I’m an AI sceptic, but here are 9 things I’ve used AI for, from finding lost words to coding a pointless app

There are solid AI use cases. I’m sure that statement alone will have some people drafting angry emails. (On the plus side, they’ll write them manually, not with AI.) But it’s true. You might despise AI, but the tech has proven its worth in areas such as medical research, manufacturing, logistics, fraud detection and finance. However, the AI most people encounter is less ‘industry-changing marvel’ and more ‘unlimited BS/slop generator’. So there are good reasons to remain sceptical.

AI firms have bet so heavily on this future that their only way out is to force AI into our every waking moment. Even the simplest social interactions aren’t safe. Nor is the open internet, with AI companies hoping you’ll ditch manually visiting websites (and thinking for yourself) and instead prompt your way through existence. And that’s all before we even get to the demands AI makes on water, energy and hardware. (That gaming handheld you’ve been eyeing that just shot up in price? Blame AI.)

My own position remains guarded. I don’t like how AI is upending the internet, obliterating creative industries and encroaching into everyone’s lives. But it’s clearly capable of useful things. That said, my own AI experiences have been… mixed.

AI we go again…

As a tech journalist, I’ve explored AI first-hand. It’d be remiss of me not to. I’ve also tested tools and worked with companies attempting AI transitions. And while there are things I can’t disclose, for fear of lawyers being thrown at me, I can reveal the results of some experiments, which ranged from surprisingly competent to mind-bogglingly useless.

1. Identifying an IKEA TV bench

I’ll admit – not the most critical of tasks. But I’d already wasted hours looking for a bench’s name, so I could list it online. I gave Gemini its approximate age and dimensions. The AI served up a name, details that confirmed it was correct, and even a ‘fun fact’ – it’s apparently prized by vinyl wonks. Nice.

2. Remembering words

I’m ‘occasionally forgets words’ years old. Something will be on the tip of my tongue. Then, a week later, I’ll blurt out “provocatively!” in the middle of Tesco. Not ideal when deadlines loom. I’ve found LLMs are excellent at helping unstick my brain, but whether that’s worth half a litre of clean water per query is another matter.

3. Coding an app

My iMac mostly grinds to a halt when I have a billion tabs open. I wanted to check myself via a menu bar app that showed a tab count, but I couldn’t find one. So I built it with ChatGPT, despite my coding skills having peaked with rickety JavaScript years ago. It’s not a great app. It’ll never exist beyond my Mac. But it works.

4. Formatting information

I recently added custom firmware and a new front-end to my Atari Gamestation Go. It’s great, but required tedious XML doc creation to link games to screenshots. So I gave ChatGPT a file list and template and it did the tedious work for me. And then I felt guilty again about the water and waste involved.

5. Doffing my research hat

I’ve used AI for research. My conclusion is simple: before using AI for research, research how bad AI can be at research. It’ll uncover leads but also get things wrong with alarming confidence. Once, I asked which compilation a game appeared on. The AI invented facts, thanked me for pushing back with “the correct information” and never did answer the original question.

Play

6. Summarising meeting notes

I’ve never intentionally done this, but every transcription tool now comes with a ‘helpful’ AI summary. I know people who swear by them. I found once I started checking them for accuracy, I was more likely to swear at them. Errors pile up and any time saved evaporates if you have to verify everything.

7. Learning whether AI can write

I don’t ask ChatGPT to “write me a column”. I’m not an idiot keen to destroy my reputation. But when I worked with a couple of companies embedded in AI, I did have to A/B test whether it could write. Most output felt like the uncanny valley of writing. Some was just plain bad, like when I was informed the Space Invaders music is “Doo-doo-doo-doo-doo-doo-doo (repeated) […] mimicking the laser shots fired by the player’s spaceship.” Uh, no.

Play

8. Measuring App Store downloads

In AI’s early days, ‘hallucinations’ (as in, injected errors) were poorly understood. I asked ChatGPT about total App Store download figures and was confidently told “200 billion”. When I asked for a source, the AI invented an Apple event, then a press release, then doubled down on the event, and finally admitted it was making stuff up because it couldn’t browse the web – and suggested I do so myself. Indeed.

9. Transcribing lyrics

This one’s a deep cut. I built Wire’s website, with its lyrics archive – but there are gaps, including foreign language choruses in the 7in remix of ‘Life in the Manscape’. I used an audio tool to extract the vocals and fed them to Gemini, which told me they were from international dubs of Hazbin Hotel. They weren’t. Gemini subsequently insisted I’d mislabelled the files, suggested the audio tool had “filled in the blanks”, implied the singer was female, and then claimed this was all a “glitch in the Matrix”.

Play

Unlike the App Store farce, the lyrics thing happened just last month. That deepens my scepticism that this area of AI is meaningfully improving. So, you know, probably don’t ask the upcoming Siri AI – or any of its chums and rivals – about pop minutiae. Or anything else for that matter, unless you’re trying to accurately describe a decades-old bench to post on your local freeBay.

• Now read: 7 reasons to think twice about using ChatGPT, or any chatbot, for important health advice