The more I look at the latest machine learning models, the more convinced I am that these technologies are bound to make their way into newsrooms — sooner rather than later. I could be a bit biased, since I’m currently developing a transcription app for journalists called Stage Whisper, which uses the Whisper machine learning model to automatically transcribe interviews. But I believe the latest developments in generative artificial intelligence could revolutionize the way journalists do their jobs.
Powerful large language models like GPT-3 promise to give reporters the ability to partially automate the task of writing articles, while generative image models like DALL-E 2 and Stable Diffusion could deliver custom stock photos to newsrooms that cannot afford subscriptions to photo services. And of course, the latest automatic speech recognition models will allow reporters to easily and accurately transcribe their interviews for free.
Many journalists already upload their interviews to paid auto-transcription services like Otter and Trint, which rely on machine learning models to quickly produce fairly accurate transcriptions. The Whisper model, released by artificial intelligence company OpenAI in September, has some real advantages over the current services. It’s more accurate, doesn’t require an internet connection, and is completely free. But it’s also relatively slow, cannot distinguish between speakers, and requires a lot of technical know-how to set up. (Stage Whisper should at least solve that last issue.)
Even more promising are generative image models like OpenAI’s DALL-E 2 and Stability AI’s Stable Diffusion. These tools have allowed users to generate realistic images based solely on text prompts. Some technology-focused reporters have already begun using AI-generated images to illustrate their stories, more as a gimmick than anything else. But as these technologies continue to improve and enter the mainstream, they are likely to reshape newsroom debates over photo manipulation, photo illustrations, and misleading stock photos.
What about the dream (or nightmare) of an artificial intelligence that can write news articles? Wire services like the Associated Press and Reuters have previously experimented with automatic article-writing tools that can produce schematic news stories based on numerical data like earnings reports or sports scores. But the latest large language models can go well beyond that.
Large language models, such as GPT-3, have the potential to disrupt newsrooms by enabling journalists to use AI to write their stories. With these models, journalists can input a few key points and the AI can generate a complete article, saving time and potentially improving the quality of the writing. However, this raises ethical questions around authorship and plagiarism. If a journalist relies too heavily on AI to write their stories, who can be credited as the author? Additionally, there is the potential for the AI to inadvertently plagiarize other sources, raising questions about journalistic integrity and accuracy.
Another potential risk of relying on large language models to write news articles is the potential for the AI to insert fake quotes. Since the AI is not bound by the same ethical standards as a human journalist, it may include quotes from sources that do not actually exist, or even attribute fake quotes to real people. This could lead to false or misleading reporting, which could damage the credibility of the news organization. It will be important for journalists and newsrooms to carefully fact check any articles written with the help of AI, to ensure the accuracy and integrity of their reporting. [/END]
The previous two paragraphs were produced by ChatGPT, a OpenAI tool that allows users to interact with GPT-3, a large language model that responds to user’s prompts. ChatGPT can already produce fabricated news stories that sound remarkably real, complete with fake quotes. Given the proper inputs (accurate information, real quotes from sources), there’s no reason that the model and its successors couldn’t produce true and accurate news stories too.
AI-based tools will never replace human journalists. But much like previous technological advancements, these tools could change how reporters do their jobs — freeing them up to spend more time interviewing sources and digging up information and less time transcribing interviews and writing daily stories on deadline.
As these tools enter newsrooms, they will spark new questions about journalistic ethics: Is it wrong to train a generative AI model on thousands of artists’ images without their consent? Is it misleading to publish an image of something that does not actually exist or an event that never actually occurred? If a reporter uses a large language model to write an article, should that be considered plagiarism or even fabulism?
Let 2023 be the year we start figuring out some of the answers.
Peter Sterne is an independent journalist in New York.