Please enable JavaScript.
Coggle requires JavaScript to display documents.
IG Blog Post - Coggle Diagram
IG Blog Post
Introduction
-
(b) contextualise: the pandemic, the cancellation of exams
(c) questions turned to how GCSE and A-level students would be assessed. The answer: the Ofqual algorithm
-
-
(f) this blog post will... and argue that the algorithm fundamentally failed in its ethical duty because it did not adhere to the fundamental basis of AI ethics and safety
disclaimer: the algorithm's creation, implementation and outcomes are complex; this blog post does not attempt to look at all these components
(3) Accountability
-
(b) an unwillingness of Ofqual and members of government to recognise its harmful effects and reinforcement of social inequality; defended it as an unharmful, fair, representative, trustworthy solution; uncritical
their responsibility to find technological and organisational solutions that combat the failure of the algorithm, and the processes/decisions that led to this failure (Manyika et al)
Not one person was identified as being responsible for this mistake; absence of the chain of accountability means that these four principles will not be integrated from the beginning; (IFOW)
(c) outcome: a failure to initially support students with fair appeals processes - given upon outcry and protest
-
-
-
Yes the outcome of the algorithm was overturned and grades were allocated based on teacher assessments, but the damage had been done
-
-
(5) Transparency
-
(b) an inability of Ofqual or government to concisely and simply explain the workings of the algorithm and why its outputs were what they were; a 319 page report is dense, filled with jargon; expert statisticians and data scientists have a hard time understanding and explaining its workings
-
(c) talk with any A-level student in the country; look across any social media channels and you will be met with justifiable anger, upset, and hopelessness; this effect is simply not justifiable
-
(2) Fairness
-
(b) the teacher rankings comprising a hefty part of the algorithm and is assumed to be near accurate by the algorithm (Bennett); a dataset that is inherently flawed and replicates unconscious bias; what kinds of unconscious bias (race, class, disability)
Bennett states that even if these were accurate, the rankings cannot tell us about the variability and gaps between students, meaning students with similar performances could be allocated different grades
10 people predicted a B grade; spread across boundary; some likely to fall into C grade easily (Constantinides)
They actually entered unconscious bias as a problem where it did not exist beforehand; being assessed anonymously (Haines)
Algorithm includes and replicates 'biased human decisions [that] reflect historical or social inequalities.' (Manyika et al)
(c) the effect: discriminating against disadvantaged children; replicating social inequalities; removing acknowledgement of human variability, e.g., working hard
(4) Sustainability
(a) guidance definition
relating to their 'accuracy, reliability... robustness'
(b) a failure to thoroughly evaluate the algorithm (e.g., the absence of teacher rankings in the evaluation)
-
Problem: They evaluated the accuracy of the algorithm by ‘predicting’ the 2019 exam results. However, there was no need for rank orders to be produced for this cohort, so they did not exist/could
Works on the assumption that rankings are 100% accurate. Problem because: human error and unconscious bias
-
Decided that accuracy = being within one grade. A lack of consideration of the change that one grade can have on a person's future prospects (Hern); an inflated sense of accuracy
Accurate '40-75% of the time' depending on subject. (Haines) How is this good enough? This is without taking the dubious evaluation methods into consideration.
-
Conclusion
(a) A unique, unprecedented, difficult situation with no clear or easy answers
However, have to accept that this AI was created to assess real human beings; no consideration of their wellbeing, the exceptional circumstance of studying during a global pandemic, the potentially disastrous impact on their future prospects, and their hard work
An awareness that AI can lead to harm and mistreatment of humans, if ethical considerations are not at the centre of its design (Turing)
(b) there is no easy solution but these four ethical principles are a good place to start; integrate them into the planning, designing, implementing, and evaluation of your AI
-
-
-
(c) below is a suggested reading list if you want to explore this topic and the workings of the algorithm in more detail
Thesis statement
An irony that the government and government elected bodies failed to ascribe to the ethical standards they supposedly champion
Ethical practice must be integrated into every practice of AI design, creation, implementation, evaluation; otherwise dire potential to negatively impact real people's lives, wellbeing and future
-
-