It all starts with some words. Today’s image generators create a visual based on just a few keywords provided by the user. For example, the four images below were generated by Midjourney from the phrase: “AI helping journalists in producing content”.

*Image generated by Midjourney from the phrase “AI helping journalists in producing content”. Credits to* *Midjourney*

It’s easy to see how this can revolutionise a newsroom’s work. Rather than hopelessly sifting through endless stock photos, journalists can just type a few words to get a custom made image. But what are the ramifications, how does it impact budgets, copyright law and bias avoidance? The Fix looks at the implications of AI image generation and tools available.

Cheaper (and better) than stock photography

Anyone who has worked in a mid-sized media knows just how dependent you are on stock photography. Yes, you might have an in-house photographer. Some pieces might have a so-called “author’s photo” provided. But in most cases you will still end up using something from a stock gallery like Getty Images and iStockPhoto.

In this context, budget matters. Image-stock portals are generally not free to use, and many news outlets tend to reproduce the same purchased image for several similarly-themed news stories. But if the set of available images is limited, this means the potential to accurately illustrate the content of a story is limited as well.

As a result, a growing number of publishers are looking at AI generated imagery as a competitive alternative. One of those is independent Russian media outlet Holod.

“At Holod, we constantly follow the news on high-tech media solutions. We started using AI imagery in August 2022, and we believe that it is a great addition to the traditional illustrations, such as author’s photos and stock images,” Taisia Bekbulatova, Editor in Chief of Holod, told The Fix.

*Image generated by Stable Diffusion used by Holod to illustrate a piece about Russians’ relationship to authorities*

“AI imagery allows publishers to save both time and financial resources, which is especially important for small and medium-sized newsrooms. While working on longform stories, we usually hire freelance illustrators; however, when it comes to texts that lose their relevance very fast, it is important to find an attractive image in the shortest possible time – and that is where AI imagery may be a perfect solution,” Bekbulatova noted.

Stylistic diversity and finding the right fit

AI powered image generation is not just about cheap visuals. Rather, it is increasingly seen as a chance to have a wider portfolio of options. It is also a chance to play with new styles and forms of visual storytelling.

“AI technologies allow a fundamentally new approach to illustrating certain types of texts – for example, columns. As a rule, columns are accompanied by a photo of their author. However, once we had a situation where the only picture that the author could provide was a shot of him standing at attention against the wall. It did not particularly match the idea of the text (devoted to the potential ban on issuing Schengen visas to Russian citizens), nor would it attract the reader’s attention,” Bekbulatova explained.

“Eventually we decided to use an AI tool to generate an appropriate image – and the result exceeded our expectations, as the illustration perfectly corresponded to the grim tone of the column. Subsequently, we used this solution a few more times while working on texts of this genre.”

BBC News Labs is prototyping a new AI-driven tool that uses AI-powered content automation to generate new story forms that appeal to different audiences who prefer graphical storytelling to classic news reporting.

“Anyone who has worked in a newsroom knows that finding good images is time consuming. So, we use Extractive Summarisation here”, said journalist Joe Whitwell to the World Association of News Publishers in an interview for WAN-IFRA. It works like this: AI scans the text of an article, finds relevant keywords and matches them with graphical elements available in their own repository. However, this process requires the involvement of a human-in-the-loop for the final decision on publication.

*Image generated by Stable Diffusion used by Holod to illustrate a piece about EU countries visa ban debate vis-a-vis Russians*

Traditional stock image companies have not been oblivious to the changing environment. Getty Images announced back in 2018 that it was working on an AI tool that would help editors by scanning the text of a news article and suggesting options for the visuals.

Considering a computer can’t know better than humans, the company’s business development manager Benjamin Beaven, assured “Panels” by Getty Images would be leaving the editor the final choice of the image to tell the story. However, the prospects of the project seem unclear at present.

Key tools and platforms

There are multiple tools one can use for image generation. Here are the most popular and advanced ones:

Midjourney

MidJourney creates images from user prompts. They describe themselves as a “research lab” exploring “new mediums of thought”. You cannot try it right away, rather the team accepts sign-ups via its website and lets users try the beta via Discord, with a limit of 25 images per user. After that, you need to buy a subscription. Midjourney was used to create the front cover for The Economist, a comic strip for Italy’s Corriere della Sera, and even newsletter images in The Atlantic.

Dall-E-2

OpenAI’s Dall-E-2 (an updated version from the original one) is arguably the best known AI image generator, with lots of viral posts on social media. It creates realistic(ish) images from text prompts and is now available in beta to those who join their waitlist. It has generated its first-ever image used as a magazine cover by Cosmopolitan.

Stable Diffusion

Stable Diffusion, another text-to-image model, has been developed by the London-based company Stability.ai (although originally based on a neural network developed by researchers at the Ludwig Maximilian University of Munich). Unlike Midjourney and Dall-E, which can only be run via cloud services, Stable Diffusion’s code and model was publicly released and can be run by even relatively modest graphics processing hardware.

Mini Dall-E (or Craiyon)

Want to try AI image generators without having to wait or having a limit of images you can try your keywords on? Head over to the Mini Dall-E website, which is currently free, open source, and specifically made for community use.

A risk for human workers?

As always with AI-related topics, some have raised concerns about replacing human workers entirely. But Sarah Grillo, Axios Visuals Managing Editor of Illustration, told The Fix she doesn’t see AI entirely replacing journalists too soon.

“The work that my team and I do is only about 50% executing an image – the rest is working with reporters to understand stories, conceptualising approaches, and brainstorming visual metaphors to help bring stories to life. There’s also a rigorous process of feedback and critique before any visual goes out the door. How is it reading? Could there be anything problematic or offensive about a visual on a sensitive subject? How are we establishing a sense of visual hierarchy with the elements?”, Grillo commented.

Even though one could theoretically use some of the available AI tech to help deploy ideas, Grillo is convinced it’s still “crucial to have a creative professional as a key part of the process of image creation, especially in an editorial context”.

Likewise, Will Chase, Axios’ Associate Editor of Data Visualisation, pointed out that a good metaphor for these tools is self-driving cars:

“Five years ago we were predicting that they’d replace truck drivers, but we’re finding that they’re not able to match the level of improvisation and judgement of a human”

Chase argues the same will hold for media graphics, noting that “in the cases where these AI tools are replacing illustrators and designers, it’s really only in the lowest-common denominator stuff: publications that use minimally-supervised AI to generate the text of news articles that are generated purely for SEO to get ad revenue from clicks”.

Nonetheless, he concedes that tight budgets may lead some newsrooms, especially hard hit local ones, to look to AI to cut costs. “Budgets are being increasingly squeezed and they may not have the money to hire illustrators and designers”, Chase summed up.

Problems with copyright, bias

Given that AI typically creates new images from user prompts, you might wonder how this links up to the idea of intellectual property rights and copyright. Turns out, it can be quite clear: some recent DALL-E outputs featured a distorted Getty Images watermark, suggesting it used web-scraped stock images, without payment or explicit credit.

US copyright law, which plays a leading role, allows for copyrighted images to be used for transformative purposes – typically for parody or caricature. When AI software uses existing pools of data and copyrighted works, it often mines elements of original artwork without a clear reference to the original work, hence defying the transformational element.

Another problem concerns actual faces of people used, unbeknownst to them, to train the algorithm to output very realistic images of non-existing people. Images might be based on an initial dataset explicitly not allowing such use, or single individuals not having explicitly given the consent to reuse their pictures.

As a result of these problems, on September 22, Getty announced they were joining PurplePort, Newgrounds and Shutterstock in banning AI-generated art and photography on their platform, and stated the company going to be using the Coalition for Content Provenance and Authenticity, an industry-created development project, to filter out AI-generated content.

In addition to copyright issues, you can run into the problem of bias. In this illustration made by Clement Delangue, co-founder and CEO of Huggingface, the concept of “Bias Laundering” theorised by Dr. Margaret Mitchell is schematised.

*The concept of Bias Laundering illustrated. Credits to* *Clement Delangue*

It refers to when our biases exist in the datasets used to train AI models, which then become public into outlets. That, in turn, is ranked and filtered according to other biases and ends up feeding back into biases in datasets. All this amplifies bias, so it becomes harder to identify and acknowledge, and eventually fix.

However powerful it may be, and even accurate to some extent, AI in this context needs to be carefully evaluated before being used for any public purpose.

AI image generators

Latest in AI image generators

What AI image generators can do for newsrooms