Is AI Coming for Your Job? The Predictions vs. The Reality

Published 25 January 2026

Main Feature: AI CEOs predict job disruption, but benchmarks show the tech isn't ready - what does that mean for you?
AI News This Week: ChatGPT ads, Wikipedia payments, AI security funding, and the UK's benefit algorithm mess

The headlines are scary. The benchmarks tell a different story.

Two of the most powerful people in AI sat down at Davos last week and made a prediction that sent ripples through LinkedIn. Demis Hassabis from DeepMind and Dario Amodei from Anthropic both expect AI to hit entry-level jobs and internships hard this year.

Amodei was blunt. He's already seeing it at Anthropic. Fewer junior roles. Less need for the traditional entry points that have defined professional careers for decades.

If you're reading this with a knot in your stomach, you're not alone.

The benchmarks tell a different story

The same week those predictions dropped, a research group called Mercor published something interesting. They built a benchmark called APEX-Agents to test whether AI can actually do white-collar work. The kind of tasks consultants, lawyers, and investment bankers do every day.

The results? Not a single AI model scored above 25% accuracy.

Twenty-five percent. On professional tasks.

The study found that AI struggles particularly with anything requiring reasoning across multiple domains. The moment you need to pull context from Slack, cross-reference something in Google Drive, and apply judgement - it falls apart.

So we've got CEOs predicting massive disruption, and benchmarks showing the technology can't reliably do the job. Both things are true at the same time. Welcome to 2026.

The gap between demo and deployment

This is something I see constantly in my work with businesses. There's a massive difference between what AI can do in a controlled demo and what it can do in the messy reality of actual work.

A chatbot answering customer queries in a test environment? Brilliant. That same chatbot handling an angry customer who's referencing three previous conversations and a policy change from last month? Chaos.

The Mercor benchmark confirms what many of us suspected. AI is genuinely impressive at narrow, well-defined tasks. It falls over when things get complicated.

So why are the predictions so dire?

Hassabis and Amodei aren't stupid. They're not making these predictions to scare people. They're seeing something real.

The answer, I think, is that entry-level work often involves a lot of narrow, well-defined tasks. The kind AI actually can do.

Summarising documents. Drafting initial responses. Pulling together research. Creating first drafts of presentations. These are exactly the tasks that used to fill a junior employee's day. And they're exactly the tasks AI handles reasonably well.

The benchmark failures happen at the senior level. Multi-domain reasoning. Context management. Professional judgement. That's not what interns do.

So both things can be true. AI struggles with complex professional work AND it threatens entry-level roles. The technology doesn't need to be brilliant to change the job market. It just needs to be good enough at the easy stuff.

The anxiety is real - but misdirected

Allister Frost, a business transformation specialist, makes a point I think is worth sitting with. He says workforce anxiety is one of the biggest barriers to successful AI adoption. And most of that anxiety comes from misunderstanding what AI actually is.

People see AI as a replacement. A competitor. Something that's coming for their job.

Frost argues we should see it as a pattern-matching tool. Something that handles repetitive, low-value work so humans can focus on the stuff that actually matters.

I know that sounds like corporate spin. "Your job isn't being replaced - it's being enhanced!" We've heard that before.

But I think there's something to it. The question isn't whether AI will change work. It will. The question is whether that change happens to you or with you.

What this actually means for you

If you're early in your career, the traditional path - internship, entry-level role, gradual progression - might look different. You'll need to demonstrate value that AI can't replicate. Judgement. Relationship building. Creative problem-solving. The stuff the benchmarks show AI can't do.

If you're managing teams, you're going to need to think carefully about what tasks you delegate to AI versus humans. The Mercor research suggests anything requiring multi-step reasoning across different tools is still firmly in human territory.

And if you're just trying to stay relevant, the best thing you can do is get comfortable with AI as a tool. Not because it's going to replace you. But because the people who know how to work alongside it will have an advantage over those who don't.

The predictions are scary. The reality is more nuanced. And somewhere in between is where most of us will spend 2026 - figuring out how to navigate a world where AI is genuinely useful but not quite as capable as the headlines suggest.

That's not a comfortable place to be. But it's probably the honest one.

AI News This Week

ChatGPT is getting ads. OpenAI announced it will start showing targeted advertisements to users on its free and Go tiers. The ads will appear at the bottom of conversations, tailored to what you're discussing. You can dismiss them or disable personalisation, and OpenAI promises they won't influence the chatbot's responses. Whether that holds up in practice remains to be seen. TechCrunch

Big Tech is now paying Wikipedia for training data. Amazon, Meta, and Microsoft have joined Wikimedia's Enterprise programme, paying for access to the same content they've been using for years. The Wikimedia Foundation says it needs the money to sustain its open knowledge model, especially as AI-driven search threatens traffic to the actual site. It's an interesting shift - companies paying for data they technically already had access to. The Decoder

VCs are betting big on AI security. Witness AI just raised $58 million to help enterprises monitor "rogue agents" and "shadow AI" - unauthorised tools employees use without IT approval. The startup tracks AI usage across organisations and flags compliance risks. With predictions that AI security software could hit $1.2 trillion by 2031, expect more money flowing into this space. TechCrunch

Only 5% of AI pilots deliver measurable value. That's according to new research on enterprise AI adoption. Most projects fail to reach production because of inadequate infrastructure, limited data access, and rigid integration processes. The fix, apparently, is "composable and sovereign AI architectures" - which is consultant-speak for systems that are more flexible and don't lock you into one vendor. MIT Technology Review

UK young adults want AI financial advice. Research from fintech Cleo found that 20% of 28-40 year olds are curious about using AI to manage their money, with younger respondents more open to the idea. The main barriers? Trust and a preference for starting small. Interestingly, 37% admitted they struggle with impulse spending - which might explain the interest in having a robot tell them to stop buying things. AI News

The UK's DWP algorithm got it badly wrong. A cautionary tale in AI accountability - the Department for Work and Pensions deployed an algorithm that incorrectly flagged numerous benefit claims as fraudulent. It's a reminder that autonomy without oversight leads to real harm. The broader point: trust in AI is declining, and rushing deployment without proper governance isn't helping. AI News

Would you like to join our Thursday AI cook-along session? Register: Here