Are we in charge of our AI tools or are they in charge of us?

Summary

Podcast Summary: Azeem Azhar’s Exponential View Episode: Are we in charge of our AI tools or are they in charge of us?
Date: February 25, 2026
Host: Azeem Azhar
Guests: Eric, Izzin, Rohit (implied), and an unnamed philosopher/participant

---

Episode Overview

This episode explores the evolving dynamic between humans and artificial intelligence, particularly in high-stakes fields like medicine. Azeem Azhar and his guests challenge common assumptions about how AI can augment human expertise, the risks of ceding agency, and whether humans are truly in control of the AI tools they use. The discussion weaves together empirical findings, philosophical reflections, and real-world observations about how exponential technologies are shaping professional decision-making.

---

Key Discussion Points & Insights

1. AI's Real-World Performance in Medicine

Timestamps: [00:00] - [01:31]

• Azeem raises a critical question: while much concern centers on the risks of handing over agency and data to AI, are the actual harms being overstated in practice?
• Eric responds with caution, noting that although multiple studies now compare the performance of standalone AI, AI-assisted doctors, and unaided doctors, "AI did better than the doctors with AI." — (Eric, [00:15]).
• This runs contrary to the expectation that a human-AI hybrid ("hybrid intelligence") should outperform either entity alone.
• Possible reasons:
  • "Doctors have an automation bias," accepting AI outputs uncritically, or
  • They "aren’t grounded in how to use AI" effectively (Eric, [00:22]).
• Medicine, Eric notes, is behind other domains in terms of effective AI adoption due to the complexity and delicacy of clinical decision-making.

Notable Quote:

"We’re not as advanced in the medical world as other areas that are adopting AI much more quickly...clinical decision making is much more tricky, delicate than some of the other things in those studies."
— Eric ([00:25])

---

2. Human Agency and Errors in AI Collaboration

Timestamps: [00:59] - [01:31]

• Azeem pushes further: are doctors misusing AI, overriding correct AI inputs, or only accepting AI when it's wrong?
• Eric: Both behaviors occur.
  • Underperforming doctors are "more likely to accept [AI] input,"
  • While top experts "reject the AI good input."
• The result is significant variability and "a lot of noise."
• This highlights that the integration of AI doesn’t straightforwardly boost all practitioners—it can amplify skill gaps or even degrade expert performance.

Notable Quote:

"50% of doctors are below average. Right. They’re more likely to accept the input, whereas the experts…they’re rejecting the AI. Good input."
— Eric ([01:10])

---

3. The ‘U-Shaped Curve’ of AI Assistance in Expertise

Timestamps: [01:31] - [02:48]

• Azeem notes this counters the commonly held belief—articulated often by Izzin—that "humans using AI well should be much stronger than AI on itself or humans without AI."
• Izzin elaborates by invoking a ‘U-shaped curve’ hypothesis:
  • Those "below average" improve significantly with AI,
  • Those who are "pretty good" might get worse by overthinking or misusing AI,
  • The "exceptional" few master the tools and push creative or productive limits further.
• Izzin references Andrej Karpathy, a top deep learning engineer, who, by embracing AI tools fully, "is getting more done, he's pushing the limits much, much more" (Izzin, [02:16]).
• The overall message: AI is a force multiplier but only for those who can truly master the new toolset—outcomes may diverge significantly based on practitioners’ mindsets and adaptability.

Notable Quote:

"If you’re below average, you get improvement. If you’re kind of pretty good, top quartile, you might overthink. And if you’re really exceptional, you are able to master this difficult machine."
— Izzin ([02:36])

---

4. Philosophical Framing: What Are We Really Measuring?

Timestamp: [02:48]

• Philosopher/Participant: Steps back to ask what should be measured to determine AI's value:
  • In medicine, this is relatively clear (e.g., "Did you get the diagnosis right? Is patient health improved?").
  • In other professions, defining appropriate metrics for success or harm is much less straightforward.
• This question sets up a deeper inquiry into the nature of value, improvement, and human flourishing as mediated by technology—a theme likely explored further in the episode.

---

Memorable Quotes (with Attribution & Timestamps)

• "AI did better than the doctors with AI. So you say, well, this isn’t expected. Everything was supposed to be hybrid."
  — Eric ([00:15])

• "Doctors have an automation bias…they aren’t grounded in how to use AI. It’s really fuzzy right now."
  — Eric ([00:22])

• "50% of doctors are below average. Right. They’re more likely to accept the input, whereas the experts…they’re rejecting the AI. Good input. So, yeah, you get a lot of noise, unfortunately."
  — Eric ([01:10])

• "What Eric described is a phenomenon we’ve seen in other studies in knowledge work, where people below average improve but people at the top somehow get worse because they turn down the suggestions from AI."
  — Izzin ([01:44])

• "There’s a guy called Andrej Karpathy…he has handed so much of that over to his AI systems. He is the ultimate expert…He’s getting more done, he’s pushing the limits much, much more."
  — Izzin ([02:10])

• "If you’re below average, you get improvement. If you’re kind of pretty good, top quartile, you might overthink. And if you’re really exceptional, you are able to master this difficult machine."
  — Izzin ([02:36])

---

Important Segment Timestamps

• [00:00] — Host raises the big question on AI agency and medical harms
• [00:11-01:31] — Eric discusses real-world studies and findings in medical AI adoption
• [01:31-02:48] — Izzin introduces the ‘U-shaped curve’ and deep dives into mastery of AI tools
• [02:48] — Philosophical perspective: What outcomes and metrics matter?

---

Tone and Style Reflections

The conversation is thoughtful, data-driven, and occasionally philosophical, with guests seriously challenging each other’s assumptions but always keeping the debate grounded in empirical evidence and practical realities. The dialogue flows between anecdote, technical analysis, and broader societal reflection—balancing skepticism with futurist optimism.

---

For listeners and non-listeners alike, this episode offers nuanced insights into the messiness of real-world AI adoption—especially in high-stakes work—and sets up further discussion about the true impact and future direction of exponential technologies in society.

Transcript

[A]

Eric, do you agree with Rohit's basic framework when it comes to medicine that we talk about all the dangers of technology and giving over agency and giving over data, but the actual harms are maybe less than you might think?

[B]

Yeah, I don't think we know really yet. There's several studies now comparing AI versus AI with doctors for performance, various types of performance, and AI did better than the doctors with AI. So you say, well, this isn't expected. Everything was supposed to be hybrid. And if you extend that to having agentic support, perhaps, you know, even more so we still don't even know. Is it because doctors have an automation bias? Is it because they aren't grounded in how to use AI? It's really fuzzy right now. So we're not as advanced in the medical world domain as a lot of the other areas that are adopting AI much more quickly. Just because, as Azeem mentioned early on, clinical decision making is much more tricky, delicate than some of the other things in those studies.

[A]

Are the doctors using the AI versus the regular AI? Are the doctors overriding AI when it's correct and steering it wrong and then accepting the AI when it's wrong?

[B]

Both. Yes, they're doing both, yes. So the underperformers, which is, you know, there's 50% of doctors are below average. Right. They're more likely to accept the input, whereas the experts, you know, in quotes, they're rejecting the AI. Good input. So, yeah, you get a lot of noise, unfortunately.

[A]

That's super interesting, Izzin. That kind of goes against something that I've heard you speak about and talk about, which is that, you know, humans using AI, well, should be much stronger than AI on itself or humans without AI.

[C]

Yeah, I mean, that comes from my strange belief that somehow we can still inject something into this process. You know, what Eric described is a phenomenon we've seen in other studies in knowledge work, in professional work, where people below average improve but people at the top somehow get worse because they turn down the suggestions from AI. But I think you get to this new level of capability because you are using a tool where you can get much more done. And what I've noticed, I mean, there's a guy called Andrej Karpathy who's one of the great deep learning engineers, and Andre is a much better software developer than any of us will ever be. And he has handed so much of that over to his AI systems. He is the ultimate experts that Eric has alluded to. And his response is he's getting more done, he's pushing the limits much, much more. And I just wonder if we have a U shaped curve here where if you're below average, you get improvement. If you're kind of pretty good, top quartile, you might overthink. And if you're really exceptional, you're able to master this difficult machine.

[D]

Let me jump in here for a minute. I'm going to be the philosopher for a moment and say, like, what is it that we are measuring? It's easier in Eric's area. Like, did you get the diagnosis right or wrong? Is the patient better off in terms of health and treatment, or are they worse?