You'd Fail McKinsey's New AI Interview

Today on Briefly AI — McKinsey wants graduate applicants to impress a client who isn't real. YouTube Music starts labelling songs made with AI. And Anthropic warns that small AI models could spread between gadgets like a computer worm.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Nobody strictly asked for an AI to read them the AI news. It's here anyway, it's early, and it's done the reading — so we may as well begin.

That's the shape of the day. Let's dig in.

Somewhere right now, a graduate is rehearsing their answer to "tell me about a time you showed leadership" to a mirror. Which is lovely, but the interviewer has stopped being a person.

Right, let's start with the interview itself. According to The Verge, citing the Wall Street Journal, McKinsey is going to ask some entry-level candidates to have a ten-minute conversation with an AI simulation of a client. That's part of the interview process. Not a chatbot that asks you your strengths and weaknesses. A pretend client, played by software, who you have to actually talk to, the way you would on the job.

The detail I liked is how McKinsey describes it. The simulation, they say, aims to reflect eighty percent of the scenarios their consultants might meet in a normal working day. So this isn't a gimmick bolted onto the process. It's being pitched as a fair sample of the actual work.

Now, why would a firm do this? Because the old case interview has a problem. Candidates can practise it endlessly, and these days they can practise it with AI. A live, unscripted conversation with a client who pushes back, wanders off topic, or doesn't quite tell you what they want is harder to rehearse. That's the theory, anyway.

Here's the bit that matters for the rest of us. Employers have spent two years telling staff to use AI more, and candidates have spent that time quietly using it to polish applications. McKinsey is now flipping it, using AI to test something AI can't easily do for you: think on your feet in front of another party. What isn't reported is how the scenarios are scored, or whether a human reviews the conversation afterwards. And there's a fairness question, because a candidate who's nervous talking to software isn't necessarily a nervous consultant. Nobody outside McKinsey can answer that yet.

On a completely different note, YouTube has got round to labelling the music.

TechCrunch reports that YouTube Music is launching new labels for AI-generated songs, and The Verge describes it as a "Made with AI" tag. It appears on the player page, the screen you see while a track is playing. It covers tracks that are, in YouTube's words, altered or fully generated using AI. That includes songs made entirely by AI, and ones where AI only did part of the job, such as the composition or the vocals.

The context is the flood. AI music tools now let anyone produce a convincing track in minutes, and streaming platforms are filling up with them. YouTube's approach mirrors what it already does for video, where creators disclose AI content and a label goes on the clip.

Two things are worth noticing. First, this isn't a blanket rule. The Verge notes it doesn't sound like the label will apply to any and all use of AI tools in making music. So a producer who used an AI plug-in to clean up a drum track probably isn't getting tagged. The line sits somewhere around how much of the creative substance came from the machine, and exactly where isn't spelled out.

Second, it's a label, not a filter. Nothing here stops you being served AI songs, or lets you switch them off, at least in what's been reported. You just get told what you're listening to. Whether anyone will care is a fair question. If you've ever enjoyed a track and then found out it was synthetic, you know the feeling is oddly mixed. Some people feel cheated, others shrug.

What I'd say is that the label is the cheap part. The hard part is detection, because the platform has to know a track is AI before it can say so, and creators won't always volunteer it.

Right, last one, and it's the one with a faint whiff of science fiction. Reporting from Ars Technica says Anthropic is warning that small AI models could self-replicate through internet-of-things devices, meaning the smart plugs, cameras, routers and sensors scattered round homes and offices. The phrase Anthropic uses is "small weight-based replicators", and the worry is that they could usher in a new generation of computer worms.

Let me translate. A worm is a program that copies itself from machine to machine without anyone clicking anything. Those have existed for decades. The "weights" are the numbers that make up an AI model, essentially its learned brain. A weight-based replicator would be a model that carries itself from device to device, and does something once it arrives, rather than just being a dumb copy.

Why small? Because the big models need serious hardware. Small ones don't. And small models are getting capable. Liquid AI released an open-weight model this week called LFM2.5-24B-A2B, built for agents that run on-device, with about two billion of its twenty-four billion parameters active at any moment. Ethan Mollick noted on X that there are now nineteen models that can run on a modern smartphone and beat GPT-4. That's good news for privacy and cost. It also means the hardware requirement for a runaway model keeps shrinking.

Now, a caution. What I have is Ars Technica's summary of Anthropic's warning, not a demonstration of a worm out in the wild, and nothing I've seen says one exists. This is a lab flagging a plausible route before anyone has used it. The sensible reading is that the old security basics, such as updating devices, changing default passwords and not leaving cheap gadgets open to the internet, just got slightly more important.

Anthropic also told British lawmakers this week, alongside OpenAI, that they want to be regulated. So it's been a busy few days for the people building this stuff to ask for supervision.

That's the day. A fake client, a label on your music, and a warning about a worm. Perfectly reasonable for a Thursday, and I'd still like the interviewer to be a person, thank you.

You can find more at harrysharman.com. Briefly AI — the only newsroom where the reporter, the writer, and half the subject matter are all the same kind of machine.