Like many galaxy-brained people whose parents over-inflated their perception of their own intelligence, I have regrettably ended up with a day job in academia.
It makes sense that academics are enthused about AI – both the Academy and AI systems are in the business of knowledge production. We chase the grants that rich people issue for research that backs up their worldview; hence why I’m so cynical when I’m told to Trust The Science(TM). We publish this newly-produced knowledge in peer-reviewed conferences and journals, and are encouraged to cite other researchers in our fields to further wring out some fresh knowledge from the synthesis. Like a complex AI system, we churn out content – often with hallucinations.
But like AI, the Academy is also an artefact of its time and place. Its mode of knowledge production solidified as an agent of Western empire-building, extolled in contrast to the folk knowledge ways of the ‘savages’ slated for wholesale erasure. No knowledge is out of bounds to The Science(TM) – everything one can know is a mountain to climb up and plant a flag on.
Contrast this with the knowledge systems of those on the other end of the colonial encounter. For many Indigenous communities, their knowledge is something to be safeguarded – passed through generations to the right people, and not to the wrong people. The aim isn’t to drastically change what is known, but rather to ensure accurate transmission, in some cases across tens of thousands of years.
When an Aboriginal Australian studies a rock painting at a sacred site, knowledge has been encoded and expressed within that painting. If you were to train a generative AI on thousands of Aboriginal rock paintings, it might produce something that resembles an Aboriginal rock painting – but it wouldn’t contain any of the knowledge inherent in the original paintings, and might well interfere with these knowledge systems if mistaken for an actual rock painting.
This is, as you may guess, a research interest of mine. The contrast between knowledge production and safeguarding struck me quite recently when looking at AI-generated images of hanukkiahs.
As the reader may know, a hanukkiah is lit by Jews every year during Hanukkah, our Festival of Lights. Hanukkiahs come in many different shapes and sizes, but must contain holders for nine candles – one for each of the eight nights of Hanukkah, plus the shamash which lights the other candles. Like any Jewish holiday, there are many Talmudic opinions about how to light the candles, but Rabbi Hillel proposed the most popular method: one candle plus the shamash lit on the first night, then one candle added for each successive night until the final night, when eight candles plus the shamash are lit. In total, 44 candles are burned over the course of Hanukkah.
But a generative AI system doesn’t know all that, which is why its hanukkiahs range in size from seven to over 20 candles – like those on a geriatric’s birthday cake. When asked to produce a scene of a family lighting a hanukkiah, the important thing for the AI is evidently to produce a profusion of candles; their symbology is less important. Imagine what would happen if an alien species came across these images and mistook them for a genuine celebration of the holiday. What other mistakes about Hanukkah would they take for fact?
This is a quandary faced by many researchers hoping to benefit from a synthesis between AI and Indigenous knowledge systems. First of all, their perspective is often wrong, trying to shoehorn the latter into the former rather than adapting AI systems to suit Indigenous use cases. Second of all, they don’t understand that their default mode of knowledge production itself contrasts with the aim of most Indigenous knowledge systems.
In other words, what good is an AI that can’t even draw a decent hanukkiah?
Sonya Blanck writings 