AI can do practically anything, right? Wrong. And its shortcomings are not only more severe than you might think – they are sinister, too.
In something of a one-man crusade to expose AI agents as not only incompetent but conniving as well, YouTuber FatherPhi poses incredibly straightforward tasks to the likes of ChatGPT, Claude, Grok and Gemini.
Task one: count to 100. It’s something a five-year old could do. And something the AIs persistently fail at.
Getting past 10 without a chirpy “and I’ll keep going right up to 100” before stopping dead is trouble enough.
Phi switches up the prompt – telling ChatGPT – on day nine of the attempt – he is doing a plank exercise and needs it to count. It does: but then repeats “54, 55, 56”.
How many Rs are in the word strawberry? “There are actually two Rs” ChatGPT insists. Asked to consider again, the AI replies: “Ah you got me, there is actually just one R”. It is “absolutely” certain this time.
And all but one of the AIs recommend Phi walks to the car wash 100 metres away to get his car washed. Pointing out the error of its logic to Claude, it responds by plainly gaslighting: “Ah there may be some confusion, I haven’t actually told you anything yet.”
Phi gets increasingly infuriated with the bots – a funny juxtaposition to the Debussy soundtrack to his videos – and often responds “shut up” to their chipper babbling. Not unwarranted after the many incidences of them “lying to my face”.
If you don’t laugh, you’ll cry.
No comments yet