Taking a different tack, we can ask why this question is different from "When can we say that machines produce heat?" or "When can we say that machines generate rotational force?".
While most people accept consciousness as being real, what makes this question hard to answer is that, in contrast with most real things (e.g. heat and force) consciousness is undetectable. Therefore we have to resort to methods such as those presented above (observable functional products of consciousness or logical reasoning) to deduce the existence or absence of consciousness. The problem with these approaches is that the undetectability of consciousness makes it impossible to obtain empirical evidence of their validity.
Now let's try a different question: "When can we say that machines store binary values?"
While this may seem to belong to the second category, it actually belongs to the first. We can detect magnetic bit states, but we can only deduce the binary values associated with them on the basis of knowledge of the design of the computer. We cannot detect the binary values themselves with any physical device.
This applies even more strongly to the higher level functions of digital computers such as storage of images or computer programs. In these examples specific coding systems are used to allow these higher level information phenomena to be created, and also (through decoding) to be reconstructed on the basis of detectable bit sequences for reproduction (for instance) on the computer monitor. However, give only a bit sequence without knowing the coding system used it is impossible to detect the image or program. And yet they certainly exist, because otherwise they couldn't be reconstructed.
The problem with all high-level brain functions (including consciousness) is that they are not created using formal coding systems that could be used to decode them, but like all living systems arise though self-organizational processes. Indeed, even the basic information associated with synapses and neural events is undetectable, only deducible on the basis of functional output. It is for this reason that the question "When can we say that brains of other animals are conscious?" is equally difficult to answer.
This issue, and potential approaches to solving it, is discussed by Daniel Boyd (2022).
My own hunch (more than that I cannot claim) is that presuming consciousness in machines on the basis of comparable output to that of conscious brains is akin to presuming that tractors pull ploughs using muscles because that is how horses pull ploughs. Biological brains are so radically different by nature than digital computers that it would seem extremely unlikely that they would use the same methods to generate the same results.
Theoretically if you could create a 100% digital simulation of the brain, then it would be reasonable to assume that it would be conscious. The question is, however, at what level the brain would need to be simulated. That is certainly at a far more detailed level than simply the synaptic connectivity patterns. The simulation would need to include details at least down to the intracellular/molecular level (see Beniaguev, et al, (2021)).
A computer capable of real-time parallel simulation of this number of dynamically interacting components is quantitatively unimaginable on the basis of current technologies. Current Artificial Neural Networks, in any case, fall short by many orders of magnitude.
References
Beniaguev, D., Segev, I., & London, M. (2021). Single cortical neurons as deep artificial neural networks. Neuron, 109(17), 2727-2739. https://doi.org/10.1016/j.neuron.2021.07.002
Boyd, D. (2022). Achieving Transparency in Adaptive Digital Systems. New Explorations: Studies in Culture and Communication, 2(3). https://jps.library.utoronto.ca/index.php/nexj/article/view/39030