It’s now a known thing that tools like ChatGPT “hallucinate” (ie make things up). But how do you fact-check the stuff they say?
In journalism circles, there’s a meme of sorts which goes like this: if one person says it is raining and one person says it is not, it is not the reporter’s job to simply quote them both. Instead, the thing to do is look out the window and check the weather conditions.
The originator of this piece of wisdom (as far as I can tell) is 1940s Texas journalist Hubert Mewhinney. In a 1974 book, journalists Sam Kinch and Stuart Long quote Mewhinney as saying this about two powerful politicians of the era:
“If Jimmy Allred says it’s raining, and W. Lee O’Daniel says it isn’t raining, Texas newspapermen quote them both, and don’t look out the window to see which is lying, and to tell the readers what the truth is at the moment.”
There’s a more recent, and more colourful, version of this basic fact-checking tool: “Look out the f*cking window”. A Reddit poster who was in his class says journalism professor Jonathan Foster used this as a staple teaching tool for decades.
Did I look out the window for this story?
I did indeed. I used three AI tools (Le Chat, Gemini and Perplexity) to find the origin of the story. All three reference a site called Quote Investigator, which is the work of a former computer scientist and author Gregory F. Sullivan, working under the pseudonym Garson O’Toole.
The Quote Investigator article on the subject of whether it is raining or not listed seven different sources, with links. Perplexity gave me the link to the Reddit thread, and the Foster quote is also found on the Goodreads site.
All that cross-checking is something that happens with no conscious plan on my part. It is just something I always do, because I spent many decades in journalism, where fact-checking is (or used to be) as basic a function as breathing.
It has occurred to me, often, that urging people to check the truth of things they see in possibly (or probably) inaccurate generative AI outputs is not particularly helpful if you don’t have that basic journalistic turn of mind.
A fact-checking roadmap (made with the help of AI)
Over the years, I’ve written a bunch of blog posts that relate to fact-checking (there’s full list of links at the bottom of the post). I took all those posts and asked NotebookLM to turn them into a practical list that people can use to keep their wits about them in our AI-slop-filled world.
It did a decent job, which I did not have to fact-check since NotebookLM only looks at the sources you give it. And I wrote all but one of the sources.
Here’s that list, edited for brevity – and there’s a downloadable checklist here.
Step One – Mindset
To fact-check generative AI, you need to start with what I call the “narrowing of the eyes”. Your default position should be to activate an inner voice that asks “Really?” That’s because generative AI is a prediction engine designed to produce smooth, plausible-sounding patterns of words; it has no built-in understanding of truth or reality. It isn’t “hallucinating” when it gets things wrong – it is bullshitting, which is the act of persuading you without any regard for the truth. Also: be clear about what you are looking at. If it’s an opinion, you don’t necessarily have to do all the steps below. If the opinion is derived from things that are stated as facts, then go to step two.
Step Two – Cross-check against other sources
Never rely on a single AI output for important information. If it hasn’t give you links to sources, ask for them. Do an internet search to see if the information is being reported by multiple authoritative websites or reputable news outlets. Consistency across multiple reliable sources increases the likelihood of accuracy. (And don’t just take the word of an AI-generated overview that you might see in a search engine: go straight to the sources listed for that overview. That’s because they too are often wrong)
Step Three – Investigate the source’s bones
If the AI points you to a specific website, scroll to the bottom. Does it have an “About Us” or “Contact Us” section with a physical address? Is there a clear privacy policy or a place to report mistakes? If not, you may be in dodgy territory.
Step Four – Check for time and human accountability
Look for a date and a byline (the author’s name) at the top of the text you have landed on. Is there a link to a profile for the author? Can you find them on another platform (like LinkedIn or Wikipedia)? Reliable sources are usually staffed by real people who are willing to be held accountable for their mistakes.
Step Five – Look for credits and links
Reputable information providers cite their sources (often with hyperlinks) and provide credits for pictures. If an article or report lacks links to background evidence, treat it as suspect.
Step Six – Look for regulatory memberships for news sources
Check if the source is governed by media regulatory bodies, such as a national press council or a broadcasting complaints commission. This demonstrates a willingness to follow established codes of conduct.
Step Seven – Beware the anointment effect
Don’t be blinded by the mention of a prestigious university or a famous publication. AI can manufacture research-like artifacts that look legitimate but might not exist. And a news article about a study by a reputable organisation could be wrong – find the link to the original study and compare what the study actually said with what the report is saying. (This advice is from the sole source not written by me: read Professor Toby Stuart on the subject here).
Step Eight – Use AI to check AI
If a study or report seems too complex to verify quickly, you can ask a different AI tool to identify flaws in the methodology or interrogate the conclusions. Or you can take the output from one AI tool and ask another to fact-check it. Remember that you need enough subject knowledge to know if the second AI is bullshitting you, too.
Finally: Moodie’s second law
If an issue is serious – like a war or a pandemic – assume that anything you see on social media or in an unverified AI prompt is nonsense until you prove otherwise. And don’t share it until you’ve done the fact-checking.
Moodie’s first law? There can never be too many roast potatoes. Which is of course an opinion. But I stand by it, just as I stand by my instinct to look out the f*cking window, every single time.
Link to that downloadable checklist again
Main picture: Alvin Leopold, Unsplash
Previous Sensible Guide articles to check out
What journalists do (part one): the narrowing of the eyes – What is that journalists do, anyway? It’s less about the tools of the trade and more about a certain cast of mind.
What journalists do (part two): the checking of the facts – In a sea of confusing information, journalism is one way of sifting truth from nonsense. How do journalists do that?
Abandon trust. We all have to be journalists now – In our strange new world, there’s a tried-and-tested way to deal with the information overflow…
Step-by-step: How to tell if a news article is reliable – The Internet is big… so big. How is anyone supposed to tell whether a news website is to be trusted or not? Some quick tips to test for legitimacy…
What you need to know about AI hallucinations (hint: they aren’t hallucinations) – When AI makes mistakes, it’s called hallucinating. There’s a better, if ruder, word for it. That means this post contains what the dictionary calls vulgar slang. You have been warned.
Facts, opinions and the Jacques Pauw affair – A post from 2021 about the difference between facts and opinions.
Research skills: Five ways to assess and use reliable sources – With our wondrous phones in our hands, we can all find any information we want. Here’s how to apply research skills to that information.
Recommended reading:
LinkedIn post by Toby Stuart about the anointment effect
If any of this is relevant to your work…
You might find one of these useful: