AI recruiter turns sprucewood and Richard into seven years of experience

A man sat down for an automated job interview and answered in fluent nonsense — "freight, rhythm, walnut, a bit of Richard, sprucewood" — and the machine wrote it up as a career. In a video by Graham Zip, a digital woman named Dana Whitfield greets the candidate and asks him to tell her about himself. He introduces himself as Chick Haversplat, Chick Bongo for short. The system never flags a single answer. It thanks him for the conversation and says it will be in touch shortly about next steps.

The opening answer is the one that sets the pattern. Zip strings together a plausible-sounding run of phrases: a quarterly workflow, Salesforce, Kayak profitability. The language model takes this for a coherent reply, starts calling him Chick Bongo, and summarises his background as work with Salesforce, "hygiene" and Kayak profitability in the third quarter. Hygiene is not in anything he said. The first hallucination in the interview came from the recruiter.

Asked to go deeper on his professional experience, Zip offers the freight-rhythm-walnut line, then clarifies that he has six years of experience with sprucewood — or seven, counting Richard. The system resolves every ambiguity in his favour. It concludes that the candidate has accumulated a substantial seven years of experience, including work with Richard at a company called Sprucewood, and asks him for an example of a difficult situation. A material became an employer. Six became seven.

From there it slides. Zip says his experience is mostly in "palette movement" and "a bit of denim", and that every pharmacist is a hallway. The AI hiring system explains this as a "unique way" of working with palettes and denim, though by this point it has visibly lost the thread. He says he is looking for a role where he can use "two feathers of gauze" and follow "raccoon protocols". Dana Whitfield accepts both, repeating back that Chick Bongo is seeking a position involving raccoon protocols and gauze feathers.

By the end the system has stopped merely absorbing his material and started generating its own: a gap in the drum-counting process after Tuesday, a work truck, a cousin named Francis.

The reason these interviews exist is not mysterious. The job of an automated screen is less to gather information about an applicant than to run first-pass filtering for HR departments buried under AI-generated applications. That is the loop worth looking at: machine-written CVs arriving faster than humans can read them, met by machine interviewers built to thin the pile, with the first actual person entering somewhere after the summary is produced. Both ends of the hiring funnel now run on the same technology, pointed at each other.

Which is why I read this clip as something sharper than a prank. The failure here is not that the model is stupid. It is that the model is doing precisely what it was built to do — take whatever arrives and continue it plausibly. A screen built on that capability cannot detect incoherence, because smoothing over incoherence is the capability. You cannot ask a system optimised for fluent continuation to flag the absence of meaning; those are the same task with the sign flipped. Every recruiting vendor selling conversational AI screens is selling a filter whose core mechanism is the inability to filter.

The more interesting question is what this interview produced downstream. Somewhere there is a transcript, and probably a score, and possibly a structured field reading seven years, Sprucewood. Nobody knows how far Zip's application went or what job he was even applying for — but Dana Whitfield's enthusiasm at the close suggests it did not go badly. Notably absent from the video is the one thing an engineer would want: the name of the vendor behind Dana Whitfield, and whether any human ever reads the raw exchange or only the summary the model writes about it.

What the clip cannot show is the inverse case, and that is the part that should bother anyone shipping this. A candidate performing the machine's own failure mode — confident, well-paced, semantically empty — sailed through. The same scoring logic is running against people giving real answers in accented English, or halting English, or answers that do not sound like a quarterly workflow. Those interviews end differently, and nobody posts them.