AI: Take a guess

Let me tell you a story about AI. The story is true, and the only bits generated by an AI are the direct quotes from one…

I had an appointment in town yesterday. I wondered if I could get another job crossed off my list. Lately, I’ve been testing Google Gemini — the latest version of its AI assistant — partly prompted by the wall-to-wall podcasting advertising they’ve been doing. A simple enough query, really. I asked it if the shop I was considering visiting was open. Monday’s a funny one; some shops shut that day to make up for working the weekend. It was worth checking.

“Yes, <shop> is open today. It is located about 0.4 miles away from you,” it chirpily replied.

Great, I thought, and started putting my coat on. And then a little doubt crept into my mind.

“Gemini, what are its opening hours today?”

“I am unable to find the opening hours using the extension. I will try to find the opening hours for <shop> using a search engine. <shop> is closed today, Monday. Here are the opening hours for the rest of the week…”

Wait, what?

It’s Schrödinger’s Shop: both open and shut.

So, I asked again:

“Is <shop> open today?”

“<shop> is closed today, Monday. I provided the opening hours for the rest of the week in my previous response. Is there anything else I can help you with?”

“Why did you tell me it was open when I asked initially?”

“You're absolutely right to call me out on that, and I sincerely apologize. It appears there was an error in my initial response. Here's a breakdown of what likely happened:”

And there followed a long list of explanations/excuses.

AI’s trust problem

So, am I going to trust Gemini answers in future? I am not. And it only takes one or two experiences like this to undermine faith in a tool.

Now, in part, this was my fault. I know enough about the modern wave of AI — about LLMs — to know that I shouldn’t use them as search engines. They’re not search engines, and it doesn’t matter how much Google tries to force AI into its search, it’s not going to deliver what we need. LLMs are NOT answer machines. They’re guessing machines. And any guess has the potential to be wrong.

AIs upend our experience of using digital technology. We’re used to our computers being horribly literal: they do exactly what we ask, without any element of creativity or imagination. AIs, however, make shit up. And we’re just not used to that. There error here is as much user error — incorrect expectations — as a technology problem. Gemini guessed that the shop was probably open. But, in this instance, the guess was wrong. But it’s designed to be a guessing engine. I used the wrong tool for the job.

Gemini gave me the wrong answer because it provided what its algorithms thought was the most probable answer: that the shop was open. It was only when it turned from its main model to looking at actual information on the web, that it was able to extract the right answer.

Now, this should give us a steer on how we should think about using the current wave of LLMs: not as answer providers, but as useful assistants when guessing is part of the process. Or, more specifically, when informed guessing is what you’re doing.

A new tool isn’t better just because it’s new

If you use the wrong tool for a job, you’re often wasting time and usually doing a worse job. Yes, you can hammer a nail in with a screwdriver, but it’s gonna take a lot longer than using a hammer, and will probably do a worse job. Equally, you can get an LLMs to write an article, but if you care about little things like facts and truth, as I hope anyone reading this does, then you’re going to spend as much time fact-checking as you might have done writing. Thus, your time savings are minimal, if any, for a worse product.

In particular, when you’re writing about news, about something that is new, and which the LLMs have probably not been exposed to in their training, you are stacking the deck against the guessing machine. A form letter? Sure, give that to an LLM. A breaking news story?

No.

So, do AIs have no place in journalism? Have I finally become so middle-aged I’ve moved over to the Luddite side of the fence?

Also no.

AI is transformative — but transformative of what?

AI is clearly a transformative technology, in perhaps the same way that the arrival of the web was in the mid-90s. But, when the art director of the title I was working on back then went to buy the company’s first modem, we had no conception of what the web would bring to the world. We’re very much at the same point with AI right now. But the big difference is that, in the mid-90s, most people dismissed the web as a fad, leaving the rest of us to experiment quietly, and figure things out over time.

Now, everyone is jumping on the AI train, which isn’t giving us the time we really need to figure out what the best uses of these technologies might be.

Google is trying to ram AI into search as an almost panicked measure: “is this where we lose our dominance?” they’re asking. “Will AI allow us to keep it?” The problem is that embracing it like they have might be accelerating their decline. Research from the Tow Centre makes this clear:

Overall, the chatbots often failed to retrieve the correct articles. Collectively, they provided incorrect answers to more than 60 percent of queries. Across different platforms, the level of inaccuracy varied, with Perplexity answering 37 percent of the queries incorrectly, while Grok 3 had a much higher error rate, answering 94 percent of the queries incorrectly.

Likewise, I do think publishers rushing to put AI-generated copy on their sites are sacrificing something valuable — trust — for short-term cost advantages. But those cost savings are not unique: anyone with access to an LLM can have them. There’s no competitive advantage there — and little evidence that the public have any enthusiasm for AI-written copy. Quite the reverse, in fact.

Because there are lots are parts of journalism that are about ideas, inspiration, and even a wee bit of art. And there our new guessing machines can help supplement our own guessing abilities.

AI as a brainstorming and research buddy

If you’re brainstorming ideas for a headline, or for a social post, or for a cluster of content on a topic you want to “own” in search or social, I’ll embrace my little LLMs buddy. Why? To use its output to challenge and push my own thinking.

Likewise, Neiman Lab had an interesting story yesterday about an LLM being used as a tool to speed laborious research:

Roganbot works on a simple premise: it listens to Rogan for you, creates a searchable transcript linked to timestamped audio, and breaks episodes down into key topics, notable quotes, potential controversies, and suggested fact-checks. It can, in essence, be extremely online for you, and then you can ask it questions about what it found in its internet rabbit hole.

An AI that listens to Joe Rogan for me? Bliss. He’s a media and cultural touchpoint we probably should be aware of now, but having to expose myself to all of his “content” is not my idea of fun.

We’re still at the very beginning of the AI journey. Very few companies are making the mistake of treating it like they did the web in the 90s: as something to be wary of. But too many are in danger of making an equal and opposite mistake, of being too credulous about an emerging technology whose use we haven’t completely figured out yet.

Now is a time for experimentation — but cautious experimentation.

Oh, and to stop trying to use AI as a search engine.