Generative AI is proving to be a Rorschach test for publishers globally. Some editors stare at the inkblot and see an existential threat or a parasite that’s attached itself to their journalism. While others see incredible possibilities, with generative AI products used to automate onerous – even impossible – tasks or to create hyper-personalized content at scale.
There are certainly elements of truth in each perspective, but the reality of generative AI is neither subjective nor open to interpretation. ChatGPT, Dall-E, Bard and their kin aren’t another “bright, shiny thing” easily discounted and ignored. They’re extremely powerful tools in their infancy. Generative AI has the potential to be transformative to many industries, changing how we work in the classroom, the courtroom and, yes, even the newsroom. In a recently released Adobe workforce survey, 92% of respondents say AI is having a “positive impact on their work” and more than one-quarter (26%) call AI a “miracle.”
With the right prompts and guidelines, generative AI can excel at tasks like summarization, content optimization and transformation between content types. Because generative AI understands language syntax so well, the results are surprisingly good at turning a longform news article into a podcast script or a Twitter thread. And the promise of conversational search and prompt-based data analysis is exciting. In fact, when we at Taboola mapped about four dozen newsroom tasks around discovery, creation, optimization and distribution, we found generative AI had the potential to assist with more than 90% of them.
But it’s not engineered for every task. As is well documented, generative AI can be unreliable with facts and citations, subject to “hallucinations“. This is one of the primary reasons publishers have such different postures toward GenAI — and why it’s critical to have a methodical approach to adoption.
Establishing Rules of Engagement
First and foremost, it’s important you begin by articulating your company’s initial position on the use of GenAI text, images, audio and video. You have a social media handbook detailing what you allow and don’t. Generative AI requires the same care. What will you use generative AI for? Where will it never be used? Be specific, for both internal and external audiences. For example, maybe you will allow generative AI images – but only on tech or medical articles to illustrate complex and theoretical concepts. And remember, this won’t be carved in stone. The technology is quickly evolving, so be prepared to reevaluate your position and guidelines periodically.
Insider, for example, published a letter from the editor detailing its position on AI, taking an approach leaning into experimentation and innovation. “I’ve spent many hours working with ChatGPT, and I can already tell having access to it is going to make me a better global editor-in-chief for Insider,” Nicholas Carlson said in a memo. “My takeaway after a fair amount of experimentation with ChatGPT is that generative AI can make all of you better editors, reporters, and producers, too.”
The Guardian published a similar memo, but took a more cautious stance. “When we use genAI, we will focus on situations where it can improve the quality of our work, for example by helping journalists interrogate large data sets, assisting colleagues through corrections or suggestions, creating ideas for marketing campaigns, or reducing the bureaucracy of time-consuming business processes.” The New York Times was even more conservative, with Executive Editor Joe Kahn saying at the recent World News Media Congress in Taipei that the paper is “eager experimenters but extremely cautious in what we would be willing to present as a finished product to our readers.” In all cases, both employees and readers know what to expect.
One of the most common provisions for newsroom use of generative AI is external transparency. If an article or video is created or augmented by AI, it’s labeled as such. The Associated Press for years has relied on data-to-text AI to create short templated articles based on structured data like earnings reports. In all cases, they contain a tagline detailing its provenance and original source: “This story was generated by Automated Insights (http://automatedinsights.com/ap) using data from Zacks Investment Research. Access a Zacks stock report on ATRC at https://www.zacks.com/ap/ATRC.” U.S. broadcast network Tegna, in an email newsletter, includes a disclaimer on generative AI assistance: “The subject of this email and all text below were written by GPT.” Messages like these can help set reader expectations and maintain trust, while attempting to provide a layer of insulation against errors. A few publishers like the Washington Post have taken AI transparency so far that they are labeling non-algorithmically recommended areas as “hand curated” or “picked by editors.”
The News/Media Alliance guidance for AI also calls for full disclosure: “Generative outputs should include clear and prominent attributions in a way that identifies to users the original sources of the output and encourages users to easily and directly navigate to those products, as well as to let them know when content is generated by GAI.”
Another common guardrail publishers are attaching to the nascent technology is that human review is generally required for any output seen by the public. In a business based on trust and accuracy, tools and workflows should include the opportunity for a validation step to prevent the dissemination of misinformation. Over time, as the technology improves, more end-to-end automation will absolutely be possible. But today, full automation would require some tolerance for risk.
Setting a Development Roadmap
Once you’ve established your framework for engagement with generative AI, the next step is to experiment with a purpose. Pick a project or two aimed directly at unlocking value or solving a problem in the newsroom. Make sure it has a reasonable development timeline – no XXL T-shirt-sized projects for your first endeavor. You want to get your hands dirty here and understand the boundaries.
Next, set very clear and measurable goals. If you are trying to optimize headlines, what’s the definition of success? Will you track adoption and usage metrics, like an increase in the number of tests? Or maybe a performance metric like CTR. And where does the product live? If it’s yet another browser tab your audience team has to have open, it might face an adoption barrier. And training a newsroom in a new product carries a cost as well beyond development hours or story points. Also make sure whatever product team or innovation squad developing generative AI tools isn’t isolated. Bottom-up suggestions on workflow efficiencies come with the benefit that the people doing the work are stakeholders in that new tool’s success.
One early adopter to AI initiatives is Forbes. The publisher’s content management system, Bertie, can recommend article topics to contributors based on their earlier offerings. A politics writer would get suggestions on politics while a tech writer might be spoon-fed ideas from Silicon Valley. It also can suggest headlines and images to reduce production time.
Lastly, you aren’t in this alone. There’s a lot of good work globally being done to help education and support publishers large and small. The Center for Cooperative Media published an excellent guide on prompt usage and use cases for generative AI specifically aimed at publishers. The London School of Economics and Political Science’s JournalismAI initiative offers everything from classes to case studies.
What can generative AI do for me?
When we at Taboola approached this question, we started by roughly mapping the end-to-end production workflow in a newsroom. While all newsrooms have different roles, generally the process is the same – you get an idea for something you want to cover, you gather info, you create it, publish, and then you work hard to get an audience to that piece of journalism. Following this progression, we identified many tasks and areas where AI shows promise.
1. Discovery tasks
How can generative AI help a reporter find what article to write today? There are actually many ways, surprisingly.
Gen AI can help with data analysis, source identification, analyzing trending topics and brainstorming/suggesting ideas, like with Forbes. I also can imagine a system where it could even help route news tips/messages. But these tools have obvious limits. Often the most powerful articles, the ones with impact, require human intelligence and a journalist’s intuition. Generative AI won’t directly help with source development, building that one-on-one rapport with individuals, even though it might help you find them.
2. Creation tasks
Here’s where generative AI and conditional AI have great potential but also very clear limits. AI can help with creating articles from structured data, transcribing notes, researching topics, editing to style, packaging content, transforming content into alternative story forms and smartly sharing content across networks. With oversight and fact checking, I’ve also seen generative AI help with research and create compelling data visualizations. But generative AI can’t take photos or video from a protest scene or conduct interviews with demonstrators. Again, there’s no replacing boots on the ground.
3. Optimization/delivery tasks
We found this is perhaps the richest area for workflow efficiency. Generative AI can help with translating content, SEO/SEM, optimizing headlines, social shares, browser and native app alerts, email alerts and newsletters, content curation and real time data analysis.
If you take a solution-based approach, you found many areas where generative AI can help a newsroom, suggesting related links or headline options, hopefully freeing up reporters to do impactful work and not commoditized tasks.
What happens next?
Really, even if your newsroom does nothing with generative AI, the world is changing around you.
Google’s Search Generative Experience (SGE) is very much a work in progress, but even with recent changes highlighting more sources, it will likely prove to reduce referral traffic. How impactful is that? In a new Taboola survey of 3,700 sites live from 2019, Google in the first half of 2023 accounted for 34% of all referral sessions. For many individual publishers, Google can account for half or more of all referrals. Any search erosion means a significant audience hit.
Specifically, we believe utility searches – ones related to getting an answer to a specific question – likely will drop first and fastest. Questions like “when is the World Cup?”, “what’s my horoscope?” or “Sydney travel destinations” will be answered without a click. Searches related to brands, authors and breaking news events should be more durable. E-commerce and affiliate links tied to search performance may erode, as questions like “what’s the best blender?” could be aggregated and shown without needing to visit a site. However, Google Bard currently excludes medical, legal and financial advice, so those verticals may represent a longer-term opportunity. You also probably won’t be able offset search loss by using generative AI to create content at scale. By definition, Google says AI-created content is spam and would result in demotion or removal.
In addition to search challenges, subscription-focused publishers also are likely to see headwinds. First, search traffic reliably converts, far ahead of social, for example. So any decrease in search will have impact. Second, SGE/GenAI interfaces isolate information from its source, reducing brand trust and affinity. The world may be less aware where that information came from and, long term, but less aware of how valuable you are as a publisher.
What does Taboola recommend?
First, experiment with using generative AI, whether you develop it yourself or use any of the thousands of tools available. This is transformational technology that isn’t going away. Use it to make the jobs of journalists easier, freeing them up from mundane tasks. Also, have some fun with it. See what it can do.
Second, some search traffic loss is inevitable, probably not in 2023 but soon. Prepare yourself by performing a risk analysis of what would happen if SGE-vulnerable search went away. Because it might.
Lastly, become a destination site. Be a source of credibility in a world of chaotic content. Publishers that are able to maintain and grow direct relationships with readers and viewers will see greater success. Use and build tools that zero the distance between you and your audience. And that’s advice that’s as good today as it will be a decade from now.