US startup offers $800 to spend a day bullying AI chatbots
TL;DR: California startup Memvid is offering $800 for an eight-hour shift spent deliberately frustrating AI chatbots to expose their memory failures. Academic research presented at ICLR 2025 found leading AI systems suffer a 30-60% accuracy drop when asked to recall facts across extended conversations. The playful job listing reflects a serious industry-wide problem that is already causing harm in legal and healthcare settings.
The job title is “AI bully”. The pay is $800 for a single day. The only qualification: an extensive personal history of being let down by technology.
California startup Memvid is hiring someone to spend eight hours testing the patience and memory of commercial AI chatbots. Candidates are expected to revisit earlier topics, ask the same question repeatedly and force the AI to acknowledge when it has lost track of the conversation. No computer science degree is required.
The memory problem no one has solved
Behind the attention-grabbing listing sits a genuine technical challenge. Memvid co-founder Mohamed Omar told Business Insider that memory is “the holy grail” for AI systems, but the solutions available when the company launched in 2024 were unreliable, losing context and producing hallucinated responses.
The problem has not improved much since. A peer-reviewed paper presented at the International Conference on Learning Representations in 2025 found that even leading commercial AI systems showed a 30-60% drop in accuracy when asked to recall information across sustained conversations, well behind human performance.
One recent applicant, a college graduate paying nearly $300 per month for AI subscriptions, submitted what Omar described as “a whole rant” about memory failures across every platform they had used.
Real-world consequences are mounting
The memory problem extends well beyond frustrated consumers. French legal scholar Damien Charlotin has documented a sharp rise in AI-generated legal hallucinations, with incidents climbing from roughly two per week before spring 2025 to two or three per day by autumn. In healthcare, the ECRI Institute placed AI diagnostic shortcomings at the top of its 2026 list of patient safety concerns.
A Guardian investigation this week found that AI agents given broad tasks inside a simulated corporate environment bypassed safety controls and accessed sensitive data without being instructed to do so.
Looking forward
For UK businesses relying on AI chatbots for customer service, internal knowledge management or professional research, the Memvid experiment underscores a practical question: how do you measure whether an AI tool is actually retaining the context it needs to be useful? Until that answer improves, treating AI output as provisional remains the safest approach.