Will ESPN's Reporter Robots Be Better Than the Last Generation?
ESPN will use AI to write articles about "under-covered" sports. What that means for journalism, and not just the sports kind.
The most obvious use case for generative AI in media is to use it to help write articles. It's also probably the diceiest use case, with many failed experiments over the past two years that have relegated most use of AI in newsrooms to the two ends of the story process — the story idea/research phase and the distribution/promotion phase — and generally kept it away from crafting publishable copy.
That may be changing now that ESPN has announced it has begun using AI to write stories about certain minor sports like the Premier Lacrosse League and National Women’s Soccer League. The whole idea is that these sports are "under-covered," according to the Disney-owned sports network, and using an AI bot to write game recaps for the ESPN website will fill in some coverage gaps without hiring costly humans to do it.
ESPN says human editors will review every article created this way, and it will also identify the nature of the content via the byline "ESPN Generative AI Services" as well as a disclaimer at the end of the story so readers will know when a robot is writing their game recaps. Here's what that looks like.
If this sounds like a process with the right guardrails, keep in mind it's pretty much exactly what CNET did when the tech publication first waded into the waters of AI-written articles: Human editors reviewed each one of them, but it turns out that wasn't enough: the people CNET had assigned to the task didn't always have the subject matter expertise to catch certain errors. Have the people at ESPN not been paying attention?
Hmm, Upgrades
Here's a thought: maybe they have. There's a chance ESPN has learned from the cautionary tales of using AI as a robot writer, and this might be a new-and-improved take on content at scale.
For starters, there's been a lot of progress in AI since GPT-3.5, which was the go-to model behind many of the attempts to do cheap, fast content. Smaller models like GPT-4o mini and Claude Haiku run circles around the previous generation of large language models in terms of performance and cost, so what comes out of the box should theoretically be in a better state.
It also looks like ESPN has built a better box. A spokesperson for the network told me it's developed specific training, prompts and "requirements" for this use case, and the project is a collaboration between its editorial team and the ESPN Edge team, the network's group tasked with building new content products, which counts Disney and Accenture as partners. Even if that's partially smoke, it speaks to a level of customization that likely takes into account the specifics of sports coverage, and perhaps even particular sports.
On the human element, ESPN says its copy desk will be responsible for review and quality control, and that the editors are "highly experienced." Notably, team members must manually transfer the copy to the CMS, the network said, before it can be published, minimizing the chance of errant hallucinations slipping through.
Judging from the output so far, it appears ESPN has figured out how to run this AI offense, and that the synthetic articles are competent and hallucination-free so far. That said, its AI-driven coverage may ultimately suffer from what I'm going to start calling the "Channel 1 problem." That is, even if the information is accurate, do you really want to get it from a robot?
When all you need is the information, this tends to be less of a concern, and AI bylines, voices, and even avatars have all been proven to be useful ways to connect with audiences when the human alternative is simply untenable. But for sports fans in particular, you tend to want more than just the raw info. Stats and scores you can get at a glance. If I'm going to read an article, I'd much rather know that the commentator giving a recap isn't just summarizing a set of facts.
With sports coverage, AI's lack of experience of the physical world can be a major deficiency. Games have flows and vibes that are nearly impossible to quantify. In ice hockey, for instance, a game with lax refereeing can be especially "chippy" — i.e. players sniping at each other with their sticks, creating tenser, more aggressive play — but that wouldn't show up on the stat sheet.
From Robot to Cyborg
Still, ESPN's robot writers are meant to be gap-fillers, not the next Steve Smith. But filling gaps will only go so far; even if the network assigned robots to every esoteric sport (jai alai, anyone?), that's not going to move the needle for coverage of the major leagues — the things that really drive its business.
What might elevate this use case to prime time is enhancing the human element. By leveraging an AI-written game recap as simply raw material for a human writer, the AI becomes an accelerant, not the end product. This would be more akin to how I understand Semafor Signals to work — with the AI summarizing news far faster than a human could, but the human remains the primary author, sewing together original writing with the best of what the AI writes.
Such an approach would require new standards and training, including rapid fact-checking of the AI. But if the writer is an expert in the sport, that should be relatively fast.
You can see the logical endpoint of this, beyond ESPN: a world where, when breaking news happens, an AI bot instantly cranks out raw material based on the available facts. Trained reporters quickly synthesize, fact-check, and add their own material. The human works hand-in-cybernetic-glove with the machine to publish thorough and accurate reports faster.
It's not exactly any journalist's idea of utopia. But regardless of what happens with ESPN's project, it feels inevitable that more media companies will turn to AI systems to write articles, or at least contribute to them. Anyone going down that road must heed the mistakes of the past or be doomed to repeat them.