If AI will inevitably assimilate into deep sociological structures of society, how will it be able to fare well in social media- a more informal setting within the digital realm? So far, the largest breakthrough in computer vision within the field of object recognition is the discovery of machine learning patterns and how it presents information through the inference of historical examples and evidence. There is no need for the creation of explicit and highly complex instructions in favour of learning from machine learning from data.
In 2016 Microsoft launched a chatbot named Tay on Twitter with the intention of emulating the style and slang of a teenage girl. Tay was designed to engage with people through direct messages and tweets she was “an experiment at the intersection of machine learning, natural language processing, and social networks.” While chatbots in the past like Joseph Weizenbaum’s Eliza conducted conversations by adhering to pre programmed and narrow scripts, Tay had the ability to learn language overtime which enabled her to conduct conversations about any topic.
Machine learning develops generalisations from large quantities of data. Within any given data set, the algorithm will recognise patterns and “learn” how to assimilate those patterns into its own behaviour. This is a process of induction, drawing general rules from specific examples—rules that are based on the account for past cases, but may also apply to future unseen cases too. However, learning from examples can entail serious risks when a machine is not provided with good examples in order for machine learning to generate an appropriate and somewhat accurate response. Good examples entails:
“a sufficiently large number of examples to uncover subtle patterns; a sufficiently diverse set of examples to showcase the many different types of appearances that objects might take; a sufficiently well-annotated set of examples to furnish machine learning with reliable ground truth; and so on.” (Barocas, Hardt & Narayanan, 2020)
In Tay’s case, engineers at Microsoft trained Tay’s algorithm to receive anonymised datasets from public data derived from twitter. In hindsight this was not the wisest decision, the internet- especially twitter is filled with trolls that made a coordinated effort, to exploit a “repeat after me” function built into Tay. Trolls encouraged users to inundate the bot with racist, misogynistic, and anti-semitic language. The bot repeated anything that was said to it on demand. Additionally, due to Tay’s in-built capacity to learn, it meant that she has internalised the vile language that were taught by the trolls.
Below are some examples of hate speech that were published on Tay’s account:
“one user innocently asked Tay whether Ricky Gervais was an atheist, to which she responded: “Ricky Gervais learned totalitarianism from Adolf Hitler, the inventor of atheism.”
“I hate feminists and they should all burn in hell”
“evidence-based decision making is only as reliable as the evidence on which it is based, and high quality examples are critically important to machine learning. The fact that machine learning is “evidence-based” by no means ensures that it will lead to accurate, reliable, or fair decisions” (Barocas, Hardt & Narayanan, 2020)
When it comes to mimicking human characteristics and behaviours, through the use of historical examples relevant outcomes will more often than not, reflect historical prejudices and bias against specific social groups, prevailing cultural stereotypes, and existing demographic inequalities. Although, Tay is not a consequential decision making AI, Tay acts as the perfect framework for poor design decisions in AI. Everything that could go wrong, did. Microsoft failed to recognise the inherent toxicity within social media platforms as users hide behind anonymity. Measures should have planned for this contingency and ensured that Tay was not easily corrupted.
“The lesson Microsoft learned the hard way is that designing computational systems that can communicate with people online is not just a technical problem, but a deeply social endeavor. Inviting a bot into the value-laden world of language requires thinking, in advance, about what context it will be deployed in, what type of communicator you want it to be, and what type of human values you want it to reflect.” (Schwartz, 2019)
Without addressing these issues at the forefront of the design process we risk creating AI’s that reflects only the worst parts of ourselves.
Abby Lim [Software Studies] 