Humans don't fix AI hiring bias – they copy it
New research shows we follow AI recommendations even when they're wrong
Welcome to FullStack HR, and an extra welcome to the 135 people who have signed up since the last edition.
If you haven’t yet subscribed, join the 10100+ smart, curious leaders by subscribing here:
Happy Saturday,
November has been insanely intense. I’ve done 29 (!) AI lectures in November, sure, it is intense, but I also love what I do, and it is marvelous to meet so many organizations and people. You all help me learn so much!
This would not be possible without all the AI agents I have built, which help me with everything from writing offers to nudging me about my upcoming events.
December is looking almost as intense, and we have our sold-out AI Day on Wednesday, which I’m very much looking forward to! We’ll be 500 people coming together to talk about all things AI-adoption.
🧠 Topics I’m engaging with
Why you shouldn’t count on humans to prevent AI hiring bias.
Perhaps the most important article this week. Human oversight was supposed to prevent AI from warping hiring, but new research says it’s not enough. In a University of Washington study, 528 people worked with simulated AI systems to pick candidates for 16 different jobs. The findings were stark: when picking without AI or with neutral AI, participants picked white and non-white applicants at equal rates. But when they worked with a moderately biased AI, they simply adopted its biases. If the AI preferred non-white candidates, so did they. If it preferred white candidates, so did they. As lead author Kyra Wilson put it: “Unless bias is obvious, people were perfectly willing to accept the AI’s biases.”
My take: I find the study valuable. It’s one of the very few studies we have in this area. But I see two clear limitations. First, they didn’t use true state-of-the-art models; they relied on open-source embedding models, not frontier systems like Gemini, ChatGPT, or Claude. That limits what the results say about today’s best AI. Second, participants weren’t recruiters. They used general online panel users, not trained recruiters. So the decisions don’t reflect real hiring expertise or context.
That siad, what the study still makes very clear:
We can’t trust AI recommendations blindly. People follow AI even when it’s clearly wrong.
We must train specifically for AI-supported hiring. General AI knowledge is not enough. We need practice spotting and resisting AI bias in real decision flows.
We need to push our vendors harder. ATS and screening tools must prove how their models behave, how they test for bias, and how they monitor it.
If humans follow AI this strongly, the burden on vendors must increase, too. And this is HUGE if you’re within the EU, this is exactly what the EU AI Act is designed to regulate. High-risk AI systems in employment decisions will require transparency, human oversight that actually works, and documented bias testing.
Microsoft WorkLab: “AI at Work: Which future of jobs are we building toward?”
Jared Spataro lays out four possible futures based on growth and employment. The early data is clear: companies using AI to augment their workforce are hiring faster than those automating for efficiency alone. Entry-level tech jobs are shrinking, while experienced workers who can direct AI tools are seeing wage premiums. His conclusion: we can achieve high growth AND high employment – but only if we invest in people as aggressively as we invest in models.
This is exactly what I see with my clients. Those treating AI as a headcount replacement are missing the point entirely. The ones who see it as an amplifier of human capability, and actually invest in training their teams, they’re the ones winning. The question isn’t “what will AI do to jobs?” It’s “what are we willing to do to get the future we want?”
Stanford/Carnegie Mellon Study: AI agents work 88% faster but with quality gaps
A new study shows AI agents complete tasks 88% faster at 90-96% lower cost than humans but with 32-49% lower success rates. It’s also worth noting that when humans and agents worked together, task completion increased by nearly 70%. Agents handled the programmable steps, humans took care of what required judgment or creativity.
Interesting data! Not “AI takes our jobs” or “AI is hopeless” but nuanced evidence that hybrid models work best. 70% improvement when human and machine collaborate. I strongly believe that this is how we should do it during the coming years, designing workflows where the right tasks go to the right “worker”, whether that’s a human or an agent. Stop asking “will AI replace us?” and start asking “how do we design the collaboration?” and then design that intentionally.
IBM: “AI Agents in 2025: Expectations vs. Reality”
IBM researchers cut through the hype and deliver a realistic picture. Marina Danilevsky: “You’re still going to have cases where as soon as something gets more complex, you’re going to need a human.” She describes a human-in-the-loop vision where agents settle into an augmented role rather than a replacement one. Key insight: only 13% of employees see agents deeply integrated into their daily workflows yet.
I appreciate this sober analysis. We’re still in the experimentation phase and that’s okay. But organizations starting now with “narrowly defined, low-risk domains” like FAQs and leave requests are building experience that becomes invaluable as the technology matures. The worst thing you can do right now is wait. The second worst thing is to go all-in without understanding the limitations.
🔁 Important updates
Anthropic launched Claude Opus 4.5 – their most powerful model yet. It beats GPT-5.1 and Gemini 3 Pro on coding benchmarks (80.9% on SWE-bench, first model to reach that) and pricing dropped dramatically to $5/$25 per million tokens.
AI could double US labor productivity growth
Anthropic analyzed 100,000 real Claude conversations to estimate AI’s productivity impact. Key findings is that tasks that typically take 90 minutes without AI are completed 80% faster with AI assistance. Anthropic Software developers capture 19% of total productivity gains, followed by operations managers and marketing specialists. The economic extrapolation is that current-generation AI models could increase the U.S. annual labor productivity growth rate by 1.8%, doubling the average rate of growth since 2019.
We’re moving from the age of scaling to the age of research.
Interesting interview with Ilya Sutskever on why we’re entering an “age of research”, deeper ideas about learning, not just more data and compute. Why current models generalize worse than humans, emotions as a model for robust learning, and why the path to superintelligence runs through continual deployment, not lab breakthroughs.
Inside Spotify’s AI Transformation - Spotify CHRO Anna Lundström
My former colleague, Anna Lundström, discusses Spotify’s AI journey. I’ve been puzzled by their silence lately. They used to be a driving force in adoption and new-tech thinking, but it’s been quiet. Until now!
Their approach has been to frame AI as making employees “future-ready” for their careers, not just for Spotify. They’ve opened their (famous) hack week to everyone, not just engineers, and said “work on something AI related”. Plus training, tool licenses, and a cross-functional steering group to coordinate.
Early wins so far have been time savings on admin work, freeing people for strategy and the stuff you never get to.
I found that people are typically quite bad at "pushing back" on AI, i.e., they instantly assume that AI is right and their own instinct is wrong. This could explain the U of Washington study results. And in my opinion, it reinforces the need for strong training before AI Tools are deployed in the workplace.