How generative AI is playing out in the media industry
Many writers, illustrators and musicians see generative artificial intelligence (GenAI) as a threat, something that exploits their creative work to produce algorithmic knock-offs, undermining their ability to make a living.
Industry bodies, companies and trade unions are campaigning against UK government plans to let AI firms use copyrighted material without permission or payment unless creators have opted out.
The Make it Fair campaign has involved many newspapers devoting their front pages to its slogan, and various musicians including Kate Bush and Max Richter released a silent album in protest.
But creators can also use GenAI to extend their abilities. Methods vary, but some organisations and individuals are finding ways to harness such systems to do new things in ways that support and develop their businesses.
Many media organisations are making some use of GenAI, although this means coping with the technology’s frequent errors. Research by the BBC, based on getting journalists to check AI-generated answers to questions on the broadcaster’s own stories produced by four services, found that half had significant issues and a fifth had factual errors.
However, some publishers are using GenAI to produce material for publication. Reach, which publishes national and local newspapers and associated websites, has developed an internal AI tool called Gutenbot that lightly rewrites existing stories, agency copy and in some cases press releases, with a human checking the results.
The Financial Times, which has 1.5 million paying subscribers, is taking a different approach. In May 2023, its editor, Roula Khalaf, noted that current AI models are prediction engines that can fabricate facts and references while perpetuating biases. “That is why FT journalism in the new AI age will continue to be reported and written by humans who are the best in their fields and who are dedicated to reporting on and analysing the world as it is, accurately and fairly,” she wrote.
However, Khalaf also said The Financial Times would experiment with GenAI. In January this year, it set up a computational journalism team comprising two journalists, two engineers and a data scientist. On 10 April, the newspaper published an article on the involvement of Elon Musk’s Department of Government Efficiency in job cuts at the US National Highway Traffic Safety Administration (NHTSA), which has eight active investigations into Musk’s vehicle maker, Tesla. This drew on the computational journalism teams’ analysis of 10,800 free text complaints made to and published by NHTSA.
The team used GenAI to categorise the complaints automatically, based on samples, manually checking a statistically significant number. The analysis found a spike in complaints about “phantom breaking”, where Teslas stopped unexpectedly at the start of 2022, as well as ongoing levels of complaints about driver assistance tools.
“The really big thing we have learnt about AI is that it is good and reliable at, and auditable on, categorisation tasks,” says Chris Cook, who leads the FT’s work on computational journalism. Many datasets are hard for journalists to use given their size and lack of structure, but AI could change that. “You can apply it to complaints and regulatory filings – the sky is the limit,” he adds.
Automation also means processes can run repeatedly rather than just once. “Part of this is going to be about enabling us to monitor and surveil things,” says Cook, such as using machine learning to analyse satellite imagery. “It would be a lot of work for a reporter to check this every month,” he says. “We can keep half an eye on a lot of things and be tipped off.”
Using GenAI for regular searches requires it to operate in a consistent way. The team has used a version of Bert, Google’s open source AI model released 15 months ago, rather than updating it. As well as supporting consistency, Cook says hosting its own open source software stack helps to control costs.
To succeed, AI-driven research services need to become easier for journalists to use, he adds. “Data journalism has traditionally been people like me scrambling around in Python on a computer,” says Cook. “If you want to create systems that are more robust that you run every month, reliably across a range of topics, you have to have properly engineered solutions.” As well as its in-house work, as part of an agreement with OpenAI, FT journalists have access to the enterprise version of ChatGPT. Some are using this to run regular searches for story ideas, although hallucinations are an issue with this.
The Financial Times has been trialling AI-generated summaries of articles and other financial news services go further by generating some stories automatically, typically those based on clearly defined inputs such as financial announcements. Beyond this, Matthew Garrahan, head of digital platforms, doubts readers would value journalism produced artificially. “There has to be a hefty human involvement to give it value,” he says, including through journalists talking to other people and providing context. “Data journalism is about crunching data to find a signal in the noise to find a narrative and tell readers things they don’t know. With all the enhancements of AI that we are trying to harness, I still think there is huge value in a human being speaking to another human.”
Tel Aviv-based Fiverr Go’s bid to harness AI
Questions over copyright abuse have accompanied the recent growth of AI image generation services, including whether they expose users to legal risks. In response, commercial image libraries including Shutterstock and Getty Images have launched GenAI services that draw purely on their own licensed material.
In February, Tel Aviv-based freelance platform Fiverr took this a step further by offering some of its illustrators Fiverr Go, an AI tool that let them use their own portfolios as training data. Illustrators can pay a flat rate of $25 a month to generate AI images from text prompts based on their existing work. Customers can buy downloads of the images, although many illustrators include in the price a round of revisions that they carry out personally. “You still need human creativity in the loop to be there to take you to the finish line,” says senior product manager Alon Naftali.
Fiverr provides subscribers with up to three models, and suggests they use separate ones for different styles or types of output. It recommends that illustrators use up to 20 images to train each model, having found that more make only marginal differences, and allows subscribers to retrain models when they wish. Some prefer to use these models behind the scenes to generate drafts and ideas, while others treat them as an always-available sales tool that lets potential customers try before they pay for a human-produced commission.
Illustrators set the prices, own the results and can put limits on what their AI models produce, so a Muslim illustrator can refuse to allow images relating to alcohol if this infringes on their beliefs, for example. Fiverr started the service for specific types of illustrations including children’s books and tattoo designs, but plans to expand it to other kinds of images, as well as into voiceovers, jingles, writing and research.
Given that AI image generation is not going away, Naftali says it makes sense to find ways for it to support creators. “We want to make sure that in the core situation we are not using AI to replace creators, but quite the opposite, to empower creators and make them more productive,” he says. “What would motivate people to keep creating in the future if everyone can just steal their work? That’s really dangerous for human creativity.”
Berlin-based Imagine uses GenAI as tool
The idea that GenAI can be a tool rather than a threat is shared by Shai Caleb Hirschson, chief creative officer of Berlin-based creative agency Imagine. It focuses on sonic branding, the audio equivalent of the logos, colours and typefaces used in visual branding. Sonic brands typically include a short sonic logo, such as Intel’s chime or McDonald’s whistle, as well as a longer melody. While paying a composer to write a version for a high-profile advert may make sense, organisations increasingly want numerous versions.
Imagine is developing an online tool called Harmoniq, which uses the Bronze AI engine to generate bespoke audio tracks based on clients’ core sonic branding. Customers set parameters including length, for example, allowing them to generate background music for several social media posts a day. The agency can check the output, but Hirschson says that for routine uses, this may not be necessary. “It is always on-brand,” he says. “We are enabling the brands to utilise these assets to their full potential.”
To support the human composers who compose sonic logos and melodies, Imagine pays them a fee each time their material is used to generate a new iteration. “Your work is being reworked by the AI, but benefitting you as the composer,” says Hirschson. “We see a gap in the market to make this a very good revenue source for composers and songwriters. By utilising technology in a correct and fair way for the humans, we can bring man and machine together in a profitable and unique way in the industry, therefore we will attract better talent.”
He says GenAI works best for melodies as there are numerous recognisable iterations. “Rhythm is very hard to define for the AI,” says Hirschson. “It lacks the understanding of a human. If I do this on the table” – he bangs his hands on the desk twice then claps, recalling Queen’s We will rock you – “you know what I’m talking about immediately, I don’t need to say the song name.”
Expanding or otherwise changing a rhythmic sonic logo can break such recognition.
GenAI can also be used to create voiceovers. Imagine offers voiceover artists the option of licensing their vocal patterns to a brand for a year, receiving a payment for each AI-generated message. Hirschson says AI can now generate routine speech straight from text, but adds that anything with emotion works better through speech-to-speech, where someone reads the words in the required style and the system converts this into the chosen voice. This can be used to experiment with how an advertising jingle might sound if sung by a famous singer, although the results cannot be used publicly for copyright reasons.
In future, Hirschson sees GenAI removing dividing lines between human and artificial creativity. “Creativity is how you solve a puzzle, comprising multiple pieces of things in your life that come together in a new formation,” he says. “I don’t see how AI will be any different. It’s only because it’s not breathing and doesn’t have flesh, but it will get there. I see it as a co-creator, not a threat.”
Source link
