Sport produces an enormous amount of structured data: scores, scorers, minutes, cards, standings. That makes it one of the first places where newsrooms tried to let software write. Large language models (LLMs) such as the ones behind today's chatbots have made automated writing far more fluent, but fluency is not the same as reliability. Here is how automated sports journalism developed, a cautionary tale, and what the technology is genuinely good for.
Before LLMs: templates fed by data
The first generation of "robot journalism" did not understand language at all. It filled templates with numbers from a data feed: "[Team A] beat [Team B] [score] on [day], with [player] scoring twice." Rules chose between phrasings depending on the data, for example "edged" for a one-goal win and "routed" for a big one.
The Associated Press was an early large-scale user. From July 2014 it used software from Automated Insights to write stories on US company earnings. In 2016 it announced it would use the same approach to cover minor league baseball, producing automated game stories for more than 10,000 games a year that it had never had the staff to cover, using data from MLB Advanced Media.
The appeal is obvious: coverage of games that would otherwise get none, produced in seconds.
A cautionary tale: the [[WINNING_TEAM_MASCOT]]
In August 2023, several newspapers owned by Gannett, including the Columbus Dispatch, published high school sports reports written by an automated service called LedeAI. Readers quickly noticed problems. One story began:
"The Worthington Christian [[WINNING_TEAM_MASCOT]] defeated the Westerville North [[LOSING_TEAM_MASCOT]] 2-1 in an Ohio boys soccer game on Saturday."
The template's placeholders had been published as they were. Other reports were repetitive, missed the details that matter to local readers and used odd phrases. After the stories were mocked on social media, Gannett paused the experiment in all the local markets that had been using it.
Key point: The failure was not mainly technical. A template with a missing field is easy to catch, if someone reads the story before it goes out. The real mistake was publishing without a human check.
What LLMs change
Template systems can only say what their authors anticipated. A language model can write a fresh report from a box score, vary its style, summarise a press conference or translate a story in seconds. It can also make things up. LLMs generate the most plausible next words; they do not check them against reality. In sport that can mean:
- a goal credited to the wrong player;
- an invented quote, injury or transfer rumour;
- a confident "record" that is not a record;
- old information presented as current, because the model learned from data that has since changed.
Watch out: A fabricated quote or injury in a sports story is not a small error. It can affect a player's reputation, a club's relationship with the media, and readers who make decisions based on it.
Where language models genuinely help
Used carefully, LLMs are useful tools in a sports newsroom:
| Task | Why it works | What still needs a human |
|---|---|---|
| Turning data into first drafts of routine reports | The facts come from a verified feed | Checking names, numbers and tone |
| Summarising long documents, such as disciplinary rulings | Saves reading time | Confirming the summary against the original |
| Transcribing and translating interviews | Fast and cheap | Checking quotes word for word |
| Suggesting headlines and social posts | Many options quickly | Choosing and checking them |
| Searching an archive | Finds related stories | Verifying what it finds |
The common pattern: the model handles the volume, and a person remains responsible for every fact that is published.
Questions readers can ask
As more sports content is produced with AI help, readers can look for a few signs of a careful publisher:
- Does the site say how it uses AI?
- Are sources and data providers named?
- Are corrections published and dated?
- Do articles add something a data feed alone could not, such as context, analysis or reporting?
For a look at the other side of AI in sport, the models that try to forecast results, read How AI Predicts Football Matches.
- Automated sports writing started with templates fed by data; AP used it for earnings reports from 2014 and minor league baseball from 2016.
- Gannett paused an automated high school sports service in 2023 after flawed stories, including unfilled placeholders, were published.
- LLMs write fluently but can invent facts, so every published fact still needs human checking.
Further reading
- The Associated Press will use automated writing to cover the minor leagues — Poynter, 30 June 2016
- Gannett to pause AI experiment after botched high school sports articles — CNN, 30 August 2023
- Automated journalism — Wikipedia
The AI Sports team writes about data, models and technology in sport.