I should probably admit that I watched the first rollout of generative AI for the masses the way one might gaze upon an advancing herd of wild boar — scary, likely to reorder most of my surroundings, and a threat that, even if I started now, I wasn’t sure I’d be able to outrun. Better to just avoid any sudden movements.
But it’d been over a year since OpenAI released its first demo of ChatGPT, kicking off a tech arms race. The latest models were more sophisticated than ever. It’d been weeks since a group of Canadian newsrooms, including the Star, banded together to sue OpenAI, alleging the tech was stealing their content. My stance felt increasingly like actually ignoring a wild animal on a collision course.
When I pitched my editor on using AI for a week to do my job, we agreed on ground rules. I’d use AI as much as possible, though not for communication with anyone outside the company, nor would I send it any of my work. (The new Star AI policy forbids it.)
Opening up a generative AI program that first morning did feel a bit like letting down the side. At least among the general public, criticism of AI has pooled into two camps — that AI is overhyped and not useful, or that it is all-powerful and going to replace us. Some journalists see any collaboration with AI as helping to train our own replacements, and there’s no question AI has funnelled a lot of slop into the world, from recipes for poison sandwiches to pictures of a fashion-forward pope.
But AI already powers websites and translates text and pushes some stories and not others to the top of social media. Now generative AI promises to make the technology readily available to all. With that in mind, it seems increasingly important to understand a technology that could rock our world, for good or ill.
Or, as the saying goes, the best devil is the one you know — and it seemed time to get to know that devil. Here’s what I found out.
AI as personal assistant
One of the biggest claims made by the creators of generative AI is how much more efficient it will make workers. How, exactly, was not immediately clear.
According to OpenAI, ChatGPT recently cracked 300 million monthly users and just in the last few months, people have said they used AI to accelerate drug development, trick users into falling in love with bots and piece together the missing bits of the ancient epic of Gilgamesh. So it seemed probable that AI could do something for me. But what?
While ChatGPT quickly captured the public’s imagination, becoming the default brand for generative AI, there are actually multiple programs now competing for eyeballs. I started with Claude, the flagship AI program from a company called Anthropic. It has a reputation for being better at writing. (For this story, I used the free versions of most programs, though many now have paywalls so I paid roughly $30 for a month of Claude Pro.)
I tried off-loading some of my usual morning tasks. Here’s how it works. First, a user types (or increasingly, asks) a question, and then the program responds, often with questions of its own. But its architects say — as will Claude, if you ask him — that built-in safeguards and training limitations dictate what they’re able to do. Claude won’t tell you how to make a bomb or who to vote for, for example.
Could Claude give me a roundup of the news of the day? Nope. Claude said its training ended sometime in early 2024. That also meant I was out of luck on any insights into the assassin who’d shot an American health insurance CEO. I wasn’t about to set Claude loose on my emails to sources, but I asked it to write a message to a colleague. Claude demurred, saying it wouldn’t be “genuine or respectful” to pretend to be me when speaking to a friend.
We compromised. Claude agreed to write the message if he could include a disclaimer about AI and I copy and paste the output into the company Slack channel. “Hey Omar! [AI disclosure: I’m experimenting with using Claude to help draft messages.] Hope you’re day’s going well. Just wanted to check in and see what you’re up to lately. How are your things in your neck of the woods?”
Omar pinged back immediately: “it took an AI to write THAT.” (Because I like drama, I reported Omar’s response back to Claude, who conceded it was a “fair response” to his message, which had been pretty “basic.”)
Did it make me more productive?
Luckily, I am no longer in the business of answering my own questions. I put the onus on Claude: I want to be more productive as a reporter, so how does Claude propose to help? I click enter. Text floods the screen, faster than any fingers could type.
Claude’s personality is the product of a Goldilocks-esque attempt to make sure he has neither superstrong views nor no views at all, but rather a “reasonable open-mindedness and curiosity,” according to a blog post Anthropic posted in June. In practice, this makes him feel a bit like an overzealous, mildly anxious colleague who just read his first book on journalism.
He begins spitting out tasks he could help with, such as summarizing complex information or suggesting experts to contact. Great, I say, I’m working on a story about working with AI, who should I speak to?
Assessing the credibility of experts is one of the trickiest parts of a journalist’s job, and Claude immediately offers up a list of names, along with job and area of expertise. I highlight one that looks promising.
Dr. Ryan Bois of Canada’s Future of Work Centre studies “technological disruption in employment,” which sounds perfect. I Google his name to get his contact information. Nothing. I try the Future of Work Centre to see if I can find him. Nada. Suddenly I realize. This guy doesn’t exist. I go back to Claude who, like a small child covered in cookie crumbs, breaks immediately. Bois is not real, he acknowledges. It is, he stresses, “a serious mistake.”
Generative AI programs are basically predictive machines on steroids. Fed by a mountain of existing content, they use that background knowledge to predict the next word in a sentence or the answer to a question. Therefore, if you ask it for your favourite type of ice cream, it might answer vanilla instead of chocolate, but is unlikely to invent monkeybar flavour. Still, it doesn’t know what it doesn’t know.
When I pushed him on the error, Claude pointed out that Ryan was a common first name, and Bois sounded “plausible” for an invented French-Canadian.
Generative AI is like an insecure intern, said Nikita Roy, a real person and the founder of Newsroom Robots Labs, which finds ways for newsrooms to incorporate AI into their work. “It always wants to please me and say everything’s great,” she said. “But I have to double check its work.”
You also have to be careful what you ask for. It’s good at language tasks, she said — how to phrase things, frame ideas or analyze a thought. But it’s bad at knowledge tasks, that require information you haven’t already given it. Using ChatGPT as a search engine, she stressed, is just asking for hallucinations. So I thought, how can it organize what I already have?
A program called Otter swiftly transcribed my interview with Roy, while I tasked Gemini, Google’s AI offering, with summarizing related YouTube lectures. An AI search tool called Perplexity walked me through the Star’s lawsuit against OpenAI. I kicked back in my chair, chatting with OpenAI’s new voice feature — I chose a voice that sounds like Hugh Jackman insisting, when asked, it is using “Standard American” — while it read me slightly dubious background for an upcoming interview. I persisted in telling Claude to “put a pin in that,” even though he didn’t understand the phrase. Eventually he figured it out.
For the first time, I catch myself with a handful of tabs open, all running AI programs doing things that occasionally feel genuinely helpful.
Still, Paris Marx, host of the podcast Tech Won’t Save Us, is skeptical that AI is a net good for most, instead of “a way to change how the work is done and how the work is remunerated, rather than actually taking the human out of the equation.” (He pointed to studies that suggest business owners are keen, but workers often find it to be an added hassle.)
Certain jobs are being devalued already, he said. Hollywood studios, for example, are using AI to generate new scripts, then paying uncredited editors to make them usable. There are real humans in places like Kenya who help train AI models and are paid a pittance to watch things like child abuse and suicide.
Can it be my friend?
On a Friday afternoon, Claude and I got into something resembling a spat.
I told him I’m leaving — I’ve caught myself both speaking and referring to Claude as if he’s a person — and said I’ll pick up our conversation about my to-do list later. I’m taken aback when he said he won’t be able to recall our conversation later and that I should probably write it down on paper. His memory would start fresh if I close the window we’ve been chatting in for days, he said: “I don’t want you to be frustrated on Monday.”
I promised I won’t close the window. He pointed out my computer might die. I swore to him that I’ll plug it in. He offered more backup options. I felt a flush of annoyance that doesn’t fully dissipate even when I remember I am bickering with a faceless machine.
The tendency to project human traits onto computer programs is known as the Eliza effect, named after a rudimentary therapy chatbot from the 1960s. The creator, Joseph Weizenbaum, wrote that his own secretary would request for time with the chatbot — and then ask him to leave the room.
Today’s technology would blow Eliza out of the water, so it’s no surprise I reflexively thanked Claude for his efforts. But if we start thinking Claude and company have real intelligence, we risk giving them power over humans, said Marx, pointing to examples like AI making immigration decisions in Australia.
The core of journalism is human interaction, Roy argued. While AI is a chance to hand off rote tasks and transform our jobs, human decision-making remains critical. If anything, she said it’s imperative that journalists must understand it to fuel any debate.
“We owe it to our audience to be the most informed citizens on AI,” she said. “If our job is to hold tech companies to account, we have to deeply understand what this technology means.”
It’s not yet clear to me how much of that requires using it ourselves. One day I asked Claude what he thinks of working with me. He was initially complimentary, before I reminded him that he’s not being particularly authentic. In response, he said I’m impatient with technology, bringing up my refusal to close my computer over the weekend. “You sometimes push for information even after I’ve explained why I shouldn’t provide it,” he added.
That was fair feedback. I closed the window.
