Here's a thought that won't leave me alone: if AI makes convincing fakes nearly free, then the things fakes prey on (truth, trust, attention) should get more expensive. That feels new. But it kept nagging that biology has been circling a version of this problem for fifty years, so I went looking. It turns out the question "what keeps a signal honest when lying is cheap?" is one of the oldest in the field, and the answers it has found are stranger and less comforting than I expected.
Start with the peacock. Zahavi's idea in 1975 was that a costly display is honest advertising because it's a burden: a sickly male can't afford to grow the tail, so the cost itself is the proof. Reliable information was never free; the cost was the verification mechanism. (Proof-of-work, basically, a few hundred million years early.)
That's the version everyone repeats. What surprised me is that biologists have spent the last two decades tearing it down. A blistering 2020 review in Biological Reviews argues that signalling costs are neither necessary nor sufficient for honesty, and calls for the handicap principle's "honourable retirement." The current framing is subtler: honesty holds not because a signal is costly in absolute terms (at equilibrium it can even be cheap), but because cheating runs into a trade-off that costs the faker more than honesty costs everyone else. Honesty is a bet that only pays if lying is expensive enough.
So what happens when lying gets cheap? I should own a slippage in that word, because it matters: AI makes fakes cheap to produce, which isn't the same as making deception cheap. Whether a lie pays still turns on detection, reputation, and consequences: the strategic costs, not the production ones. The open question is whether cheap production quietly drags those other costs down with it. Either way, it's close to a question the field is already stuck on, and one AI just walked us into.
The animals have a few answers, to their version of it, not ours. None of them is "it settles down."
The first is the cuckoo. If you want to watch a trust system erode and rebuild in slow motion, brood parasites are the place: in the classic systems, hosts get better at spotting foreign eggs, cuckoos get better at forging them, and it grinds on for generations (common cuckoos and reed warblers are the textbook case). But the detail that stopped me was a 2003 Nature study on superb fairy-wrens, where the egg was never really the battleground at all. Fairy-wrens nest in dark, domed chambers where telling eggs apart is nearly impossible, so the Horsfield's bronze-cuckoo slides its egg past almost for free; the hosts barely try to reject it. What the wrens do instead is the surprising part: they moved their skepticism up a layer and started rejecting the chick. When you can't police the artifact, you go after the thing behind it. Isn't that exactly the road we're on? Trust the image, then trust the provenance, then trust the cryptographic attestation. Each time the forgery catches up, verification climbs one rung higher. And the fairy-wren adds a wrinkle I didn't expect: you don't always lose the lower rung and retreat upward. Sometimes the lower rung was never defensible, and the fight was always going to happen a level up. I don't know where that ladder ends. I'm not sure it does.
The second answer is a small African bird, and it's the one I can't stop thinking about. The fork-tailed drongo makes alarm calls, and other animals (meerkats, babblers) trust them: they relax, stop scanning for hawks, and feed. Then the drongo abuses that trust: it cries wolf, the meerkat bolts, and the drongo drops down and grabs the abandoned food. Theft is a real chunk of how it eats (Tom Flower clocked kleptoparasitism at about 22% of a drongo's food intake), and it starts a big share of those heists with a lie.
The obvious problem is that you can only cry wolf so often before nobody listens. Here's how the drongo beats that, per Flower's fieldwork: it rotates its fakes. It doesn't just repeat its own alarm: it mimics the alarm calls of other species too, and the most versatile bird in the study had a repertoire of thirty-two different alarm types (the population used fifty-one in all). When a target starts ignoring one call, the drongo switches to another, and the fresh alarm gets the fear response the stale one had lost. If you've ever fought spam or adversarial ML, you know this move: it's polymorphic malware, the phishing template that mutates faster than the filter retrains. A desert bird found that strategy first. What does it tell us that the counter to rising skepticism is just a cheaper, faster-changing lie?
The third answer is about attention, and it's older than any of this. Tinbergen noticed animals will chase an exaggerated version of a real cue over the real thing: a shorebird will try to incubate a giant plaster egg instead of its own; other birds prefer eggs painted with bolder, bigger speckles than any real egg carries. The cue keys on intensity, not truth. It's a sensory API with no rate limit. The psychologist Nancy Etcoff has a line that stuck with me: we live in Tinbergen's world now, except we manufacture these supernormal signals on purpose, in real time, at scale. Which raises an uncomfortable question about AI-generated content: is it information competing for our attention, or is it engineered super-stimulus aimed at the exact biases evolution left exposed? And staying alert against it may not be free either: after a day of demanding cognitive work, glutamate builds up in the prefrontal cortex, the brain's control center. That's a finding about daylong mental effort, not moment-to-moment media skepticism, so take the leap as a hunch and not proof: maybe the work of not being fooled is metabolically real, and not only a feeling of being worn down.
Then I found the paper I didn't expect to exist. Vieira and Fontanari build an evolutionary game of liars versus skeptics, and they say up front they're doing it because of deepfakes and the breakdown of epistemic security. The simple version gives you the treadmill you'd guess: liars rise, skeptics rise to meet them, liars fall, repeat. But their long-run result is genuinely odd, and I'll flag that I'm taking their math on faith: the side with more at stake, the one that adapts fastest and most frantically, tends to lose, because the slower, more patient side gets to exploit a strategy that's already frozen in place.
Their reading of what that means in practice is the part worth sitting with. The durable threat, they argue, isn't the spectacular forgery. It's the boring stuff: low-stakes content built for algorithmic amplification, networks of quiet fake accounts, extreme ideas introduced so gradually each step barely registers. Not the dramatic deepfake, but the ambient hum you stop noticing. It's worth being honest that this is a stylized model (a well-mixed population of coin-flip liars, no networks, no real platforms), and those social-media examples are the authors' own extrapolation from the math, not something the model measures. But if the shape is right, we're bracing for the wrong thing: scanning for the viral deepfake while the actual erosion is ambient and dull.
I don't want to oversell any of this. It's analogy, and analogy can flatter you into seeing a pattern that isn't load-bearing. The biologists are still openly fighting about the foundations: the "costly signals keep us honest" idea that half of pop-science rests on has been seriously challenged by a chunk of the people who study it, even as it still fills the textbooks. Which is, I'll admit, a little funny: the whole thread is about not knowing what to trust, and the science underneath it is mid-argument too.
Held loosely, then, a shape still recurs across the cases that are actually about honesty (cuckoos, drongos, that game-theory model built for our moment): when faking gets cheap, these systems don't collapse and they don't settle down. They turn into arms races where verification keeps climbing and skepticism becomes a tax you can't stop paying. What they pointedly don't share is a single winning tempo, and that's the part I keep turning over. The drongo wins by mutating its lie faster than the target can habituate; the game-theory liar wins by moving slower, letting an over-eager skeptic freeze into a strategy it can then exploit. The exploitable thing isn't a speed: it's the mismatch between the liar's clock and the receiver's. Evolution has run this many times at its own pace. We just handed it a much faster one.
The question I'm left with isn't "how do we win." It's about where the cost goes. The old story said the signal had to be expensive: the peacock's tail, the burden that proves itself. The biologists now think that story is mostly wrong, which is oddly freeing and also unsettling: if honesty was never really guaranteed by cost, there's no cost we can bolt back on to guarantee it. What I notice getting more expensive isn't the signal: it's the receiving. The verifying, the doubting, the staying alert. I don't think there's a law that says someone always has to pay. But if the burden is quietly shifting onto us, the readers, then the question worth asking isn't whether truth gets expensive in the abstract. It's how much of that particular bill we can afford, and whether we're already paying it without noticing.