14. Machine Learning and Irresponsible Inference: Morally Assessing the Training Data for Image Recognition Systems

A few clarifications about the relationship between stereotypes and presumptuous judgment may be helpful. First, not all cases of presumptuous judgment involve stereotypes. Stereotypes involve associating an individual with a group (Blum 2004; Beeghly 2015). But it is possible to make a presumptuous judgment without relying on a group association. For instance, I might make a presumptuous judgment about a person’s intentions just on the basis of the assumption that her goals are the same as my own. Second, not all uses of stereotypes involve presumptuous judgments. This is simply because not all stereotypes are about persons’ intentions. Finally, regarding the moral features of stereotypes and presumptuous judgments: Presumptuousness, all else equal, tends to be morally undesirable, but it’s controversial whether this is true of all stereotypes. Beeghly (2015) argues that not all stereotyping is morally objectionable, and Lippmann (1922) saw positive and negative aspects of stereotyping. In contrast, Blum (2004) holds that stereotyping is always morally objectionable to some degree. My contention here, that presumptuous judgments manifest inadequate respect for persons as individuals, is consistent with Beeghly’s explanation of when and how stereotypes fail to respect persons as individuals. However, my thinking about why such a failure of respect is morally objectionable shares more with Blum’s analysis than with Beeghly’s. In the context of the present paper—with its focus on the moral evaluation of training data for machine learning systems—it is enough for my purposes if at least some judgments are morally objectionable precisely because of their presumptuousness.

As Hodosh et al. (2013) point out, “Gricean maxims of relevance and quantity entail that image captions that are written for people usually provide precisely the kind of information that could not be obtained from the image itself, and thus tend to bear only a tenuous relation to what is actually depicted.”

Though crowdwork raises ethical issues of its own (Marvit 2014).

This comes from the online appendix to Hodosh et al. (2013).

This suggests another way to explain what is wrong with presumptuous judgment. To judge a person’s mental states according to a standard like we would use for any other sort of judgment not involving persons, is to take what Peter Strawson (1962) called the “objective attitude” rather than the “participant attitude” toward the person.

Of course, the image recognition system could report the falling water, and we could rely on some other process to infer from the falling water that it must be raining. But this would be to limit too much the capacities of image recognition systems. A scene can be one that looks rainy, and looking rainy may be both more intuitive and more useful information than the report that it looks like water is falling from above.

There’s nothing special about the specific probability values of 0.02 and 0.98, besides the former being small and the latter being large. These values are just convenient for purposes of illustration. Values of 0.01 and 0.99 or 0.05 and 0.95 would have worked just as well (although values that were too extreme or too moderate would indeed alter the examples).

I do not intend this as a criticism of the Flickr 8k data set. Violations of the instruction I am recommending seem to appear only rarely in the data set. However, this image and the next are valuable for illustrating the worry I that is my focus.

I do not mean to imply that Dennett himself is guilty of making this assumption.

Cf. Blum (2004).

And, of course, a further worry about this strategy concerns the thorny issue about how we might go about categorizing intentions as attractive or unattractive in the first place.

Such work is already underway. See, e.g., Park et al. (unpublished ms).

Along these lines, Dennett argues, “the class of indistinguishably satisfactory models of the formal system embodied in [the] internal states [of an entity toward which we might take the intentional stance] gets smaller and smaller as we add such complexities [such as a wider range of behaviors]; the more we add, the richer or more demanding or specific the semantics of the system, until eventually we reach systems for which a unique semantic interpretation is practically (but never in principle) dictated” (1989b). Notoriously, according to both Quine and Davidson, some indeterminacy may be ineliminable. However, along with Dennett, I doubt that any remaining indeterminacy poses any practical or ethical problems in the context of machine learning systems. For discussion of indeterminacy and its (in)significance, see Davidson (1984b).

This is a specific version of the type of problem James Moor (1985) has famously called “invisibility.”