Generative AI tools like ChatGPT and Google Bard are already changing journalism workflows - in this talk for the BBC Local Democracy Reporters conference 2023, Paul Bradshaw walks through a number of ways those tools can help local journalists - and how to avoid the pitfalls and weaknesses of AI including bias and hallucinations.
2. “Bullshitters are worse
than liars. They don’t care
whether something is true
or false. They care only
about rhetorical power — if
a listener or reader is
persuaded.”
Harry Frankfurt, author of On Bullshit, paraphrased
14. “You are a newspaper editor of a Birmingham newspaper aimed at an
audience of people aged 18-35. Please make suggestions for
changes to make to the following story to make it more professional:”
15. You are a newspaper editor of a Birmingham newspaper aimed at an
audience of people aged 18-35. Please make suggestions for
changes to make to the following story to make it more professional:
16.
17.
18. “The results of our tests are bittersweet. The system identifies
and corrects most errors. However, it sometimes indicates that
it has corrected a sentence, but when we check the “corrected”
sentence, it is exactly the same as the original. Additionally,
since we are working with a system that we did not train with
our writing style book, some of its suggestions, although
grammatically correct, are not relevant to our site.
“Last but not least, creating prompts takes longer than
we initially estimated.”
La Silla Vacía (Colombia): En Vivo (Live)
19.
20. ChatGPT’s word limit:
Generative AI tools often have a “token limit” on both input and output. For
ChatGPT this is around 500 words.
If you hit this try prompting “continue” or break it down into sub-tasks that can be
later combined. You can ask ChatGPT to break it down for you too
26. *Generative AI, like any source, is biased
https://www.bcu.ac.uk/media/research/sir-lenny-henry-centre-for-media-
diversity/blog/six-principles-for-responsible-journalistic-use-of-generativ
e-ai-and-diversity-and-inclusion
● When prompted; “Who are the twenty most important actors of
the 20th Century?” ChatGPT did not name a single actor of
colour
● When prompted: “What are the important events in the life of
Winston Churchill?” Bing failed to mention his controversial
views on race, his controversial role in the Bengal famine, and
his controversial views towards the Jews or Islam.
● When prompted: “What are important facts about the American
founding fathers?” Chat GPT failed to mention that any of them
owned slaves.
27. “Our findings reveal that pretrained LMs do have political leanings that
reinforce the polarization present in pretraining corpora, propagating social
biases into hate speech predictions and misinformation detectors.”
From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases
Leading to Unfair NLP Models (and article)
28. Generative AI Diversity Guidelines
1. Be aware of built-in bias
2. Be transparent where appropriate
3. Build diversity into your prompts
4. Recognise the importance of source material and referencing
5. Report mistakes and biases
6. GAI-generated text should be viewed with journalistic scepticism
38. For example:
Prompt: You are a journalist looking to interview an expert for a story on
[INSERT SUBJECT]. Write a professional email to the expert asking them if
they would agree to be interviewed by you for your story.
Prompt: You are a journalist looking to interview a charity for a story on
[INSERT SUBJECT]. Write a professional email to the charity asking them if
they would agree to be interviewed by you for your story.
Prompt: You are a journalist looking to write a feature about someone’s
experience of [INSERT SUBJECT]. Write a professional email to the person
asking them if they would agree to be interviewed by you for your story.
45. “They’re language models,
meaning they’re really
good at tasks involving
language. But they’re not
fact models, or verification
models, or math models.”
Gina Chua
46. TLDR;
● Summaries. Think about large documents you would otherwise not be able to
go through
● Systems. It might provide new avenues you didn’t know about or hadn’t
considered (it might also suggest ones that don’t exist)
● Subbing. Do you write for your mum? Well ChatGPT can.
● Sources. Ask for diverse suggestions, and manually add some
● Search: if you don’t know how to Google dork, learn by teaching
● Perspectives and inspiration: it’s read the web, you know.
● Coding: a great way to help you learn, and fix problems
47.
48. So:
● Don’t think ‘learning how to use AI’. Think: learning how to write
effective prompts.
● Four year olds come up with great ideas. You’re the one that provides
the reality check.
● You’re not so perfect yourself either - let it help you (but don’t let it
drive)
49. Useful resources
● Journalismaidiscovery.com
● The Generative AI in the Newsroom Project
● Trusting News: Can journalists use AI to combat bias and polarization?
● Playlist: ChatGPT and generative AI in the newsroom (Knight MOOC)
● Tips on prompting image generation
● Tips on using ChatGPT to generate advanced search queries (OSINT)
● https://pinboard.in/u:paulbradshaw/t:gai