the disputed report at all.”
The editor called this platform policy, although it had not been communicated to journalists. More strikingly, Farouk says, the editor told her in a friendly conversation that the platform deliberately withheld reasons for rejection and its AI policy “so journalists do not deceive us and take precautions.” It relied on special detection tools. Farouk ended the collaboration.
She was not alone. A colleague encountered the same response from the platform after resubmitting, in late 2025, a report written in 2021 without changing a single letter. It was rejected as AI-generated. Farouk questions detectors’ reliability: the report predated the widespread adoption of generative AI among journalists, yet the platform insisted its tools detected use without providing evidence or explaining verification. She says the same approach affected several contributors without published criteria or evaluation rules.
Why text detectors cannot be trusted
El-Sayed says text detectors cannot conclusively establish AI authorship. They rely on statistical and linguistic patterns such as sentence length, repetition, syntax and predictability, and can label human writing as generated. Text lacks a digital fingerprint comparable to images, he argues. A 2023 study in the International Journal for Educational Integrity tested 14 detection tools and concluded that they were neither accurate nor reliable. All recorded accuracy below 80%, and five did not exceed 70%. Human text could be mislabeled as AI-generated and AI text as human. Six of the 14 produced false positives: the likelihood ranged from 0% for Turnitin to 50% for GPTZero. False-negative rates ranged from 8% for GPTZero to 100% for Content at Scale.
He adds that improved language models, extensive training and frequent use have made outputs closer to human prose. Well-crafted writing may resemble model output, leaving journalists and writers vulnerable to false accusations and loss of rights.
If Farouk’s experience shows how AI becomes an accusation without policies, Mariam Mohamed*, 29, illustrates the other side: using tools without institutional training and controls can compromise journalistic accuracy.
“I identified around ten cases in one year in which AI provided inaccurate or fabricated information,” she says. “I do not trust any information it gives before checking it.” After a year’s use, she found erroneous information, invented sources, outdated data, language errors and privacy risks involving user-data retention. She checks every fact before using it.
Mariam began journalism five years ago but stopped several times because she had no permanent newsroom position. For the past two years she has freelanced for Arab websites. About a year ago she began using ChatGPT, Claude AI, Perplexity AI, Google Pinpoint and Elicit.
Organizations she works with allow use but have no governing policies, she says; review is left to the responsible editor. None trained her. She paid for training through specialist programs organized by the Journalists Syndicate with Arabi Facts Hub.
She uses AI for research, idea development and translation and says it improves speed and quality. But missing policies increase mistakes and place journalists in difficult professional situations. Training on limitations and risks is necessary, she stresses.
Nermin Mohamed*, 34, reports a similar experience. She began in 2014, stopped for several years and returned in 2022, accumulating five years of active experience across print, digital journalism and radio. She now freelances as a journalist and fact-checker.
Since 2023 she has used Gemini, ChatGPT, DeepL, Claude AI, Perplexity AI and NotebookLM for research, translation and idea development.
She has encountered inaccurate information, mistranslated technical terms, biased presentation, hallucinations, weak source verification, intellectual-property risks, data leakage and language errors more than ten times in the preceding 12 months, making verification essential.
She received training through Egyptian and international organizations she worked with, including Al-Manassa in Egypt and London’s Muwatin network in partnership with ARIJ. She sees less interest in AI training among most Arab organizations than international ones.
The journalists’ accounts support El-Sayed’s view that the crisis lies in how AI is used. Current models are general-purpose rather than journalism-specific. Development requires training for particular tasks while protecting data, sources’ privacy and journalists’ safety and limiting algorithmic bias.
Examples of mistakes
El-Sayed says newsroom studies reveal missing written policies and reliance on self-teaching, contributing to professional and factual errors. He cites coverage of secondary-school results: AI-generated stories featured links such as “View the secondary-school results PDF,” in a ChatGPT-like style, but contained misleading information. Egyptian and Arab outlets also circulated a video supposedly showing a truck being rescued after a tire burst. Most of it was later found to be AI-generated, yet it had been published as real—a preference for reach and excitement over accuracy and objectivity.
The crisis is not confined to misuse in newsrooms. El-Sayed says organizations have deployed AI for administrative, economic or political aims. Some announced adoption as early as 2018, but reviews showed reliance on general models, threatening journalists’ and users’ privacy and reflecting a poor understanding of the technology.
He says some have used AI without discipline to fabricate voices and videos, justify arbitrary dismissal of journalists and evade responsibility by blaming the technology.
Journalists remain the weakest link, El-Sayed argues, with mistakes rooted in limited training and experience. Risks include translating news uncritically and entering sensitive information about stories and sources, endangering both. Journalism’s future belongs to those who develop their skills and become knowledge producers, he says. He rejects insufficient budgets as an excuse for poor training, calling it an excuse worse than the offense, and notes that many experts would offer training free if organizations provided the right environment.
He says the response begins with written editorial policy, supported by governance defining responsibilities and accountability, and mandatory training for journalists and editors. He calls for digital-verification units, secure data-protection tools and updated journalism ethics codes.
Missing policies
Abdelmaksoud agrees that Egyptian news organizations have yet to develop integrated written institutional policies and that existing efforts amount to limited guidance. In her view, an unwritten policy does not exist.
A policy should define priorities, prohibited uses, ethical dimensions and applications consistent with institutional strategy, she says. Journalists who rely on outputs without checking or consulting original sources risk hallucinations. Replacing journalistic analysis and interpretation with AI, culminating in wholly generated content without human review, is among the most serious professional errors.
She has observed platforms relying on inaccurate information and secondary sources, copying AI-generated or misleading stories from other outlets without verification or attribution. This spreads misinformation and obscures its origin until fact-checkers or specialists intervene. She has identified such practices at more than one Egyptian outlet.
She cites a story ending with the phrase: “If you would like us to offer a different version in the form of a press release…” Such wording is typical of AI suggestions; leaving it in the published story exposes the lack of human review.
She rejects weak budgets as the main training obstacle. The problem is the absence of a strategy based on newsroom needs rather than global trends.
Training opportunities now exist through news organizations, training bodies, international organizations and self-learning. She therefore does not consider lack of training the fundamental problem, although some courses fail to accommodate different skill levels. Despite more Arabic-language training, journalists still do not master every tool. Written policy is only a beginning: it must become guidance, practical applications and training, developing into institutional practice and culture.
Farouk lost more than a report: she lost the right to be assessed under declared rules. Between journalists using AI without policies, institutions using or prohibiting it without standards, and audiences losing trust, the investigation locates the crisis in missing governance that regulates use and holds people accountable for errors.
This investigation was produced with support from the European Union and the Dutch government through the AI in Media Fellowship for the Middle East and North Africa. Its content does not necessarily reflect the views of the European Union or the Dutch government.
Generative-AI tools assisted in preparing this work. Its content was fully reviewed, and the accuracy of every fact and source was checked before publication. Efforts were made, as far as possible, to remove bias in language, framing and source selection. Full editorial responsibility rests with the author.
