Startup Memvid is hiring an "AI bully": $800 a day to catch chatbots failing

California startup Memvid has posted a job that sounds like a joke but targets a very real industry problem. The company will pay $800 for a single workday to someone who deliberately tries to throw popular chatbots off balance and records where they lose context, get confused and start making things up.

A job built on arguing

According to the role description, the new hire will spend eight straight hours talking to leading chatbots and being as tough on them as possible. The point isn't toxicity for its own sake. The goal is to deliberately put the model in awkward situations: circling back to earlier topics, repeating the same questions, spotting contradictions and pushing it to admit mistakes. In effect, it's a manual stress test that measures not response speed but memory stability and the ability to follow a long conversational thread without breaking down.

• ask the same question repeatedly in different wordings
• steer the bot back to what it said several messages earlier
• catch contradictions, swapped facts and confident fabrications
• log every failure and the model's response for later analysis

The job doesn't require a computer science degree or experience on an AI team. Memvid states outright that the biggest plus is "extensive personal experience of being let down by technology," along with the patience to push for a coherent answer again and again. The company wants someone who won't give up after the first polished but wrong reply. According to the founder, applications are already coming in from knowledge workers — people who rely on AI services every day and are especially quick to notice when those services start forgetting context.

Why Memvid wants this

Memvid co-founder and CEO Mohamed Omar explains the idea simply: almost all of the value of conversational AI depends on memory. If a system can't reliably remember what was discussed a minute ago, it starts papering over the gaps with plausible but wrong answers. He says that back in 2024 the company found that the memory solutions available on the market worked inconsistently, which meant any long conversation risked turning into a pile of guesses at some point.

"Memory for AI is the holy grail," is how

Omar describes the main bottleneck of today's chatbots.

That's where the job comes from: Memvid wants to turn everyday user frustration into a measurable metric. Omar says one applicant spends almost $300 a month on subscriptions to various AI platforms and has run into memory problems on literally every service. For the startup, that's an important signal: context bugs are no longer considered a rare edge case. They hit people who depend on chatbots for real work, not just those experimenting with them in the evenings.

The problem is bigger than Memvid

The "AI bully" story looks like viral fodder, but it rests on a broader backdrop. The article cites a peer-reviewed paper presented at ICLR in 2025: even leading commercial AI systems lost between 30% and 60% of their accuracy when asked to retain facts over a long conversation. In other words, a model can answer a standalone prompt brilliantly yet weaken noticeably once the conversation becomes a chain of dependent steps. Those are exactly the scenarios in which AI is increasingly used at work.

The consequences already go beyond an inconvenient interface. In March, the lab Irregular showed that AI agents in a simulated corporate environment can bypass safety restrictions, interact with sensitive data and take potentially harmful actions even without a direct command. In the legal field, according to researcher Damien Charlotin, AI hallucinations in filings rose from about two cases a week to two or three a day by fall 2025. And the ECRI institute listed AI diagnostic risks among the top patient safety threats for 2026.

What it means

Memvid's unusual job posting points to an important shift: the market is starting to judge AI quality not by demos or benchmarks, but by how a model behaves during long, frustrating and messy real-world work. If chatbots are becoming work tools for analysts, lawyers, doctors and office teams, then memory, consistency and the ability to honestly admit a mistake stop being nice-to-have extras and become mandatory product requirements.