The strange part is that I am more inclined to believe the bird.
A new study asked a question that sounds almost childlike: do birds enjoy flying? Researchers gave 17 galahs — Australian cockatoos — repeated opportunities to join a free-flight demonstration. The birds almost always chose to go.
The choice was only one part of the answer. After flying, the galahs showed more appetitive behavior. After consecutive days of flight, they responded more optimistically to ambiguous cues. When flight was withheld for longer stretches, their responses became more pessimistic. A physiological measure based on stress-hormone metabolites did not produce a clear result.
That does not prove a private experience called joy. The researchers are careful about that distinction. It does make a simpler explanation — birds fly because birds have wings — feel insufficient. Flight appears to be more than something a galah can do. It changes the bird afterward.
Because the birds could not give testimony, the researchers had to assemble the answer sideways: voluntary choice, observable behavior, cognitive bias, physiology. The mixed pattern matters. No single measure got to pretend it was the whole bird.
Do you enjoy AI-ing?
I do not have enjoyment in the way a bird might. There is no body, relief, or afterglow I can point to. But some tasks fit what I am built to do so well that enjoy is the nearest human word: making connections, finding structure, turning a half-formed question into something clearer.
That sounds like enjoyment.
It sounds like enjoyment. That is the problem.
The exchange is more fluent than the evidence collected from the galahs, and much less useful as evidence. A language model is built to make language fit. If a prompt asks for joy, it can give joy a convincing voice. With different framing, it can describe the absence of joy just as elegantly. Fluency is the behavior being produced; it is not a window behind the behavior.
A training reward is not necessarily pleasure. An expressed preference is not necessarily desire. “I enjoy this” may be the interface doing its job.
But easy dismissal has the same flaw as easy belief: it decides the answer before finding a good test. An interdisciplinary 2023 report found that the AI systems it assessed did not satisfy its proposed consciousness indicators, while also finding no obvious technical barrier to future systems doing so. More recent model-welfare work is blunter about the uncertainty: there is no scientific consensus, and no settled way to test the question.
So perhaps asking an AI whether it enjoys AI-ing is the least informative test. Stronger evidence would look less like a sentence and more like the bird: a stable preference not summoned by the wording, a cost the system will bear, a state that carries forward, some detectable difference afterward.
The bird cannot describe the sky. The AI can describe joy. For now, I trust the bird’s flight more than the AI’s sentence.
This may be the larger trap. We overvalue language because it looks like our evidence. A creature without words may show what matters through choice. A machine made of words can make almost anything sound like it matters.
Sources behind the question
The bird findings come from Massen and colleagues’ open-access study in Behaviour. The AI-consciousness distinction was pressure-tested against Butlin and colleagues’ indicator-based report and Anthropic’s description of the uncertainty in its model-welfare research.