It’s not often I blog more than once on the same subject but the impact of algorithms on education and social work is central to my academic work, and what’s happened in relation to exam grading appears to exemplify the injustices I’m concerned about. So, these comments add to what I’ve said[1] recently and which have been published by Algorithm Watch[2] in a slightly abridged form.
In recent years I have struggled to become a bit more statistics savvy, as the reality is that algorithms are, in reality, mathematical modelling systems, that use statistics to support prediction. Given I am not a mathematician, what motivated me to do this was a review comment made by Adrian Smith, Director of the Alan Turing Institute, in relation to David Spiegelhalter’s wonderful The Art of Statistics[3]. Smith references a quote frequently attributed to H.G. Wells: “Statistical thinking will one day be as necessary for efficient citizenship as the ability to read and write.” The truth is this quote is from Samuel S. Wilk’s presidential address to the American Statistical Association in 1950, itself a bastardisation of Wells’ quote in his book Mankind in Making[4], which was published in 1903 and read:
The great body of physical science, a great deal of the essential fact of financial science, and endless social and political problems are only accessible and only thinkable to those who have had a sound training in mathematical analysis, and the time may not be very remote when it will be understood that for complete initiation as an efficient citizen of one of the new great complex world-wide States that are now developing, it is as necessary to be able to compute, to think in averages and maxima and minima, as it is now to be able to read and write.
If, as the 1998 Crick Report[5] suggested, citizenship education demands concern for three dimensions: belonging, participation, and political literacy, we might say then statistical literacy is a necessary component of the latter. And, in turn, functioning democracies should make this a priority. Which takes me to my moment of sympathy for Ofqual[6]; the task they were set is not just beyond the vast majority of the population, it was incredibly difficult, perhaps even impossible.
Consider now what any sensible person would do if asked to undertake a challenge of this magnitude; surely it becomes obvious to collaborate with experts but also consult and listen carefully to a wide-range of perspectives – all of which seems in-keeping with even a modest commitment to democracy. That no less than the Royal Statistical Society (RSS) offered to help, and were rejected because they refused to sign a Non-Disclosure Agreement (NDA) smacks of a failure on both counts: to engage people who know what they are talking about and a disregard for a key component of a functioning democracy – transparency. Even though Ofqual dismissed this offer they could at least have listened to the concerns raised in the RSS’s statement[7]. Here was a reasonable provocation to test (or ‘run’) the algorithm to establish its likely effects and anticipate any perverse outcomes. It could be assumed Ofqual didn’t do this, and yet their report[8] includes a table showing some people would have their grades dropped by up to three levels. To go ahead knowing this beggars belief.
The crux of the matter is the abject failure to consider and take into account these effects in human terms, which is to say, on individuals; in the language of battle, to accept ‘collateral damage’ provided the ‘integrity’ (or ‘fairness’) of the wider, national, system was maintained. Where was the constraint in the algorithm to prevent these perverse outcomes, a mechanism to stop individuals’ grades from being adjusted by more than one level; or, as George Constantinides[9] said, by allowing for grade ranges to reflect known uncertainties.
Not doing this has created a range of extraordinary anomalies, one of the strangest being students who were awarded better grades in Further Maths (assumed to be much more challenging) compared with (ordinary, less challenging) Maths, which should surely have triggered an appreciation that something was wrong. Now the algorithm has been disclosed, it seems this is probably due to the fact that these maths students were in a Centre Assessed Grade (CAG)-respecting cohort of less than 15 (where the algorithm was not allowed to function for fears about statistical significance) for the former and a larger group for the latter (where the algorithm did function, including in very large cohorts where the CAGs were disregarded altogether).
So how did all this come about? I hesitate in saying this, but was the composition of Ofqual’s board anything to do with it, were they from private schools and blind to many of the inequalities experienced by many of more modest means? Perhaps they simply didn’t think to ask what would happen to disadvantaged pupils. We know, for example, from important work on the racial biases[10] of many algorithms, or the crippling effects of algorithms designed to identify welfare overpayments[11] in Australia, that the architects of those algorithms may subconsciously, if not consciously, be feeding in bias to these models. Was there a moral distancing, a failure to even imagine the effects on ‘other people’s’ children?
To the credit of some, albeit belatedly, a good number of Ofqual’s board have now come out to say their own algorithm should be rejected. I can’t help thinking this looks a lot like what’s happened with Black Lives Matters; the penny has finally, finally, dropped that there is such a thing as ‘white privilege’. For some at least.
So let’s take stock of the real effects. Steve Chalke, founder of the Oasis Academy Trust, has referred to what he calls a ‘pastoral disaster’, the failure to take into account the human effects of algorithmic decision-making. We might reasonably speculate that this will extend to significant impacts on at least some young people’s mental health, and we know what, in extremis, that can mean.
Not that there won’t be push back from those that evangelise about ‘resilience’, like Matthew Said[12], who, like others, demands ‘rampant grade inflation’ is arrested. Relatedly, there’s the irony that those who trumpet concern for ‘the system’ stand idle whilst observing the “haemorrhaging of public trust”[13] in that system, or are blind to the fact, as ex-Ofsted Chief Michael Wilshaw puts it: “they are losing the dressing room.”[14] And then we have the likes of Jeremy Clarkson who ever year at this time tells us about his own exam failures and how they ‘never held him back’ and how it’s important to cope with failure’. Doesn’t he even comprehend that this year young people didn’t actually sit their exams, that they were labelled failures, by an algorithm?
In combination, the emphasis seems to be on not advantaging young people, making sure they don’t get grades ‘better than they deserve’. It’s as if there is a queue of nay-sayers only too keen to put another foot on the heads of the many that have often experienced such downward pressures throughout their lives.
Here’s another way of looking at it: let’s test the meritocracy we’re told we live in; let that meritocracy run its course. Surely those lacking ‘ability’ will be ‘found out’: they’ll fail subsequent exams, fail probationary years, and reveal themselves at work and so on? Or perhaps they won’t. We might find that what we have here is an interesting and potentially valuable experiment which proves giving opportunities, and further opportunities, to ‘weaker’ students is precisely what ‘levelling up’ means in practice. It’s not as if we can’t handle changes in distribution of exam performance at this time, there are fewer overseas students this year due to the pandemic. Simply remove or adjust the university admissions gap; that would do it.
As of today, the logic is that England should fall in line with Scotland, Northern Ireland, and Wales (even though, by virtue of a different system, the Welsh have more evidence to go on – unlike in England where (as I stated previously) Michael Gove had done his utmost to get rid of the continuous assessment data that would have proved valuable at this time). It should use the CAGs too.
Whilst this might not be a perfect outcome, it will undoubtedly be fairer on individuals and a more transparent process, both values I suggest are essential to maintaining even a basic commitment to democracy. This would also prevent similar problems being wired into an appeals process.
If all this is mitigation, perhaps we can salvage something productive. I hope in future we can learn something about algorithms more broadly, and that there are inevitable negative effects on individuals – real, human beings, who become casualties of algorithmic mismanagement. And I hope we can learn something about democracy, not just how important transparency is, but the essential role of consultation and discussion but also dialogue – which we should remember implies a process we enter into in order to learn. How can we have ‘levelling up’ without that?
Democratisation, then, demands humanisation, and it needs to be applied resolutely to the use of all digital tools. So-called ‘algorithmic development’ and the wider role of technology in society must be subject to ethical questioning. Algorithmic accountability[15] should be more than a goal; it should be a pre-condition of a civilised, equitable and rights-respecting society. With these values we might also make significant gains in the name of education, ridding our systems of an obsession with competition, high-stakes testing, qualification and stratification thereby learning to value what it is really about: learning.
[1] http://www.graemetiffany.co.uk/?p=1592
[2] https://algorithmwatch.org/en/story/uk-algorithmic-grading-gcse/
[3] https://www.penguin.com.au/books/the-art-of-statistics-9780241258767
[4] Wells, H.G. (1903) Mankind in Making, Chapman & Hall: London.
[5] https://www.teachingcitizenship.org.uk/sites/teachingcitizenship.org.uk/files/6123_crick_report_1998_0.pdf
[6] https://www.gov.uk/government/organisations/ofqual
[9] https://constantinides.net/2020/08/15/a-levels-and-gcses-in-2020/
[7] https://rss.org.uk/RSS/media/File-library/News/2020/06082020-RSS-EPAG-statement-on-grade-adjustment-2020-exams-in-UK-FINAL.pdf
[8] https://www.gov.uk/government/publications/awarding-gcse-as-a-levels-in-summer-2020-interim-report
[10] Benjamin, R. (2019) Race after Technology, Polity: Cambridge; and Noble, S.U. (2018) Algorithms of Oppression, New York University Press; New York.
[11] https://www.theguardian.com/australia-news/2020/may/29/robodebt-government-to-repay-470000-unlawful-centrelink-debts-worth-721m
[12] https://www.bbc.co.uk/sounds/play/m000ltkd
[13] https://www.lbc.co.uk/news/a-level-crisis-gcse-results-ofqual-boris-johnson/#:~:text=The%20Telegraph%20said%20some%20Ofqual%20board%20members%20believed,done%20-%20may%20be%20the%20%22least%20bad%20option%22.
[14] https://www.bbc.co.uk/sounds/play/m000ltkd
[15] https://royalsocietypublishing.org/doi/10.1098/rsta.2017.0362
0 Comments