{"id":2516,"date":"2024-02-16T13:00:47","date_gmt":"2024-02-16T13:00:47","guid":{"rendered":"https:\/\/niemanstoryboard.org\/2024\/02\/16\/journalism-ethics-artificial-intelligence-sources-fact-checking\/"},"modified":"2024-10-18T16:24:13","modified_gmt":"2024-10-18T16:24:13","slug":"journalism-ethics-artificial-intelligence-sources-fact-checking","status":"publish","type":"post","link":"https:\/\/niemanstoryboard.org\/2024\/02\/16\/journalism-ethics-artificial-intelligence-sources-fact-checking\/","title":{"rendered":"Two journalists talk to the bots \u2014 who talk back \u2014 about the pros and pitfalls of AI"},"content":{"rendered":" \n
\n
\n \n

Nieman Storyboard contributor Chip Scanlan and Casey Frechette worked together for more than a decade at The Poynter Institute where they created online courses in reporting and writing. Casey is an associate professor and chair of the journalism department at the University of South Florida.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Chip Scanlan: <\/strong>Hi Casey, I\u2019ve had a great time with you on Google Meet over the last several weeks playing and working with AI and talking about its implications for journalists and journalism. Why does the topic interest you?<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Casey Frechette: <\/strong>Hi Chip. It\u2019s been great exploring this with you. I got interested in AI last spring, when I had what I\u2019d call a \u201cwhoa\u201d moment. I started playing with ChatGPT in earnest, feeding it different journalistic things. I gave it a bunch of news stories and asked it to suggest headlines. I pasted an article from The Atlantic and one from The New Yorker and asked it to tell me which was which. I had it summarize stories in two sentences. And I asked it to review several pieces of breaking news for signs of bias, both obvious and subtle. I was blown away by how quickly it performed these tasks, at how useful the results were. The experience felt a bit magical, akin to going online for the first time and feeling awestruck by the possibilities (minus the crunchy modem-connection noises). But I\u2019m curious: What\u2019s drawn you to this topic?<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Chip:<\/strong> I guess I\u2019ve always been an early adopter. As soon as news broke on Nov. 20, 2022, that Open AI launched ChatGPT<\/a>, I signed on, first as a free user, and then got a $20 a month subscription so I could have access to ChatGPT-4, its most advanced system. I haven\u2019t used it as broadly as you have or focused it on journalism as I would expect an inspiring J-prof like yourself would.  I\u2019m writing a historical novel set during World War II and have used it primarily for research. Its performance has been impressive, although I must concede that a Google search can answer many of my questions. What sets ChatGPT apart is the chat part, a dialogue between you and a super-fast computer, one trained by scraping large swaths of the internet for content. Follow-up questions are the lifeblood of the journalist and I ask the chatbot a lot of them. The most important one is \u201cWhat are your sources?\u201d<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

In its staff guide<\/a> to AI, Wired magazine describes AI\u2019s value to journalists: \u201cAI software can\u2019t call sources and wheedle information out of them, but it can produce half-decent transcripts of those calls, and new generative AI tools can condense hundreds of pages of those transcripts into a summary.\u201d<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

But you and I have used it together to see how it operates around the discipline of journalism. Want to take it from there?<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Casey:<\/strong> Isn\u2019t it amazing how the cost of technology plummets over time? I recall learning early digital watches sold for upwards of $1,000 when they first hit shelves; a nightlight and stopwatch were the extent of their \u201csmart\u201d features. I guess we could call it the \u201cearly adopter tax\u201d but, like you, I find the allure of exploring the frontiers of innovation too enticing to pass up.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Once we accept that AI is flawed, we can use it responsibly, even relish in what it has to offer. <\/p>\u2014Casey Frechette<\/em><\/cite><\/blockquote><\/figure>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Using a chatbot as a kind of research assistant for your novel is a great idea. The interactive style you described certainly translates well to the kind of critical engagement that makes journalists so good at finding, organizing and presenting information. Applied to this realm, the possibilities are really quite staggering.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

There\u2019s a lot of focus, understandably, on the \u201cgenerative\u201d side of AI. And we\u2019ve all heard the horror stories of news organizations that have dabbled with \u201crobot-assisted\u201d stories<\/a> that sound cartoonish or make embarrassing errors.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

What\u2019s more interesting to me, though, are all the other ways AI could change news production and consumption, from analyzing vast, complex datasets, as The New York Times recently did in a report on Israeli bombings in Gaza<\/a>, to generating different story variations for each reader, each calibrated to exacting personal preferences.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I think there are a few lessons here, chief among them: Both pitfalls and possibilities abound.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Chip: <\/strong>We\u2019ve spent a lot of time testing the boundaries. I\u2019m thinking of the time I took notes from a newspaper story of mine that I reverse-engineered to teach writing to journalists. First, I inputted the text into the chatbot with a request to answer questions about the story, including my favorite: What\u2019s the story really about, in one word? In seconds, it replied, \u201cEmpathy,\u201d which totally nailed the theme. I was impressed.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

But as you\u2019ll recall from the other end of our Google Meeting, when we asked it to start turning the notes into narrative, the problems began.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

 \u201cWriting stories is another matter,\u201d  a staff guide for Wired <\/em>magazine concluded. \u201dIt turns out current AI tools are very good at churning out convincing (if formulaic) copy riddled with falsehoods. A few publications have tried \u2014 sometimes<\/a> with disastrous results<\/a> \u201c<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I\u2019ve read that AI scientists call that \u201challucinating\u201d or \u201cconfabulation.\u201d My profane translation as a journalist: It began to make shit up. It made up dialogue. It inserted facts that weren\u2019t in the notes.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

When you and I challenged it on these points, it apologized \u201cfor the oversight.\u201d It said the dialogue was \u201ca creative interpretation based on the situation outlined in your notes, as there was no direct transcript of the conversation.\u201d<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Not ready for journalism primetime, we agreed. In a longer version of our e-conversation, you told me that the current iteration of AI tools are \u201cnot built for journalism\u201d but, with the right cautions, can still be useful. Those cautions include the warning that we can\u2019t take anything for granted, and must question the sourcing of everything, especially in high-stakes stories.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Your take, as a journalism professor and technologist, holds a lot of weight with me. So let me ask. What do you see as other pitfalls of AI when it comes to reporting and writing stories? How might it be of use to writers and editors? And what does someone have to come up with so journalists can use this technology in the most intelligent and ethical way possible? (Hint: We might call it \u201cDefoe\u201d after the early 18th-century writer whom ChatGPT identified as one of the world\u2019s first journalists (historians might disagree).<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Casey: <\/strong>You\u2019ve hit on what I see as the key concern, and that\u2019s the fallibility of the technology. It makes stuff up. It gets stuff wrong. It leaves out important caveats and context. And it does all of this with the confidence of, to paraphrase a popular saying, a person putting on lipstick without a mirror.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

In a similar vein, we all need to be concerned with the biases lurking in the datasets that AI models are trained on. There\u2019s important work ongoing in the realm of responsible AI to combat this, but reporters and editors need to be aware that many AI technologies have amplified, not diminished, the unfairness and misrepresentation that ail society and undermine the journalistic mission.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Another pitfall is an extension of the conventional wisdom that phones made reporters lazy because they didn\u2019t have to go see people in person anymore. And then email made them lazy because they didn\u2019t have to pick up the phone. And social media made them lazy because they didn\u2019t have to send email. And now, I suppose, AI might make them lazy because they don\u2019t have to hop on their DMs. I tend to think this concern, and its many permutations over the years, grossly underestimates journalists\u2019 innate resourcefulness.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I do worry, though, that the \u201csmarter\u201d technologies become, the more tempting it is to outsource our cognitive capacities to them.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I\u2019ll confess that if you asked me to drive to a location I\u2019m unfamiliar with, without the assistance of Waze<\/a>, I\u2019d be hard-pressed to get there without some struggle and a good deal of cursing. The advent of calculators made long-division a rare skill. The internet wreaked havoc on our attention spans, as Nicholas Carr explored in his seminal book \u201cThe Shallows.\u201d<\/a> What will happen with AI? Will our creative and expressive muscles atrophy?<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

But let\u2019s look at it from another angle. Every reporter must now ask themselves: Did AI write what I\u2019m reading? Draw what I\u2019m seeing? Speak what I\u2019m hearing?<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I anticipate we\u2019re entering a messy phase where parts of things are AI-generated, mixed with human-generated content. <\/p>\u2014Casey Frechette<\/em><\/cite><\/blockquote><\/figure>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

We might be comfortable answering these questions now, at ease with our ability to notice a stilted voice or a telltale sentence construction. But whatever signs we think make fake content easy to spot at the moment are sure to become outdated as the technology improves. And here\u2019s a further complication: Much of the discourse about deep fakes has assumed what I see as an all-or-nothing framing. But I anticipate we\u2019re entering a messy phase where parts of things are AI-generated, mixed with human-generated content.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Here\u2019s the good news: Once we accept that AI is flawed, we can use it responsibly, even relish in what it has to offer.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

For writers and editors, I see potential in AI as a research tool; a brainstorming partner; a transcription service; an assistant able to sort, summarize and label; a writing coach who offers feedback and troubleshoots problems with copy; an editor who can review early drafts for clarity, accuracy, AP style, grammar; a data analyzer; a reader who can give feedback as either a neophyte or expert on the topic at hand, depending on the role we want it to play.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I\u2019d hasten to add that there\u2019s a certain irony in considering how journalists can benefit from AI, considering how much AI has benefited from journalists.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

We now know that many chatbots have been trained, in no small part, on stories that have been scraped from news websites. Without the deeply considered, reported and edited content that\u2019s produced almost exclusively by news organizations, chatbots like ChatGPT wouldn\u2019t be nearly as good. At least, that\u2019s the argument lawyers for The New York Times have made in their action against OpenAI and Microsoft<\/a> alleging massive copyright infringement.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

If we\u2019re envisioning an AI tool truly suited for journalists, I\u2019d imagine it would boast a few critical features:<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

It would be private and secure. It would acknowledge the limits of its understanding and the ways it might be presenting biased or distorted perspectives. It would cite its sources. And it would never make shit up.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

What\u2019s interesting is that the current generation of tools isn\u2019t intrinsically incapable of these things. Yet, as we\u2019ve seen from our experiments, it takes tremendous coaxing to get a bot to work only with supplied facts.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

As someone who has coached legions of writers, I wonder what features you might add to this wishlist.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Chip: <\/strong>I\u2019m not sure what I can add to your comprehensive list.  I\u2019ve already been impressed with it can summarize and analyze texts with detail and accuracy in a fashion that would make a coach or editor proud. This strength would help writers, too, especially before they submit a story or work without the safety net an editor provides. It could also guide revision, along with other features on your wish list.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

The aforementioned risk remains: ChatGPT will happily make the revisions if you ask, and there lie the dangers of hallucination, not to mention the ethical lapse that could, if not should, cost a journalist their job.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Still, it can\u2019t provide the emotional support I\u2019ve found some writers need to get past writer\u2019s block and overcome fears about approaching strangers with sometimes uncomfortable questions, and meeting deadlines \u2014 problems that afflict rookies and newsroom veterans alike. It may put me out of business as a coach down the road. (I don\u2019t know how far down.)<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I\u2019m convinced it can make us better writers by raising questions and pointing out flaws. The question for the future is how far should we let it go? Highlight passages in news stories that were created with AI? Share a byline with ChatGPT? What will it do to a writer\u2019s voice, the element that draws in readers and keeps them engaged if everything is regurgitated by a large language model and copied from other writers? That\u2019s a big question and one that anyone who works with words has to address.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

But first, in the spirit of this exchange, I asked ChatGPT to answer your question about my wishlist as a coach. Its response (in italics so we\u2019re certain of the source):<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n
\n
\n
\n \n

ChatGPT:\u00a0<\/strong>Source Verification: An AI tool that can assist journalists in verifying the credibility of sources and the authenticity of information, especially in an era of deepfakes and misinformation.<\/p>\n\n <\/div>\n <\/div>\n\n\n \n

\n
\n \n

Ethical Guidance: An AI that incorporates ethical guidelines specific to journalism, helping journalists navigate complex issues like privacy, consent, and the public interest.<\/p>\n\n <\/div>\n <\/div>\n\n\n \n

\n
\n \n

Collaborative Editing: Enhanced features for collaborative editing that allow multiple journalists to work together seamlessly, with AI providing real-time suggestions for improvement based on best practices in journalism.<\/p>\n\n <\/div>\n <\/div>\n\n\n \n

\n
\n \n

Enhanced Fact-Checking: An AI tool that not only helps in fact-checking but also provides context and background information to help journalists understand the broader implications of their stories.<\/p>\n\n <\/div>\n <\/div>\n<\/blockquote>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Between your and my ideas, and ChatGPT\u2019s, we could definitely give a coder some ideas. And now that we\u2019re back to human-to-human conversation, can AI replace what you do as a teacher?<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Casey: <\/strong>Replace me? No way!<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Wait, let me step back. My knee-jerk reaction to your question is probably driven more by fear of obsolescence than a reasoned view on where we\u2019re heading. Much as I\u2019d like to believe I can offer something of unique value, my sense is that AI can already outperform me on several key tasks and will soon be a worthy rival on many more, from designing learning experiences, to producing lessons, to providing feedback.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

I\u2019m not quite ready to let go of the notion that, without emotion, a computer will never be able to provide the kind of connection that instills a passion to learn. <\/p>\u2014Casey Frechette<\/em><\/cite><\/blockquote><\/figure>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

In her book \u201cMachines Who Think,\u201d Pamela McCorduck<\/a> recounts philosopher and mathematician Bertrand Russell<\/a>\u2019s reaction when the inventors of \u201cThe Logic Theorist,\u201d<\/a> often regarded as the first AI program, presented him with an AI-generated mathematical proof superior to one of his own. Russell\u2019s reaction, McCorduck wrote, was one of delight. I hope I can approach AI\u2019s expanding capacities with a similar open-mindedness and humility.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n

Still, I\u2019m not quite ready to let go of the notion that, without emotion, a computer will never be able to provide the kind of connection that instills a passion to learn. And yet, I can\u2019t help but think that this is the next chapter in our relationship with technology. I hope you don\u2019t mind if we bring our friend Claude <\/a>into the conversation.<\/p>\n\n <\/div>\n <\/div>\n \n

\n
\n \n\n\n <\/div>\n <\/div>\n \n
\n
\n \n
\n
\n
\n \n

Claude, what are your thoughts on AI\u2019s potential to provide humans with emotional connection?<\/p>\n\n <\/div>\n <\/div>\n\n\n \n

\n
\n \n

Claude:<\/strong>\u00a0I don\u2019t actually have personal thoughts or feelings. I\u2019m an AI assistant created by Anthropic to be helpful, harmless, and honest.<\/p>\n\n <\/div>\n <\/div>\n\n\n \n

\n
\n \n

Casey:<\/strong> Allow me to rephrase the question. Could AI one day provide humans with meaningful emotional connection?<\/p>\n\n <\/div>\n <\/div>\n\n\n \n

\n
\n \n

Claude:<\/strong>\u00a0I think the potential for AI to provide emotional connection to humans is an interesting area of research and development. Some key considerations around this topic:<\/p>\n\n <\/div>\n <\/div>\n\n\n

\n
\n \n