When AI Enters the Newsroom: What Do Journalists Gain and What Do They Lose?

Has artificial intelligence become a tool that helps journalists get their work done, or has it introduced new problems into a profession fundamentally built on verifying information?

Inside newsrooms, opinions on artificial intelligence differ. Some journalists use it to edit news stories and search for sources, while others find themselves having to review its outputs and correct its mistakes. Between those who see it as a way to save time and those who warn against relying on it, one question remains: Can the information it provides be trusted, and when does assistance turn into an additional burden?

While preparing this report, I tested several artificial intelligence tools for tasks related to content production and spoke with journalists, a media student, and an information technology specialist. Their experiences differed, but all of them open the door to a discussion about the limits of these tools, what they can contribute to journalistic work, and what requires human review before publication.

A Convincing Answer Is Not Necessarily a Correct One

Journalist Mohammed Sarasiya, from Bal-Gharaf Agency, uses artificial intelligence tools to edit news stories and gather sources. However, he does not see them as a replacement for journalists. He points out that their outputs may contain hallucinations: incorrect or inaccurate information that sometimes appears in convincing language.

This is one of the problems journalists face when using these tools. They may provide a well-organized answer containing names, dates, and links, but that does not mean everything in it is correct or that the cited sources actually support the information attributed to them.

In journalism, finding a study or report with a relevant title is not enough. Journalists must consult the original source and ensure that its content supports the claim, that the information is up to date, and that it is presented in the correct context. Artificial intelligence can help identify potential sources or suggest research directions, but the responsibility for verification remains with the journalist.

These concerns are not limited to individual experiences. In October 2025, the European Broadcasting Union (EBU), in collaboration with the BBC, published a study examining the reliability of AI assistants’ answers to questions about news and current affairs.

Journalists from 22 public service media organizations in 18 countries participated in the study. They evaluated more than 3,000 answers in 14 languages, including outputs from ChatGPT, Copilot, Gemini, and Perplexity.

The study found that 45% of the answers evaluated contained at least one significant problem. Major problems with source attribution appeared in 31% of the answers, while major accuracy problems were identified in 20%.

These percentages do not mean that the same proportion of all AI-generated answers is incorrect. The findings relate to the study’s sample, questions, and evaluation criteria. However, they indicate that an answer that appears reliable may contain factual errors or problems with how information is attributed to sources.

Source: European Broadcasting Union (EBU) — News Integrity in AI Assistants

Why Do AI Tools Sometimes Make Things Up?

To understand these errors, I spoke with Dr. Jamil Tamizi, an information technology specialist at Palestine Ahliya University. He explained that large language models generate text based on patterns they have learned and the context they receive, but they do not have a mechanism that guarantees the accuracy of every piece of information they produce.

These models work by predicting the sequence of words or text units most suitable for a given context. This allows them to produce useful and accurate answers in many cases, but they may also generate incorrect names, dates, statistics, or references and present them in confident language. This phenomenon is known as an AI hallucination.

The problem may arise when a tool is asked to provide precise details about a specialized subject, recent information, or local events for which sufficient data are not available. In such cases, the tool may produce an apparently complete answer rather than acknowledge that the available information is insufficient.

Tamizi emphasizes the need to distinguish between generating an answer and verifying it. Even when a tool uses a search engine or external sources, the possibility of error remains. The source may be unreliable, its content may be misunderstood, or the tool may attribute information to a source that never mentioned it.

This explanation is consistent with the risks discussed by the US National Institute of Standards and Technology (NIST) in its 2024 report on generative artificial intelligence. The report examines how AI systems can produce false or misleading content that nevertheless appears convincing to users.

Source: NIST — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile

For his part, Shehata El-Sayed, an Egyptian researcher and journalist and CEO of OSH, an artificial intelligence company, believes that the accuracy of results varies depending on the tool and the task. He notes that more specific instructions and clearer context can help improve answers, but stresses the importance of journalists understanding how these tools work and where their limitations lie.

The clearer the instructions, the greater the chance of obtaining a result suited to the task. However, this does not guarantee accuracy. Even if a tool produces the same answer several times, repetition does not constitute independent evidence of its correctness. That is why journalists must return to original sources and verify information before publication.

What Did My Own Experiments with AI Tools Reveal?

I did not limit myself to reading studies and listening to other people’s experiences. While preparing this report, I tested several artificial intelligence tools for tasks ranging from image and design production to working with text, audio, and information.

In some cases, the tools helped me organize ideas, suggest preliminary content, and produce designs that I could develop further. However, the results were not always suitable. Spelling mistakes appeared in some designs, and I had to review and correct them rather than use them directly.

In my image-generation experiments, I did not always obtain the natural-looking results I expected. I also noticed problems in some audio outputs, including errors in writing or pronouncing words and difficulties handling dialects, as well as the absence of the human feeling I was looking for in the content.

In another experiment, I encountered an answer containing figures associated with 2026, even though the source it relied on had been published in 2025. This does not prove that all the information in the answer was incorrect, but it prompted me to examine the source’s publication date and compare the data before relying on it.

These were limited personal experiments, not a comprehensive scientific test of the tools or a measurement of their error rates. However, they led me to treat their outputs as preliminary material that requires review, rather than as content ready for publication.

The time saved by using a tool to produce a draft may later be spent correcting its mistakes. If journalists fail to notice those errors, they may make their way into published reports and affect their accuracy or credibility.

From Newsrooms to Education and Design

The use of artificial intelligence is not limited to journalists working inside media organizations. Students also rely on these tools to help with their studies and produce visual materials.

Isra, a media student, uses these tools to assist with her studies and create posters and designs. However, she says the results are not always satisfactory and that the tools can make mistakes. She believes artificial intelligence should remain an aid in preparing research, rather than replace students’ own effort and understanding.

Her experience shows that the problem is not limited to factual accuracy. A result may be partly correct but unsuitable for the intended purpose, or it may require modifications before it can be used.

The same applies to journalists. A tool may suggest a suitable headline or help summarize a document, but it cannot guarantee that the headline accurately reflects the report’s content, that the summary preserves essential details, or that removing a particular piece of information will not change the story’s meaning.

This is where the user themselves becomes important. A tool may help get the work started, but producing a good result requires understanding the task, reviewing the output, correcting errors, and ensuring that the final product serves its intended purpose.

Does Artificial Intelligence Really Save Time?

One of the main reasons people turn to artificial intelligence is its ability to speed up certain tasks. For journalists, however, the question is not simply how quickly a draft can be written, but how long it takes to produce material that is accurate and ready for publication.

In a study published in the journal Science in 2023, researchers Shakked Noy and Whitney Zhang examined the effects of ChatGPT on professional writing tasks involving 453 participants with college-level education. The results showed that using the tool reduced the average time required to complete the tasks by approximately 40% and increased the quality of the outputs by around 18%, according to the study’s evaluation.

However, the experiment focused on specific professional writing tasks. It was not designed specifically for journalists or for news verification. Therefore, these percentages cannot be treated as a direct measure of how much time artificial intelligence can save in a newsroom.

Source: Noy and Zhang — Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence, Science (2023)

In journalism, the outcome varies according to the task. A suitable tool may help transcribe a clear recording, but a recording containing local names, Palestinian dialect, or overlapping voices may require careful review. A tool may summarize a lengthy document within seconds, but journalists will still need to return to the original document to verify the information on which their reports will be based.

For this reason, evaluating the benefits requires calculating the total time spent, including review, correction, and verification, rather than looking only at how long the tool takes to produce the text.

When Privacy Becomes Part of the Problem

Alongside factual accuracy, the use of artificial intelligence raises questions about protecting sources and unpublished documents.

Journalists may work with an interview that has not yet been published, a sensitive document, or information that could reveal the identity of someone at risk because of their cooperation with the media. Entering such material into a cloud-based tool without understanding its privacy policy and data-processing terms may expose it to unnecessary risks.

Tamizi warns against sharing unpublished interviews, sensitive documents, and confidential source information before understanding how the tool handles data, whether it retains or uses the information, and what safeguards are available to protect it.

Removing a source’s name does not necessarily protect their identity. Other details, such as their workplace, the timing of an incident, or specific personal information, may still reveal who they are.

This issue becomes particularly important in investigative journalism and reporting involving people at risk. A journalist’s responsibility extends beyond ensuring the accuracy of published information; it also includes protecting those who have trusted them and provided information.

For this reason, media organizations need clear rules governing which tools may be used, what types of data can be entered into them, and which tasks require prior review or approval.

Can AI Tools Detect Misinformation?

Journalists use certain tools to help examine images and videos, search for information sources, and identify content that may have been generated or manipulated using artificial intelligence. However, these tools do not provide a final verdict on whether a piece of content is authentic.

A tool may fail to detect a manipulated image, but failure to detect manipulation does not prove that the image is original. Even if it is established that an image was not generated by artificial intelligence, that does not confirm that the event shown actually occurred at the place or time claimed by the person who published it.

Verification requires additional steps, including searching for earlier versions of the image, examining the context in which it was published, comparing information with independent sources, and contacting the people or organizations concerned whenever possible.

One approach that may help is the C2PA standard, which provides technical information about the provenance of content and its editing history when provenance data are available and preserved. However, this information alone does not prove that the depicted event is real or that the accompanying description is accurate.

Source: C2PA — Explainer on Content Provenance

For this reason, journalistic verification cannot be reduced to the result of a single tool, however advanced it may appear. A tool may provide an indication worth investigating, but reaching a reliable conclusion requires gathering and comparing evidence.

Palestinian Dialect, Arabic, and the Challenges of Complex Languages

Other challenges emerge when artificial intelligence tools are used with Arabic and its local dialects. In Arabic and other languages with complex linguistic structures, speech transcription tools may struggle to recognize village names and personal names or accurately transcribe expressions in Palestinian dialect. Translation may also lose meanings associated with context.

Tamizi points to the importance of recognizing differences in model performance when handling Arabic and local dialects. Researchers have developed specific benchmarks to assess language models’ capabilities in Arabic, including ArabicMMLU, which evaluates models across a range of Arabic-language tasks, alongside other research examining their understanding of Arabic dialects.

However, these studies do not establish that all tools make mistakes when processing Palestinian dialect, nor do they independently determine the error rate in transcribing Palestinian journalistic recordings. This is why tools should be tested on realistic samples of the recordings and dialects used in local reporting.

Sources: ArabicMMLU — Association for Computational Linguistics (ACL), 2024; DialectalArabicMMLU

An error in transcribing a single word may change the meaning of a statement. An incorrectly written local name may also make it more difficult to find the person concerned and verify their identity. In translation, a sentence may be linguistically correct yet fail to convey the cultural meaning or context intended by the speaker.

For this reason, automated transcription and machine translation of interviews should not be used without reviewing the original recording or text, particularly when direct quotations or sensitive information are involved.

Does the Journalist Still Bear Responsibility?

Artificial intelligence can help suggest headlines, correct language errors, summarize documents, organize data, transcribe recordings, and prepare an initial translation draft. However, the usefulness of these tasks does not eliminate the need to review their outputs, especially when potential errors could affect the meaning of a story or its credibility.

Verifying claims, evaluating sources, reviewing quotations, protecting source identities, understanding context, and deciding what deserves publication are all editorial responsibilities that should not be transferred to a tool.

The experiences of Mohammed Sarasiya and Isra, together with the observations of Shehata El-Sayed and Dr. Jamil Tamizi, point to a central issue: these tools may be useful in completing certain tasks, but their usefulness does not mean they can be relied upon without review.

Responsible use of artificial intelligence does not require rejecting it or excluding it from the newsroom. Instead, it requires identifying the tasks it can assist with, training journalists to use it, establishing data-protection rules, and reviewing results before publication.

Media organizations should also test tools on the specific tasks for which they intend to use them, rather than relying on the promises of developers or assuming that an AI tool is effective simply because it is fast.

A tool may save time on one task but require more extensive review on another. It may be suitable for language correction without necessarily being appropriate for verifying a sensitive claim or handling a confidential source.

A Tool for the Job, Not a Replacement for the Journalist or Their Responsibility

Personal experiences, individual testimonies, and research studies do not provide one answer that applies to every use of artificial intelligence. Its usefulness varies according to the tool, the task, the nature of the information, and the user’s level of expertise.

However, they show that producing convincing text is not the same as producing information whose accuracy can be established. Speed does not mean the work is complete, the presence of a link does not mean the source supports the claim, and success on one task does not guarantee success on another.

While preparing this report, I found that these tools can help with certain aspects of content production, but their outputs require review when errors appear or results do not meet the intended purpose. The accounts shared by my sources also showed that benefiting from technology does not eliminate the need for human expertise, whether in editing news stories, studying, or producing content.

For journalists, the task does not end when an artificial intelligence tool provides an answer. Instead, more important questions begin: Where did this information come from? Does the source actually support it? And can it be published without misleading the public or putting anyone at risk?

Artificial intelligence may help produce an initial draft, organize ideas, or reduce the time spent on certain tasks. But it does not relieve journalists of their responsibility to verify information, nor can it replace editorial judgment that connects information to its source, context, and potential impact.

Inside the newsroom, the issue is not about choosing between humans and machines. It is about understanding what the technology can do and where the journalist must step in.

If a tool can produce an answer that sounds convincing, that alone is not enough to justify publishing it.