How is AI being used in journalism?

There’s no doubt that some executives in the news media industry have viewed artificial intelligence (AI) with anxiety. But a newsroom in the future without any AI tools is highly unlikely.

According to Nieman Lab¹, major newsrooms across the country have been designing new roles and initiatives around how to develop and use AI tools and processes while maintaining ethics and editorial standards.

Already, news publishers like the New York Times, Associated Press, the Washington Post, ESPN and Semafor have been open about investing in initiatives to explore how AI technology can be used alongside human journalists.

This article will look at how AI is being used in journalism, how top publications are thinking about their AI strategy, and what industry analysts think might be coming for the future of journalism.

Before generative AI got popular, news organizations were already using machine learning and some form of AI technology to assist them with social media monitoring², managing large datasets used in news stories³ and organizing engineering workflows⁴ for digital products.

Reporters have benefitted from natural language processing-based transcription services like Otter and Trint. News organizations have used AI algorithms from platforms like CrowdTangle and ChartBeat to analyze audience engagement and track trending topics on social media.

Early experiments like the AI-powered news app Artifact⁵, teased some of the potential ways AI might make news more fun, with features like summarize news in the style of gen Z.

More often than not, these types of tools enabled journalists to scale up their reporting, cut down on busywork, and decipher relationships or patterns within data they’re scraping from multiple sources.

The novelty that generative AI introduces is that it’s able to produce content such as written texts, audio, video and images. In some use cases, generative AI can help editors and reporters translate and transform their stories for different distribution channels. However, the technology can also be hijacked by bad actors to produce disinformation and deepfakes, which makes journalists’ jobs harder.

AI technology has hit an inflection point; industries are moving beyond the hype and working to practically understand what AI can and can’t do. But in order to keep pace with innovation and change, businesses must experiment with different ways of using AI. This includes in newsrooms.

Media industry analysts predict that generative AI can alleviate some of the more tedious backend work in newsrooms⁶. These include tasks like tagging, categorizing, adding metadata, headline and SEO suggestions, copyediting, organizing research, processing permissions and moderating comments⁷.

Generative AI might also be useful for language tasks that don’t require introducing new information not already present in the document it’s working on, a report ⁶ by the Reuters Institute suggested. These include summarization, translation, simplification, rewriting in different styles, and extracting copy for social media, newsletters and scripts.

Initiatives like the JournalismAI project by the London School of Economics (LSE) journalism think tank Polis has been collecting case studies⁸ on how AI is being deployed in newsrooms across the world. Here are some ways that the news industry is already using AI:

• Newsgathering

• News production

• Audience engagement

Major publications like AP, Bloomberg and Reuters have already been using computers and some degree of automation to look for news all over the world.

For example, Bloomberg has created their own large language model (LLM)⁹ trained on financial documents they’ve curated and the data on Bloomberg Terminal. It said that this model improves natural language processing tasks like sentiment analysis, named entity recognition and news classification related to financial terms.

Semafor on the other hand has collaborated with Microsoft and OpenAI to build an AI-powered “multi-source breaking news feed"¹⁰ called Signals. This AI tool is focused on research and help Semafor’s journalists search news sources in different languages from around the world. Human editors ultimately evaluate and verify sources, write summaries, and cite the original information through relevant links.

AP uses AI¹¹ to automate certain corporate earnings stories by synthesizing information from press release, analyst reports and stock performance. It had said that this capability allows its reporters to focus on more in-depth reporting. It’s also experimenting with using AI to detect breaking news events from social media alerts. Both the Financial Times¹² and the Wall Street Journal¹³ are working on AI models that can predict trending topics to inform potential stories for journalists and find gaps in coverage¹⁴.

AI can also surface relevant research as the starting point for investigative reporting. AI language models can help reporters hone in on sections of interest¹⁵ within troves of documents, regardless of the format. For local watchdog publications, AI might also identify anomalies in government audit reports in order to surface leads for its reporters¹⁶. AI can also process large datasets from campaign finance records, state legislation, civil complaints, municipal budgets and summarize the contents of those documents to help reporters¹⁷. A local newspaper in Norway, iTromsø’s, even designed its own AI-powered tool to scrap data from municipal archives, rank documents by relevance and extract key information that might turn into leads for stories.

AI-powered tools might also take and organize notes from local conferences or city council meetings, sort through tips and create transcripts of recorded videos. Many newsrooms are using some AI speech to text software for transcription and translation.

Reuters has found that AI highlights and summaries¹⁸ make it easier for reporters to search through archived videos for key people and moments. AI video highlights have also become useful in sports reporting. ESPN uses AI to identify clip highlights and generate recaps at scale¹⁹.

There are a suite of AI tools that can help journalists proofread, draft headlines and come up with outlines. In a post in October²⁰, the New York Times said that they do not use AI to write articles. However, they do use it to sift data used to investigative reporting, make audio versions of their articles and offer article recommendations. Sometimes, they can use generative AI to draft up potential headlines, summaries of articles, and first translations of their stories from English to Spanish. All of this is done with human oversight and goes through edits before publication.

The Washington Post partnered with text to speech software company Eleven Labs to offer AI-generated audio to accompany some of its written newsletters²¹. This audio can be added to a playlist in The Post’s app along with other offerings like podcast episodes and audio-first articles.

BBC has previously experimented with automated tools²² that might come up with a ‘rough cut’ of their video and audio programs.

NLP applications, like Newtral’s automated fact-checking tool and Duke Reporter’s Lab’s FactStream, can assist with identifying statements that need to be fact-checked. And editors seem to approve in general of AI-assisted proof-reading and copyediting.

AI tools to detect deepfakes are being tested and developed. However, experts warn²³ that they should only serve as the starting point of a verification process.

As publications compete for attention, AI summaries show promise to draw readership. In early tests, it boosted readership for a public broadcaster in Norway²⁴ and a daily newspaper in South Africa²⁵. A small Swedish newspaper found that including AI summaries increased the amount of time readers spent on an article²⁶.

The AI summary feature can also be presented to readers in the form of a chatbot. For example, The Washington Post has an AI chatbot trained on archive articles that can answer reader questions on climate science. The Post designed the tool with engineers from Virginia Tech²⁷. The AI uses retrieval augmented generation (RAG) and is trained on The Post’s archives. It generates a summary based on information that it pulls from articles in the archive. The model will also indicate whether it doesn’t have enough information to answer a reader question. Also, it lists the articles it references and tells readers to consult for verification.

The Financial Times has deployed a similar chatbot with help with Anthropic that answers subscriber questions on recent events as well as broader topics. The chatbot is in beta-testing, but when journalists from The Verge tested it out in early 2024, they found that its answers contained inconsistencies²⁸. A study by JournalismAI²⁹ found these tools struggled with summarizing longer articles, especially ones that were more creative rather than straightforward.

Outside of strictly language tasks, AI can help with reader engagement by managing dynamic paywalls³⁰ to increase subscriptions and retention. AI can also use reader behaviors, habits and journeys to recommend content and personalize user experiences. Good AI curation might nudge people to become more informed by getting them to stories they might like as well as ones they might not normally read. 

Despite promising case studies, there’s still a constellation of risks associated with generative AI tools. These include outstanding issues around accuracy, transparency, fairness, privacy and intellectual property infringement.

AI-generated stories have stirred up mass controversy for not only being poorly written, but also for plagiarism and factual inaccuracies. As more search engines deploy AI summaries, there are concerns that these features give the appearance of authority by including links to sources. But they can take facts out of context and give misinformation³¹. And depending on the source of their training data, AI models might amplify existing biases.

Although it’s easy to point fingers at chatbots like ChatGPT, the Reuters Institute ⁶ noted that news organization can’t easily circumvent this problem. Developing proprietary models in-house is challenging. Even the largest newsrooms might not have an archive large enough to supply all the training data that an LLM might need.

The best solution would be to fine-tune or prompt-tune existing models, but those methods can come with their own problems around safety, stability and interpretability.

Despite the impressive feats generative AI can perform, they ultimately lack a coherent understanding of the world³². As a result, AI cannot vet the quality of sources, and they can sometimes get tricked. For example, Wired³³ found that Google, Microsoft and Perplexity’s AI products have been surfacing AI answers based on widely debunked race science because there’s a lack of high-quality information on the web. On top of that, AI models can hallucinate, and they’re still learning how to convey uncertainty.

Previously, publications published their data and code alongside work that was produced by using machine learning or AI. Now, there’s an even higher demand for algorithmic accountability and explainability—audiences want to know when content is being produced by AI³⁴.  Even then, some early studies have shown that audiences tend to trust news content less when it's marked as AI-generated. 

Journalism relies on a relationship between the writer and the reader. Maintaining trust is paramount. And as AI becomes increasingly used across different levels of news production, media companies are trying to be as transparent as possible in their disclosures.

In a guidance put out by The New York Times³⁵ in May 2024, its editors said that generative AI will be used as a tool in service of their mission to uncover the truth and help more people understand the world. The technology is used with human guidance and review, and editors explain how the work was created and the steps they took to mitigate risk, bias and inaccuracy.

“The relationship between a journalist and AI is not unlike the process of developing sources or cultivating fixers,” as put by Columbia Journalism Review³⁶. “As with human sources, artificial intelligences may be knowledgeable, but they are not free of subjectivity in their design—they also need to be contextualized and qualified.”

There is a trend toward more transparency in AI systems across various industries. However, companies are still negotiating the tradeoffs between more open source codes and security.

The arrival of AI is complicating the terms of an evolving relationship between large tech companies and the news media industry. Over the past few decades, news organizations have been at heads with the tech platforms that distribute their content. Due to conflicting business models, there are ongoing lawsuits in which publishers have argued that tech giants are monopolizing ad revenue, traffic and by using their content without fair compensation³⁷. AI appears to be making this issue worse.

Some media executives told New York Magazine ¹⁹ that they’re worried AI summaries in search engines like Google make it easier for aggregators to “rip off their content,” and inundate the web with low-quality information.

A 2024 Tow Report³⁸ in Columbia Journalism Review noted that because custom AI is hard to develop in-house, news organizations must rely on technology companies. Further, AI-enhanced search might impact audience engagement and entrench tech platforms’ control over the information ecosystem.

However, media companies do have a leg up in the negotiations with the tech companies this time around. AI can only work with existing information. AI cannot go into the world and gather new information, experiences, or interact with other humans. News content is a valuable source of real-time information in a context that can improve the quality of foundation models, “which suffer from bias, misinformation and spam,” according to a commentary by Brookings Institute³⁹. In fact, news articles make up a substantial amount of the dataset used to train popular LLMs, according to an investigation by the Washington Post⁴⁰.

Importantly, AI systems are always hungry for more data. And tech companies are quickly running out⁴¹ of publicly available training data for these models. Without new high-quality data, these models might degrade and even collapse.

This means news outlets might have more power in defining the relationship that they’re going to have with the companies developing AI systems. And there is a higher incentive for tech companies to either work together with media companies, or face problems with intellectual property.

As it stands, a handful of publishers are suing AI companies for copyright infringement, while others are signing licensing agreements⁴².

This highlights a need for an updated legal framework around intellectual property and compensation. The Washington Post⁴³ found that sometimes AI summaries allowed tech platforms to get around previous regulations requiring them to pay publishers for content, especially content behind paywalls.

Although there are still legal landmarks to navigate through, there is a future where AI companies and news organizations can collaborate in a mutually beneficial way. The key is open communication and discussion. Instead of assuming they know⁴⁴ what people need, AI developers should ask individual industry stakeholders about the tools and services that would be beneficial to users.