THE POINTLESS ETHICS OF AI LANGUAGE SOFTWARE

 THE POINTLESS ETHICS OF AI LANGUAGE SOFTWARE

 

The firing/resignation of a Google ‘ethics’ expert has caused a bit of a storm, mainly because the researcher in question is a black woman representing minority concerns in the USA. Most media articles like the one quoted here from Wired focus on this issue, maintaining quite correctly that AI language software ‘can replicate biased language on gender and race found online’. I would replace the word ‘can’ with ‘does’. 

 

Gebru’s departure was set in motion when she collaborated with researchers inside and outside of Google on a paper discussing ethical issues raised by recent advances in AI language software. Supersmart algorithms won't take all the jobs but they are learning faster than ever, doing everything from medical diagnostics to serving up ads. Researchers have made leaps of progress on problems like generating text and answering questions by creating giant machine learning models trained on huge swaths of the online text. Google has said that technology has made its lucrative, eponymous search engine more powerful. But researchers have also shown that creating these more powerful models consumes large amounts of electricity because of the vast computing resources required, and documented how the models can replicate biased language on gender and race found online. Gebru says her draft paper discussed those issues and urged responsible use of the technology, for example by documenting the data used to create language models.

 

As I have noted in a number of blogs on this issue before, algorithmic, statistical matching as the basis of generating text (and associated language tasks) is a regressive enterprise (echoing Chomsky’s dictum on behaviour models of language as being ‘crass and crude’) because it treats language as a closed (but ever expanding) set of data, or as phrased in the quote below ‘by creating giant machine learning models trained on huge swaths of the online text’. Of course a lot of language use is premised on replicating (with possible variations) what has been uttered before but language competence (in Chomsky’s sense) generates an infinite set of utterances (data) via the mechanism of iteration and merge. Given this uniquely human language capacity, language itself (via AI or any other application) cannot be reduced to a computational model. Wittgenstein famously failed to reduce language to a logical, axiomatic set of rules, and instead settling on game theory to account for it. As he employed chess as an analogy, it therefor not surprising that AI has largely solved the chess problem even though on occasion Deep Blue (or whatever the next greatest version is) still loses. Note that chess has a closed set of possible moves however massive in numbers – an increasingly small challenge for number-crunching computers. Now imagine a chess board that is infinite on all four sides – this is what language is: a finite set of rules able to generate an infinite set of moves. So, what has this got to do with the ‘ethics’ reported above? If we merely protest the replication of ‘biased language on gender and race’ in the hope to become less biased if we include all data from, say 2SLGBTQ+, we will surely mitigate the problem but we will not solve it. Racism and sexism exist and will continue to exist as long as the capitalist model of life, as we know it, exists. By working for Google, Ms. Gebru and others merely support the system that creates the biases noted. Not surprisingly, Google does not like the idea of employees biting the hand that feeds them. Corporate capitalism by definition has no ethics. That Ms. Gebru headed a Google department on ethics must be a cruel joke if not a minor tragedy. AI language software, as we know it, is the ultimate business model, automating human speech (and writing) so that when you ring your insurance company to make a (predictable) claim you will be served by a robot that can answer all your questions and at the same time decline your claim, even if the whole process takes 15 minutes or more – note Turing’s famous dictum that when a robot can talk to you for 15 minutes and you did not realise you are talking to a robot, said robot has become human (or at least clever humans have succeeded in computing and replicating human speech). Sadly Turing did not anticipate algorithmic matching. As a member of the 2SLGBTQ+ community he was treated cruelly by homophobics at the time, hence it was a real personal tragedy. Patented AI language software will mean that human language becomes the most valuable corporate asset of all time, and no amount of crying wolf and ethics (the more you cry the less impact it has, as even the ancient Aesop figured out) will stop it. 

 

https://www.wired.com/story/prominent-ai-ethics-researcher-says-google-fired-her/