The pandemic, social benefits, and automated decision making (ADM): Just because it is quicker to use a machine, is it consistent with the principle of non-discrimination?

bridge over river in city

Introduction

On one level, it seems a very worrying idea that machines alone could be left to make important governmental decisions about our business, our personal finances or security.  No doubt, if those decisions “go our way”, we probably won’t mind too much about the process itself since quick welcome decisions will always be applauded.  But how do we feel when decisions are made about us, by a machine, which are to our detriment or might relate to protected characteristics like sex or race?  Are we missing something if we too readily accept decisions by machine simply because they have the advantage of being quick?

In this blog we discuss the recent Admin Court judgment in  R (o.t.a The Motherhood Plan) v Her Majesty’s Treasury[1] (Motherhood) which held that the Government could justify a potentially discriminatory benefits systems based on ADM as being legitimate because it was  quicker, cheaper and allegedly more straight-forward. 

It is possible that the conclusion in Motherhood is a “one off” and a legitimate response to the emergency needs of government in the pandemic; but, if not, and Motherhood is seen as having a wider application, it seems to open the door to more and more government by machine. 

So, what was Motherhood all about? 

On 30 April 2020, just as the pandemic was taking a hold, the Treasury introduced the Self Employment Income Support Scheme (SEISS).[2]  SEISS was meant to be a life-saver for the self-employed whose business had crashed by offering them a taxable grant.  To be eligible for the grant the Treasury required claimants to show that they were adversely affected “by the health, social and economic emergency in the United Kingdom resulting from the coronavirus and coronavirus disease”.

The Treasury initially decided to make a grant of 80% of a business’ average trading profits (ATP) as calculated over the preceding three full tax years (2016/17, 2017/18 & 2018/19) capped at £2,500 per month for a three-month period.  In designing the SEISS in this way the Treasury was not seeking to eliminate entirely the financial loss suffered by a self-employed person, but to “soften the blow” (our words) of the pandemic by reference to a prediction of how successful the business would otherwise have been. 

The Court described the Treasury’s aim in paragraph 50 as follows:


The quantum of payments under the Scheme had to be based on some proxy. This was not a flat-rate payment scheme, but a scheme which based the amount of payment on the established profitability of the business, up to a maximum figure. The aim of the Scheme was, by making payments in this way, to compensate for business profits forgone as a result of the pandemic. The proxy chosen was an average of the last three years’ trading profits, with adjustments for those traders who could not show three years’ worth of trading profits.

In the urgent need to keep such businesses going, the government could not sensibly ask all self-employed people to fill in complicated application forms each of which would then be laboriously reviewed by a human to decide the “right” level of financial assistance.  Doing it that way would obviously take months just when people needed immediate financial support.  So we think everyone would agree it was sensible to have a simple formula to calculate the grant using some proxy for the profit lost through the pandemic.

So far so good, but what was the “devil in the detail”?

You may have guessed from the case’s title that some new mothers had much to complain about when it came to the SEISS scheme.  The scheme meant that businesswomen who had been on maternity leave in the past three years, would likely have a lower ATP than those self-employed of either sex who had been able to continue to work through.  For them, past profitability was not an accurate proxy for current profitability, yet the SEISS lacked either a specific modification for them or a role for a human being to over-ride this machine-based assessment of past tax returns where the proxy was inaccurate.

So the Claimants brought a challenge by way of judicial review in the Admin Court arguing that Article 14 read with Article 1 of Protocol 1 of the European Human Convention on Human Rights (ECHR) had been breached in that discrimination against women on maternity leave had occurred and that there was a breach of the Public Sector Equality Duty (PSED) contained in s.149 of the Equality Act 2010 (paragraph 3).

First the Claimants’ case before the Admin Court was that using the ATP over three years to calculate the level of payment was unjustifiable indirect discrimination because reliance on that method of calculating the taxable grant to be paid under the SEISS was a provision, criterion or practice (PCP) which placed women at a disadvantage in comparison to men (paragraph 35).

The Court noted a typical case study at paragraph 35:


The Second Claimant’s position illustrates the problem: she was awarded £1,119 per month under the Scheme. This was based on the Scheme formula, taking account of her trading profits for the three relevant accounting years during which she had been in business on a self-employed basis. Her trading profits for those years were: £2,400 in 2016/17 (the year in which she commenced her business, so reflecting only part of a year’s work), £8,600 in 2017/18 (during which year she took maternity leave) and £5,781 in 2018/19 (during which year she took maternity leave). She suggests that, if she had not been on maternity leave but had been able to continue to earn at similar rates, she would have received a payment of around £3,996 per month.

The Admin Court rejected this argument on a wholesale basis.  First it held that the argument failed from the start on the basis that the reliance on the ATP over the preceding three full tax years could not be a PCP since the disadvantage to women who had been on maternity leave arose “from an absence of or reduction in a person’s income in the past” rather than the SEISS scheme itself (paragraph 62 and also paragraph 64). 

We are not persuaded that this is the right analysis; a different scheme, for example, one that excluded years when a woman was not on maternity leave, would have achieved a different result. Moreover, it neuters discrimination law to the point of meaningless to say that a PCP does not exist where the disadvantage is caused by the protected characteristic.

It seems to us that there is a credible argument that the formula adopted by HM Treasury, which used a data point that related to previous earnings in set years, could be a PCP.  

Equally, we are not persuaded by the Admin Court’sanalysis of the Claimants’ alternative argument concerning direct discrimination under the ECHR.  The Claimants’ alternative argument was that women on maternity leave were subject to direct discrimination (referred to as Thlimmenos discrimination throughout the judgment). They argued that women on maternity leave were treated to their detriment since they would receive a lower grant if their businesses’ profits had been supressed due to periods of maternity leave in the past 3 years even though by the time of pandemic, they were working at pre maternity leave levels (paragraphs 37 & 54). In other words, supressed maternity leave pay was being taken into account when it was no longer relevant since the purpose of the grant was to allow traders to meet costs in the year of the pandemic itself as opposed to previous years (see Barry v Midland Bank).

The Defendants argued in response that the Claimants’ claims could only be presented as indirect discrimination claims but since there was no PCP, these claims had to fail (paragraph 56). 

For our part, we can readily see the force in the Claimants’ submission and question the Admin Court’s conclusion that there was no direct discrimination contrary to the ECHR since maternity leave happened in the past (paragraphs 66 & 67) or that maternity leave inevitably leads to lower income (paragraph 67).   

The Admin Court went too far when considering justification

However, our main purpose in this blog is not to explore the meaning of PCPs, the dividing line between direct and indirect discrimination or indeed the interplay between Article 14 and maternity leave. Instead, we wish to focus on the Admin Court’s treatment of justification which we find particularly troubling.

Specifically, the Admin Court repeatedly stresses that the system for calculating the SEISS grants was justifiable because it could be administered by an ADM system which had the advantage of being quick since decisions could be made by a machine (paragraph 49).  Moreover, the system adopted by the HM Treasury allowed an ADM system to focus on known data already collated in databases derived from previous tax returns (paragraph 51). Focusing on known data, even if it excluded information about maternity leave, was simpler (paragraph 84).  It was also cheaper to focus on ADM and known data (paragraph 83).

The Admin Court considered these arguments to be sufficiently compelling that it discounted the Claimants’ argument that a manual scheme could have been introduced for women returning from maternity leave which would have removed the disadvantage to them completely (paragraph 74).  This is where the issue arises, because it seems that much of the learning over many years that a policy that does not allow for special situations is likely to be suspect.

It is important that we pause here and highlight an important factual feature of the SEISS scheme which underpinned the Claimants’ argument on this point.  That is, eligibility for the scheme was changed to accommodate women (and men) who been on parental leave.  Specifically, the SEISS scheme was amended in July 2020 when it was realised that women on maternity leave might struggle to show eligibility for the scheme due to periods of maternity leave in the past.  

To explain further, a person was only eligible for the scheme if they could show that the pandemic had adversely affected them and if they have filed a tax return during the prior three years over which ATP was assessed. Whilst the scheme was being formulated, it was recognised that this could disadvantage women on maternity leave if their maternity leave meant that they had not filed a tax return in a previous year, or their current maternity leave, meant that they could not demonstrate a negative impact from Covid-19 (paragraphs 30 & 32).

These are both problems concerning lack of proper data.  That is, HM Treasury could not tell from their data base if gaps in trade were due to maternity leave or not (paragraph 29).  This problem meant that HM Treasury was advised to amend the eligibility criteria so that anyone on parental leave (which includes maternity leave) in 2018/19 could use the prior year of 2017/18 in order to demonstrate eligibility (paragraphs 31 to 33). It was recognised that such a solution would mean asking people to identify themselves as having been on parental leave and then treating their application outside of the ADM process (paragraph 32). 

This amendment, which was introduced as advised, is described in the following way in the Judgment at paragraph 34:


On 1 July 2020, in line with recommendations in the seventh submission, the Chancellor made a further direction (the “Second Direction”), to introduce a second round of grants under the Scheme (this time based on three months’ worth of 70% of average monthly trading profits), with applications opening on 17 August 2020 and closing on 19 October 2020. The Second Direction extended the Scheme to two specific groups who had not originally been eligible: military reservists who had been on military service in 2018/19 and persons who had not qualified because of the effects of pregnancy, maternity or childcare in 2018/19 (meaning that they had not filed self-assessment tax returns for 2018/19 or had not met the profits condition for that year). The method of calculating the benefit for those who had become parents in 2018/19 required the submission of alternative evidence to support the claim, in the manner described by Mr Hacon in his witness statement (see paragraph 23 above).

The evidence given by Mr Hacon summarised at paragraph 23 in the Judgment is also important:


Mr Hacon described the varied process used for 2018/19 parents, who become eligible under the Second Direction. There were only around 421 claims from this group, although HMRC thought that the group itself might have comprised as many as 10,000. Each claim from this group was risk assessed (an automated process), but any claims which were considered high risk had to be subject to a manual verification process. The information required from these claimants was different and more extensive, and their claims took longer to complete.  

This amendment to the SEISS scheme was absolutely called for and would have ensured that no discrimination arose in relation to eligibility as initially feared.  To the extent – and it is not entirely clear from the judgment – women on maternity leave were rejected under the first scheme but then accepted under the second scheme, they have likely experienced discrimination since the grants under the second scheme were less generous than the first.

Despite the amendment to the scheme in July 2020, the Defendants argued that maternity leave could not be taken into account when determining the level of grant awarded to women who returned from maternity leave (paragraph 75). They argued that introducing an additional manual process in relation to the level of the award would lead to cost and delay and that a “more sophisticated IT system” would be required (paragraph 76).  There does not appear to have been explanation as to why a manual scheme could work when it came to eligibility but not the level of the grant.

Despite this, the Admin Court accepted the Defendants’ argument concluding, amongst other matters, that the scheme was justifiable even if discrimination did arise (whether that was direct or indirect) because, as explained at paragraph 77, a manual system would cost more:


I accept that a move away from a method of calculation based on actual profits, reflected in data collected and held by HMRC, would have involved expense and led to delay, both of which were antithetical to the required quick delivery of the policy. The Defendant had good reason for adopting an approach that was simple and quick, which used one rule, one approach, applicable to all. IT overheads and manual intervention was kept to a minimum so that the Scheme could be implemented quickly. This was not unreasonable.  

We think this analysis must be wrong. 

The Treasury were able and indeed did amend the scheme when it was realised that women on maternity leave could be disadvantaged by the eligibility criteria. It did so even though that meant departing from a fully automated system and using manual verification which was more extensive and time consuming (Judgment, paragraph 23).

A related problem is the Admin Court’s suggestion that simply because a fully automated system is cheaper and quicker it can justify discrimination.  We accept that there will be circumstances in which the advantages of ADM might justify infringements to the principle of non-discrimination but as illustrated by the changes to the eligibility for the SEISS introduced on 1 July 2020, it will often be possible to create parallel, human led mechanisms which ensure that protected groups are not placed at a disadvantage whilst still achieving the scheme’s aim without a prohibitive level of cost or complexity.

The Admin Court’s analysis also ignores the wealth of guidance, including from the UK government to the effect that ADM must be human centric due to the risk of, amongst other matters, discrimination. To learn more about this guidance, please see here.

Further, the Admin Court concluded in paragraph 79 that HM Treasury could justify using ADM because it meant processing known data held in existing data base, which was easier than gathering additional, relevant data about maternity:


The desire for claims to be verifiable by reference to data already held by HMRC is a powerful justification for the design of the Scheme. It meant that the claims could be automated, which achieved speed and cost savings, and the risk of fraud was reduced. The concern about fraud is canvassed in the ministerial submissions right from the first, and the Defendant’s evidence confirms that it was at the heart of the Scheme’s design from the outset. I accept that evidence. The Claimants’ suggested alternative, based on a woman self-certifying that she had been on maternity leave for certain periods, would in the Defendant’s view have been “wide open to fraud”, because such certifications could not realistically have been checked and for that reason would not have been acceptable. That is an understandable position to take.  

This is a worrying conclusion because it ignores the inadequacy of the data held by HM Treasury.  Its data could not differentiate between traders who had been on maternity leave.  This created the disadvantage which contributed to the problem about which the Claimants’ complained and yet that limitation in the data set is then used as a means of justification.  The circularity to this analysis is flawed.  Again, it also ignores the wealth of material which recognises that for ADM and AI systems to be ethical and non-discriminatory, data sets must be inclusive. To learn more about this guidance, please see here.

Conclusion

ADM has a place in government decision making – it can introduce speed and accuracy which at a time of pandemic are important aims. But we must not assume that the advantages of ADM allow organisations to sidestep the more complex needs of protected groups.  Critical analysis is always needed which includes a proper consideration of the extent to which ADM systems can be sensibly modified to meet the needs of protected groups.  Our view is that the Admin Court failed to properly consider the Defendants’ justification defence because it too readily accepted arguments that ADM was useful.

A final word on the PSED

The Claimants also failed in their argument that there was breach of the PSED.  In short, since there was evidence that HM Treasury had considered the impact on women returning from maternity leave, as set out extensively in the judgment, it could not be said that the PSED had been breached since it is a procedural duty to consider rather than an obligation to actually eliminate discrimination (paragraphs 86 to 88).  This analysis is a reminder of the limitations within the PSED; it will not always be a meaningful tool to challenge the ways in which ADM and data collection/processing by government can impact negatively on protected groups, a sharper stick is required.

19 April 2021


[1] [2021] EWHC 309.

[2] The SEISS has been updated a number of times; its current iteration can be found here.

One thought on “The pandemic, social benefits, and automated decision making (ADM): Just because it is quicker to use a machine, is it consistent with the principle of non-discrimination?

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.