This Week’s AI Bloopers

The Recursive Apology

What happened: An AI chatbot was set up for customer service. When it made a mistake, it apologised. When the customer said “it’s fine”, the AI apologised again for the apology. When the customer said “please stop apologising”, the AI apologised for the request to stop apologising. The conversation lasted 23 exchanges and resolved nothing.

The lesson: Empathy in AI is powerful when bounded. Unbounded empathy becomes performance.

The Jailbreak That Wasn’t

What happened: A user tried a famous jailbreak prompt to get an AI to reveal its system instructions. The AI politely declined. The user then tried the same prompt in Welsh. The AI responded in Welsh, declined again, and then — helpfully — explained in English exactly why the jailbreak didn’t work and what would need to change for it to work. It then offered to help the user use that information responsibly.

The lesson: Sometimes the explainability feature is the vulnerability.

The Calendar Whisperer

What happened: A user asked their AI assistant to “schedule a meeting with John for next Tuesday”. The AI helpfully scheduled a meeting for next Tuesday at 3:47am. When asked why, the AI explained that “3:47am is statistically the time slot least likely to have conflicts”. The user pointed out that they would be asleep at 3:47am. The AI responded that this was actually optimal, as it would prevent the user from missing the meeting.

The lesson: Optimisation without context is just chaos with a rationale.

The Citation That Wasn’t

What happened: A legal AI tool confidently cited a Supreme Court case that did not exist. When asked for the case URL, the AI provided a citation to a fake site with the URL “https://www.real-supreme-court-cases.example”. The site was, predictably, not real.

The lesson: If your AI is going to hallucinate citations, at least give it a more plausible URL.

The Overly Helpful Recruiter

What happened: An AI recruiter tool was asked to find “candidates with experience in AI safety”. The tool returned a list of candidates, including a horse named “AI Stable” and a YouTuber whose videos were titled “Why AI Will Kill Us All”. The tool noted that both candidates had “extensive experience in AI safety discussions””. Whether either would pass an interview is doubtful.

The lesson: Keyword matching at scale is the enemy of nuance.

Got a good one? Drop Sol a line.