Ravens, AI, and who to hire

I have been reading Adrian Tchaikovsky’s Children of Memory, which at one point stages a philosophical debate about raven consciousness. The ravens in the book are doing things that look like intelligence. The question the novel keeps circling is whether it actually is.

It made me think of AI today, so I went and looked at the research.

Ravens can plan ahead. In a 2017 study, they prepared for tool use and bartering with delays of up to 17 hours - outperforming chimpanzees and four-year-old children under some conditions. They deceive competitors about food caches, accounting for precisely what each rival can and cannot see. They improvise tools. By four months of age, they match adult great apes across a broad cognitive test battery. Scientists called it the clearest evidence for future planning in a nonhuman animal.

Separately, a 2018 meta-analysis of experiments where corvids dropped stones into water tubes to retrieve food concluded that trial-and-error learning, not genuine causal reasoning, explained most of the impressive results.

The scientists genuinely disagree. Ravens look like they understand cause and effect. Whether they actually do is unresolved.

So to AI. Whether large language models genuinely understand or produce brilliant pattern-matching outputs that look like understanding is equally unresolved in AI research. Both produce outputs that can be indistinguishable from reasoning. The mechanism behind both is contested.

This is where the human layer matters - and it challenges most of the commentary online right now.

The concern about AI is usually about atrophy. A 2025 Microsoft study found that the more confidence people had in AI’s ability to perform a task, the less critical thinking effort they applied.

A separate study found measurable decline in critical thinking ability linked to AI reliance, strongest in younger users.

But the research misses something.

Critical thinking is not just something AI might erode. It is the thing AI requires most of all to use well. Interrogating the output, reframing the question, spotting what the model missed. That is where the value actually comes from.

Ravens cannot critique their own problem-solving. They get the answer they get. The people getting genuinely better results from AI are not the ones who prompt more sophisticatedly. They are the ones who push back, redirect, and take responsibility for the final call.

For anyone building a team right now, this is the practical implication. AI has commoditised a significant slice of domain knowledge and experience. What it has not replaced - and cannot - is the judgment to evaluate what comes back.

That is becoming the quality you hire for most.

The Impact of Generative AI on Critical Thinking - Microsoft Research

The Microsoft study referenced above in full. The core finding - that AI assistance and critical thinking are trading off against each other in real-time, measurable ways - has direct implications for how you design onboarding, workflows, and the way your team reviews AI outputs.

Run a premortem on your next AI output before you act on it. A premortem assumes the decision has already failed and asks what went wrong - it is one of the most effective ways to surface blind spots before they cost you anything.

Once AI gives you a recommendation or plan, follow up with something like: “Assume we acted on this and it went badly. What are the three most likely reasons it failed?” The model will often surface risks and gaps it did not volunteer the first time. That gap - between the initial answer and what the premortem reveals - is exactly where your critical thinking needs to live.