Building public confidence in data-driven systems

Show video

all right hello everyone um welcome to today's  events great to have you all here a few people are   still filing in so before we dive in with formal  introductions i'm just going to give a quick few   bits of housekeeping for all of us to follow today  um the first is that you'll find the chat and q a   functions the bottom of your screen for those of  you who have never used zoom before which i would   anticipate is not many um if you would like to  ask a question during the q a please do use the q   a function below and if you'd like to introduce  yourself to the room and just share a bit about   why you're here today please do so in the chat  we'd love to get the conversation going there as   well uh you can also engage in the conversation  on twitter if you are a social media influencer   our twitter handle is at ada lovelace inst and  the uk stats regulation uh twitter handle is at   statsregulation there are closed captions  available you can turn on the subtitles by   clicking the cc closed caption button at the  bottom of your screen and alternatively we can   view you can do fully adjustable subtitles via  the stream text which will be dropped in the chat   shortly great okay i feel like most of us are here  so welcome everyone to this conversation today   on building public confidence in data-driven  systems with the ada lovelies institute   and the office for statistics regulation my name  is andrew straits i'm the associate director   of research partnerships here at the ada loves  institute those of us who don't know ada we are an   independent research and deliberate body based in  london with a mission to ensure that data and ai   work for society in march 2020 education ministers  in england scotland wales northern ireland   uh closed schools as part of the uk's response  to cobit 19 outbreak government announcements   confirmed that public examinations would not  take place in summer 2020 and the uk's for   qualification regulators off qual the scottish  qualifications authority qualifications wales and   the council for the curriculum examinations and  assessment in northern ireland were directed to   develop an approach to awarding graves grades in  the absence of exams while each qualification body   adopted a different approach all of them involved  the use of statistical algorithms to assess and   award grades now i personally just so happened  to be in westminster on the morning of august 6   2020 a day when students from across the uk march  through parliament square and to number 10 downing   street chanting and now infamous cry which i  will not repeat here about those algorithms   in the end the algorithmically determined scores  were drops and the grades in all four countries   were reissued based on the grades schools  and colleges had originally submitted   the off call of qual a-level exam results  algorithm prompted wide societal debate about   the conditions that would engender public  trust and confidence in data-driven systems   in this event today the office for statistics  regulation will share findings from their uk   wide review of the exam models deployed in 2020  focusing on the importance of confidence and   models specific factors that impacted  on confidence in the a-level algorithm   and drawing on lessons for the future the office  for statistics regulation is a key uk regulator   charge with increasing public confidence in the  trustworthiness quality and value statistics by   articulating standards and shaping good practice  the ada levels institute will also share some   insights today from a diverse profile of public  engagement and deliberation work that we undertook   during 2020 on the conditions that engender  public confidence in data and related technologies   at times of public health emergency and beyond  i'm very excited to say that we're joined by four   wonderful speakers today first election truce  gail rankin who leads osr's program of systemic   reviews for the uk and also heads up its office  in edinburgh gail's background is in engineering   and project management and she joined osr from a  career in local government data and performance my colleague riva vatel is also with us here today  she is ada's associate director of engagement   rima has worked for the organization from its  establishment as part of its founding team   she leads aida's public attitudes  in public deliberation research   and its broader engagement work on justice and  equalities ed humphrison is the director general   for regulation with the office for statistics  and regulation and along with michael hodge   the head of data and automation at the office for  statistical regulation they'll be joining us for   a q a panel at the end of the day we'll start  from some presentations from gail and rima on   the findings from the research into off qual so  i'll hand it now to gail rankin to kick us off i think we will be having a few difficulties uh  connecting gail we should be with her she'll be   with us shortly would it help uh andrew if i  just is that humpherson here i just said a few   words whilst uh gail tries to reconnect rather  than have everybody sitting around awkwardly   so first of all let me just introduce us um as the  uh as the uh osr so we are the regulatory arm of   the uk statistics authority and our formal  statutory role is to promote and safeguard   production and publication of official  statistics one crucial point that   i always make in our presentations  is that we don't actually produce   any statistics and that we're actually quite  separate from the ons and a big chunk of our work   is overseeing regulating if you will the work  of ons and those of you who are interested in   demography will know that we published quite  a critical report about ons on monday of this   week about how they compile uh population uh  estimates um for some parts of the country   so we're separate from ons and we essentially hold  all parts of government to account for the way   they produce and publish statistics and data and  you see on the right hand side of the screen there   some of the tools we use the main  one actually is the code of practice   for statistics um that's our kind of uh it's  got a sort of like foundational role for us it's   absolutely central set of principles that we go to  our and then we do a range of other things compare   the national statistics designation we do reviews  across whole systems like adult social care   and we um are known for stepping into the public  domain when we have concerns about the use of data   and there's a screenshot there from a headline  uk statistics watchdog warns government over   the use of covet data so that's us i can see  from my screen we've got gail back at gail are   you back with us i am i'm very sorry i was having  connection problems then there ed thanks for doing   that introduction on us and just to take over  we today are going to talk a bit more and about   the review of the exams process that we published  in march of this year and with this review we've   explored the approaches taken to awarding grades  last summer and we did this to identify the wider   lessons for those working with statistical  models and algorithms and i should say just   at the um start as well we really enjoyed working  with the qualifications regulars some regulators   and most of them we hadn't really worked closely  with before and we found that they all acted with   honesty and integrity in what was very difficult  circumstances and can we move to the next slide   please thank you so um as you'll all probably be  aware and last year because of coronavirus and uk   exams were cancelled and instead and grades were  calculated using a mixture of teacher assessed   grades but also as we all know statistical models  were used as well to standardize the results   and we also headlines in the papers come  august and statistics have ruined my life and   we also started at this time to hear phrases such  as mutant algorithm being used and they were being   used as a way to apportion blame and and basically  the statistical modeling decisions that were made   last summer had a very fundamental impact and for  some children a lasting impact and it's quite hard   particularly us as a regulator to think of another  time when statistical models have had such public   impact before and we probably you know most people  you know us included as the public we never really   thought about them before or particularly felt  that we had felt that we hadn't been affected   by them and also what's important to note is  our review covered all four countries of the   uk and there wasn't while there's been focus on  off calls model and there wasn't one just wasn't   just one statistical model and the approaches  actually adopted in each of the countries did   have similarities but we're also different but i  think the thing that we want to talk to you more   about today is that none were able to succeed  in securing public confidence and if we just go   back one slide dreamer we've just jumped ahead  a little bit there's one before this slide um   which just explains a little bit of the context  and of why we decided to undertake this review so   all of you who are working in this area will  know that statistical models and algorithms are   increasingly becoming part of public life and we  really believe this technology and availability of   data increases that there can be very big benefits  in the public sector using these kind of models   but it is quite clear to us that there has been a  negative rhetoric has developed around algorithms   and why we've got involved is our role at osr  is to uphold confidence in statistics so it's   very clear to us that um public confidence in  statistical models and algorithms has been damaged   by the experiences last year and our worry is that  people will be less like less and likely to use   these kind of models for fear of a public backlash  and so that's that's the main reason why we looked   in detail at the approach is it wasn't to pull  apart to judge what others have done but it's   basically to get a much better understanding and  of what caused things go wrong and really so that   everybody can learn from it and it's worth just  flagging on this slide the last bit of the slide   at the heart of what we do is the code of  practice and the code of practice is all   about ensuring that statistics serve the public  goods and what this means that when we when we   do our regulatory review it means that it's a  balanced one so our reviews our approaches focus   on the technical issues but um quite importantly  we also consider that the public you know the   broader context so the public dialogue public  understanding and public acceptance and it's   this kind of socio-technical approach that  you'll see across a lot of the work that we do   it's built into the code of practice and we feel  it's allowed us to contribute to quite a unique   position on this debate so next slide please reema  so i just wanted to spend a little bit of time on   exploring the sort of the key factors that  we found through our review and that um   impacted on public confidence so the first one is  around confidence in the model so all models have   limitations and uncertainty and i think if we take  ourselves back to last year it was a unique time   and resources there was resource constraints  there were a lot of challenge but we felt when   we looked there was a high confidence placed  in the model that that a model could deliver   a single grade to all students on a single  day whilst also not disadvantaging any groups   um in addition the next one is around um  transparency of the model and its limitations so   the regulators took lots of um lots of  communication effort um but and there was   what we felt limited transparency so full  details as we all know when reading in the press   of the model weren't published in advance but  one of the things around that was a desire   not to cause uncertainty in this new time for  students who were going through this new process   but in our view the limitations of the  general limitations of statistical models   and also uncertainty in the results  of them weren't fully communicated   to the public and possibly had there been more  public discussion of limitations and also most   importantly the mechanisms that were being  used to overcome the limitations of the model   such as the appeals process it may have helped  to support public confidence in the results   and then technical challenge  so we saw lots of collaboration   um in in terms of technical challenge and  this was largely with the qualifications   and education context experts in those fields  there was more limited um statistical challenge   in the wider field and this is where we heard  possibly some dissenting voices come through   the methods weren't exposed to the widest  possible audience but we fully do acknowledge and   in particular the report that there were time  constraints to doing this in this particular   case the other factor in to consider is the  impact of historical data all four countries had   different models but all four countries used the  previous grades at schools and centers as an input   and within um this historic attainment there  are patterns so well-known patterns um between   um how different groups the attainment between  different groups and because these were used in   the model these fed through to the results and  this created a public perception that the model   had created bias and there was limited public  discussion and limited public awareness in advance   of these underlying patterns in the data and  also how they might impact on the results the   next factor was quality assurance we saw lots of  really good examples of good quality assurance   of input data and output data and but what  we saw was largely at an aggregate level   and with limited quality assurance of individual  results and so and this meant that when   the results were published and the media focus  which was on these individual individual cases so   particularly on individuals where their grades  were substantially different from the teachers   we felt that there was possibly  that that could have been predicted   um in earlier in the model development and next  is public engagement and testing and there was   a wide range of public engagement and testing but  what we found when we looked into it that testing   largely focused on the processes of calculating  the grades and rather than the impact that was   going to have um and this um one thing that came  out for instance as you know scotland came first   um and it was only after scotland um issued  its results people felt that they were able   to see themselves in in the process and what  might happen to themselves or their child   and next one is broad understanding um the  in any normal year statistical evidence and   also expert judgment are used to set grades so we  grades you know pass rates go up and down and but   there was limited understanding of what happens  in a normal year in this case and this result   this resulted in this perception that things were  possibly more novel than they were when actually   there is statistics in underlying a process  normally and the last one is human in the loop and   humanly can mean lots of things to lots of people  and what we saw was we saw clear human involvement   in setting the model the parameters the coding and  human involvement and checking the results and and   some and feeding back the parameters but what we  found was the models tended to make the decisions   rather than to support decisions in this in this  case and it could be that there had there been um   more involvements in the final in more involvement  of human humans in the final grade this may have   improved com public confidence in the output  so that that was the key factors we find from   specifically looking at the exams and come to the  next slide please what's really important though   in this the main point of doing our review was  what does this actually mean for those who are   developing models and we find three principles um  that support public confidence and in our report   underneath each of these three principles  there's actually 40 learning points for others   and looking to work in this way so the first  principle is around be open and trustworthy   so it's about ensuring transparency about the  aims of the model and also the model itself so so   that touches on the limitations being open to and  acting on feedback and ensuring the use of models   is ethical and legal and the second um key  point is being rigorous and ensuring quality   to rights so that's around ensuring  clear governance and accountability   bringing in those subject matter experts and  technical experts when developing the model   and also making sure the data and the  outputs of the model are quality assured   at the level they're going to be used at and  then lastly is around meeting the need and   providing and public value so that's around  engaging with commissioners of the model   making sure and considering that a model is  actually the right approach and also testing the   acceptability of the model and for those of you  that know the code of practice very well you might   recognize tqv trustworthiness quality and value  in these but what is interesting is we actually   had a bit of an open book when we did this um and  actually this was a bottom-up approach but it just   goes to show how well and the code does stand  up for cases like this i'm just going to move on   nema and so as well as looking at em lessons for  others developing models we also need to consider   the big picture so what we find are some lessons  for the centre of government and we find that um   it's really not always clear when you're working  in this area what guidance is relevant and where   you can go for support and as we all know there's  a lot of different organizations in the space and   this means it sometimes can be confusing to  work in particular if you're if you're new   to model development there's also um a level  of inconsistencies with regards to terminology   and it's also quite hard sometimes to find um out  who's doing what in the space and we also found   that and there was quite limited guidance and  also practical case studies around um public   acceptability and transparency models and we also  we also feel that em there's a lot of really great   support available but it should be a bit clearer  and easier to access and so i'm just going to   close the top by talking to the changes we'd  like to see to happen so next slide please   so what we are what we have um are recommended is  at the highest level we feel that there should be   clearer leadership from government and we're  calling on the analysis function and digital   function but working with the administrations  of scotland wales and northern ireland to ensure   they provide consistent joined up leadership on  these models and to support this we recommend   and that those working this area collaborate and  i know there's a lot of collaboration that goes   on already but particularly around the guidance  area and we see cdei having a role in this and   kind of bringing together um the guidance and  that has developed and also looking at the gaps   so you know possibly in the public acceptability  in space and we also recommend that any public   body who are doing advanced statistical  models which have a high public value   should consult the national statistician and the  resources within within that structure and such   as the data science campus and the um and the  data ethics and support available there as well   so and can i just move on to last slide please  remember i'm just going to finish up by saying   um i thought it'd be really helpful to sign post  some of the other work that we're doing this area   the review of our exams supports our wider work  in ostar is part of our automation and technology   program and we've got michael here and to answer  some questions if you have that at the end   as part of this program we are going to be um  clarifying osr's role and the code when and   statistical models are used and we're currently  just finalizing some guidance about models in ai   in the context of the code and most importantly we  are really looking forward to working and with the   other organizations in this space and embedding  our recommendations of our review so that's it for   me today thank you rumah and we're very happy  with i'm also joined today by ed humphrieson   who's director general of mosr and michael hodge  who leads our automation and technology program   and will be very happy to take questions on  anything um at the end of the session thank you   thank you so much gail i really appreciate it that  was an excellent excellent uh introduction to the   to the report's findings and recommendations  i want to hand it out to my colleague roommate   patel who will walk through some of the  research that we've been doing at data live   institute on participatory forms  of enhancing public confidence   um hi everyone um so i'm incredibly interested  in in some of the findings that have come out   from from this report and found it incredibly  thoughtful and stimulating so i just wanted to   to congratulate and compliment gail on her  presentation and what i'm doing here in this   presentation is just taking a bird's-eye view  of the landscape particularly last year and   building on a lot of our public engagement public  deliberation work um in order to answer this   question what helps us build public confidence in  data-driven systems so there's a real synergy here   and next slide so just a bit about the early  love listen to we use three core approaches   and we build the evidence base to think about um  the the impact of data and technologies on people   in society we also convene diverse voices so it's  that piece of the work piece of a jigsaw that i'll   be speaking to and and these two approaches help  her change and shape policy and practice and so   it's great to be a part of this conversation  and conversation with many of you here today   on that on that matter how do we convene diverse  voices that's a really broad range of approaches   that we we use and we deploy it and we think of  our approaches both participatory and mixed method   and so from the left hand side you can see that  we undertake a range of attitude surveys and   understand people's um attitudes quantitatively  so um here some of our work on on facial   recognition technology but also the impact of  of the pandemic on different groups of people   in society is a really um key example or two  key ways in which we've done that last year we   obviously had to respond to to the constraints and  the questions of a pandemic and um uh prototypes   a new model of deliberating with the public  um through rapid online public liberation so   here you've got a an example of of a convenes  dialogue that we we pulled together and that   was agile rapid online was thinking about um the  uses of data-driven technologies such as contact   tracing and such as uh immunity certificates they  were called then at the time now vaccine passport   and and all of that work has been  written up and published today   there's also been um pre-pandemic but also  we we took these approaches into a hybrid   um during the pandemic stage there's also a  range of uh deliberative approaches that we have   used such as citizen juries citizen council and  public dialogue and and here we think of this as   a longer-term um public liberation that brings  the whole system including policymakers and   practitioners into a room to have a conversation  with um people about technologies what where what   what would be or would not be okay when it comes  to the governance of technology and a key example   of something which we published quite recently  the citizens biometric council and and and last   but absolutely not least there is a really live  conversation about the role that technologies can   uh play in exacerbating inequality but also in  addressing inequalities and and as part of that   we're working very closely with a range of um  underrepresented uh perspectives to understand   the impact that they driven systems have on on  them um so through the biometric council we we   ran workshops for instance on gender identity  and the impact of systems on gender identity   and also on um racialized groups and minoritized  groups where do we see public trust and confidence   being part of this and and the way we think about  this is as a virtuous circle so when we have data   systems underpinned by strong sense of trust and  confidence from everyone and we we know that we've   got certain things right beforehand we we know  that we've got representatives inclusive and   proportionate data and governance um and and and  underpinning that is the mandate for support and   active participation government governing data and  the reason i represented this here as a feedback   loop is because it's incredible it's incredibly  important if that for that to happen because it   then has a knock-on impact on um us being able  to engender fairer more equitable outcomes from   the uses of data system um and then that feeds  into the beneficial impact that i have that then   also impacts in turn on more representative  inclusive and proportionate approaches data and   governance and what this feedback loop highlights  is that quality is not disconnected from this   notion of trust and confidence and neither is  effectiveness that they're really um connecting   i wanted to offer a few reflections on on the  pandemic and hear the source of the data is   from the agent trust thermometer which is  a longitudinal um trust barometer that has   been going for at least 10 years and um so we can  see here that globally the pandemic did impact on   on on trust levels in technology and  um the uk was not exempt from that   so between may 2020 to january 2021 we saw a  12-point drop um compared to a range of other   other countries so this is part of the global  trend that we're we're wrestling with and um the   same trust parameter indicated that willingness um  to share uh a personal data to respond to handling   it declined over that period of time and um there  was a globally a six point decline in in the   percent of people who said they were willing  to give up more of their personal and health   and uh location tracking information to government  than than all that ordinarily is the case   so of course this in some to some extent is about  technology but it isn't just about technology   so um we are beginning to see reports from i'm  referencing here and an off-course study reports   that trust in gcses um and also increasingly  trusting the way teachers award marks have been   uh affected by by the uh exam exam situation the  the fiscal theatre around um the exam scandal and   and and so and you you have a situation where 27  of respondents agree that um gcses are trusted   this year compared to 75 in a normal year and  um so what this highlighted that facial effect   technical impact really um that certainly trust  in technology is an important thing to consider   but also the technology and the trust and  technology impacts upon um the wider system   and how how the wider system is received i  i really loved and i pulled this bit out of   of the study um that gail was speaking about  i really love this particular quotation   um that illustrates the public confidence is  not just about the key technical aspects of   the model neither is it about the quality of the  common strategy but it's about considering public   confidence as part of an end-to-end process  from designing to use the statistical model   through to deploying the statistical model um so  just just a few points that have come from a range   of our public liberation initiative there's four  key points um that that really emerged through the   lockdown debate initiative that was considered  at a time of crisis um we were asking what would   build confidence in crisis the first was about  the evidence base and so how do these technologies   work what impact do they have do we know enough  about that evaluation all of that that's really   important secondly processes to offer independent  review and assessment of the technology so   my colleague and andrew is leading a programme  of work here at the eddy lovelace institute about   how do we ensure that we're more anticipated we're  assessing the impact the technologies have before   um they're used and implemented so what does  participatory impact assessment in this context   look like what is impact assessment that looks at  affected groups and communities here look like um   the third was around boundaries when it comes to  data sharing data use rights and responsibilities   clarity about that people felt this was very  important in the context of the fact that it was a   crisis where there might be less clarity than  ordinarily if okay um and last but not least a   lot of people were really concerned about the  risk of adverse or disproportionate impact on   on on vulnerable groups on minoritized racialized  groups and largely because of questions that they   had around um the the quality of data sets that  were held about different groups and different   people in society so so those were key themes that  emerge um there's also another report that we we   highlighted quite recently that reinforces a lot  of the key themes that um gail has spoken about   and then really about the apps and technology  being judged as part of a system they're embedded   into the neutrality so um always being aware  that technologies aren't always viewed as   neutral so um they're always part of that broader  socio-technical system the um the point that trust   uh isn't just about data or or privacy uh that  technology needs to be effective and and respond   to particular needs i feel like that came out  very much very strongly and echoes um a lot of the   the work that gail's spoken about um recognizing  that data is often about people's experiences   and and expression of identity and and these  are are quite complex they may be fluid so data   systems may not always necessarily comprehensively  capture that or respond to that being aware about   that and um the proactive role that technologies  can play so they can when well designed   thoughtfully designed and help us to proactively  address bias and protects against buyer um and   and those were really some key theme that  emerged and that that's all from us i just   wanted to thank you so much for taking time to  listen i'm really interested to hear what your   questions are um and looking forward to sharing  more shortly thank you so much rima i think some   really interesting points to connect between  the two different threads one being this notion   of human in the loop and a question there being  which humans and which loop um i think there's a   very valuable insight and connection there between  participation and the types of decisions that are   made by these systems um we got a few questions  in the chat already um but if you'd like to add   any more please use the q a function at the  bottom of the screen i'd like to welcome ed   humperson and michael hodge into the conversation  to help answer some of these questions and if i   may i'd like to start with one question that uh  came up which was what measures are necessary   to put in place to ensure more accountable and  transparent uses of algorithms um there was a   point made by gail in your presentation about  transparency being very key and i'm curious if   there's a very specific learnings or practices  that uh um other public service organizations   can put in place to ensure that they're  meeting those kinds of requirements in future   uh thanks so much andrew i'll i'll give an initial  answer and then um invite gail to to kind of   link what i say back into the report  findings if that's okay so i think uh i i   fear andrew that our answer to lots of these  questions is going to revolve around the   same kind of core proposition and the same core  proposition is that if you um break the problem of   deploying sophisticated statistical algorithmic  models down into component bars and pick out   one part and say that's the bit we've got to fix  you're probably going to find that you run into   difficulties so i think if you just say that's  fix transparency of the way the algorithm works   take that out you could come up with some fixes  but i think you might be missing other things   like um the acceptability of the whole process  or um the uh the sort of understanding of the   um of the preceding environment pre the use of  the model what what is the model replacing so   it's for that reason that we'd say actually  the way to think about this is is end to end   when you're moving from one system to another and  the sec the new system has a great component of   statistical modeling or reliance on statistics you  need to think about this notion of the public um   the public kind of engagement with that  throughout so there are some specific   things on transparency but i kind of wouldn't  privilege transparency separate from all of the   other things as as an answer and get gail is is  there anything you'd like to add to yeah just to   add and yes i agree with you ed that and we are  going to be talking about um and probably quite   a lot of our answers the fact it's not about doing  one or two things brilliantly and i hope that came   out clearly in the talk that is very much about  him looking at this full end to end process and   i was just actually having a little look at  our 40 learning points and actually there's   transparency across a lot of them in a lot of the  areas um so yes it's it's one of these things that   just needs to be thought about um in a lot of  steps and a lot of processes along the way very   interesting points i want to take a question  now from the audience and i can see there's a   question here about how do you think the wider  use of data models and data during the pandemic   has impacted on public perceptions around modeling  for example modeling coveted deaths or different   epidemiological models um do you think this has  had any impact on the exam results situation   no okay i'll pass that to you ed and then  i'd love you to come in as well but that's it   yeah so um i suppose my sort of starting  point in answering that andrew is to say   i think we let me put it this way i i have you  know built the last few years going around a lot   of audiences saying you know you know what all  of this statistical analysis it's not just for   elite decision makers it's a public asset it's for  the public uh that the public are both the subject   and the beneficiary of lots of the these models so  we shouldn't forget that these things are for the   public good not for just efficient decision making  um that's just one element of the public and i   always sense when i give this a rather impassioned  like uh advocacy of of of this that my audience's   eyes glaze over or they slightly sort of think  that's a lot that's kind of like a fantasy   you know in the in the real world what matters  is kind of policy makers making policy decisions   and um what i think we've seen in the pandemic  is uh that we're right actually that we're   right that that that uh on um certain profound  issues the public deeply care um not just about   what's happening to them but what how what's  happening to them relates to what's happening   to their broader community and that of course is  the job of statistics is to to give a picture of   the broader community uh the aggregate picture  and so you see you know extraordinary levels   of engagement for example with the coronavirus  dashboard i think i heard from public health   england that they have they're up to well over 20  million hits on that on that dashboard through the   through the course of the pandemic uh you see  in the media uh in in in outlets which would   normally feature graphs and statistics and data  is probably very very prominent uh presentations   of statistics and data because there's an intense  public interest in this and i think the algorithm   story or and i as they say algorithm we always  try and say statistical models to rescue this   story from the kind of pejorative sense that the  algorithm has um when we talk about the the the   exams using statistical models i think we see this  again we see some intense public interest not just   what's happening to me but what's happening to  the community and that's that's the sort of the   the generic agriculture picture so i think that  in a way the pandemic has taught us many many   things about public health and our how we live our  lives and our exposure to sort of exogenous risk   and so on i think it's also taught us uh both the  importance of data and statistics in our kind of   civic life and also the need for humility in the  face of the data i think those two things are   really really crucial uh lessons that that have  learned and i suppose the really interesting thing   for us and i'm not sure if this even answers the  question but i'll i'll i'll i'm on a roll now so   i'm going to say a really interesting question  for us is is this going to be a one-off moment   in in kind of societal history when  suddenly data became important and then   everything went back to that scenario where  people think data for elite decision makers   are not for the general public or is this a  secular shift in interest i believe it's the   latter i believe it's a secular shift and in  some ways part of our work is to sustain that   shift and to get producers and presenters of data  to recognize that that is a sustained shift and to   and to serve it um so yeah there's a lot in that  i'm as you can see i'm i'm quite quite confused   by by the question of linking this all into the  broader pandemic yeah i mean i'm curious how you'd   respond to that a lot of the work you've done i  know has been around this notion of health and   equality is in light of the pandemic i mean how  how do you feel about the uh about this question   around um how the hook of it might change uh  the the context for the understanding of these   challenges well the reason i presented those  slides earlier on the idiom and barometer of   trust is i think it's so interesting to see that  in at the time that um that tech has accelerated   when it comes to use and deployment and used  by everybody more or less um or at least a   rather large significant proportion of population  um at that particular time you you're also seeing   this this challenge around trump and so that  really illustrates um the extent to which which   trust is is really important but there seems to  be a dynamic that's at play in the pandemic that   suggests that that has declined and um so that  is that were that's really worth interrogating   a bit further as as to understanding what it is  that is contributing to that decline the point   about legitimacy is really important as well so  um it's not just that the technology needs to work   well for us but also that people need to feel  it works well for us for it to work well for us   um and uh it i mean examples during the pandemic  illustrate that quite well we we need if if the   contact tracing app is going to be effective  then um there needs to be a sense that we we   feel that it is effective and the wii is a  really interesting point to interrogate which is   who is it that it is working for and this is  where the work on the data divide um comes into   play a lot of our public attitude work recently  demonstrated or illustrated a big um disconnect   between the different demographic groups um who  were benefiting from or downloading apps such   as symptom tracking apps and so on so so that's  a really interesting thing that is a collective   responsibility not just the responsibility of one  person and your ns is doing a really interesting   piece of work through the inclusive data task  force which we're feeding into as well um and   and working to to influence as well thank you very  much i think it's a very good point i want to take   another question from the audience um uh i feel  like i am just picking the announcements for now   but i will get to the rest i can assure you uh  i think this one's coming is quite interesting   there's been a lot of op-eds um claiming there  was no fair way of fully automating automizing   or automating the exam process and building an  algorithm in the first place was was just wrong   i'm curious what the panel's opinions are about  this would there have been a right way to build   this algorithmic model or was the decision to  use a model in the first place inherently flawed   and i noticed a bit of a tricky one um uh ed  gail michael would you like to start us off   i'll i'll give an initial view and then and see  if get gail and michael wants to come in so as   we were doing this work we you know debate raged  internally and with uh we had a very eminent um   expert oversight group featuring uh i think three  past presidents of the royal statistical society   and the debate was were the four bodies across the  uk set an impossible task or were they set merely   an incredibly difficult task that that was really  the question and at times we thought that actually   this was an impossible task because there would  always be the outlier which would trigger a degree   of uh public sense of unfairness and that sense of  unfairness would would sort of generate sufficient   loss of public confidence that it would always it  would always inevitably unravel uh never of course   limitations in the data and nobody ever tried  this before and then the other school of thought   was not impossible merely very difficult and um  that uh there was there was there was a there was   a pathway through this that might have got to um  uh a more acceptable result uh our report lands   in the latter uh without at all downplaying the  enormous difficulties that the organizations faced   and i should say not producing a mission telegram  they did they produced really good technically   uh well-thought-through approaches um there  were ways through that that might have   produced a different outcome um but they're  not downplaying the challenges they faced   but i've been really interested in my colleagues  thoughts on this because as they debate did rage   about this um gail what what are your thoughts  yeah i mean i don't have a huge amount um   yes we did you know so i think what's what seems  what should be very clear from what edwa said is   we didn't go into this review with any sort of  set expectations of what we were going to find   or with any sort of judgment and i'd hope you know  it you know we came across with this balanced view   and so it was very much we just looked at  what was there and we didn't go back and   you know it's not our within our remit to  look within the policy decisions and it was   very much a looking at this at the end you know  the processes that were involved and as ed said   very mindful this was done in a very short space  of time in in difficult circumstances but yeah we   did have those debates as i was outlining there  so oh sorry michael please yeah i'd love to hear   you're obstructing this all right it's just gonna  for me it'd be um it'd be a shame if this was   something that shut the door on the use of  algorithms um to try and get to a fairer society   it i do believe that we can develop our  algorithmic systems to be fairer but i think we   have to look at fairness from a whole perspective  fairness here was thought about in terms of   of the impact which was bad like but also  fairness has to be thought about in terms of   your design of that algorithm and what happens  afterwards and if there was a fair system in   place where people could appeal after these these  type of things have happened then i believe that   is another step into fairness so it's not just  that is that initial result fair but are the steps   we've taken before it and are the steps we're  going to take afterwards are they fair as well   yeah it's a very very good point michael and i  think it raises a challenging question of fair   to whom um uh but i think that point around the  the that focus on a systems approach to fairness   about about the socio-technical concept that this  is a complex system that's integrating into a   complex environment um which raises all kinds of  bias issues along the pathways very important one   to flag okay yeah we're just going to flag that in  one of in when i in the towards the end when i was   summing up what we'd like to see happen this um  notion of public's acceptability which fairness is   tied up into is an area that we see need some work  you know we do need to put some flesh on the bones   and i know people and probably people on this call  are looking at these issues but it is an area that   yeah we think needs to be explored further thank  you rima would you like to come in at this point   uh yeah please i'd love to hear your perspective  um but just very brief i think that there's   certainly lessons to be learned from the way that  this um you know the last year's conversation have   gone that's why we're having this this call  um i'd be reluctant to say that an algorithm   should never be used because i think going  back to that point about the socio-technical   systems it's actually about its use and  its deployment what it's designed to do   the nature of the human interactions with it and  and most importantly the way in which people are   involved in the development and use and the  designs of this of the system um it's fair to   say certainly from our perspective that that there  are we're still as a society working these things   out um i think this report the review um takes us  a step forward in in that direction and but the   crucial point is how do we use 2020 as a learning  opportunity for us to build on so that we can   do what we're here to do which is to ensure that  we have just the bare socio technical system   that's a great point rima actually flows into  i think the next question which was submitted   by peter kemp from kcl which is um thinking  of this year uh which is fair on students   is it to use this year's teacher allocated  grade or last year's cag plus algorithm   uh i'm i'm actually um going to pass that directly  onto gail because i think he probably thought   about this uh a little bit more than i have  um other than to say that i think it's uh it's   definitely a good question to ask um i suppose the  the one little kind of um nuance i place on the   question is i'm not sure whether uh fair is the  is the way that would quite think about it i think   more say which is the one which is more likely  to land in a way that um that is accepted and   acceptable by the public i think that's the  way i think would focus a little bit more on   that public confidence space but gail what what  are your reflections on i think one of them may   i think one of the main reflections i have i would  have between you know last year and this year   and this isn't a fudge answer by the way but it's  the fact that all is brought questions of fairness   to the fore now there has been discussions in  the past around fairness of the existing system   and some of those have come out and and so i think  actually it's a really good thing and because it   brings everything to you know out in the open  so we can have these kind of discussions and   the experts and the education experts and have  these types of discussions because you know at   the heart what we are very concerned around is you  know is the is the mod is the modeling and the use   of the algorithms and how that's come about and  and you know the fact that the dirty word element   um and i think you know models and statistical  models have got an element to play in this   fairness and we will go back to a situation as  well where we are using albeit this year we're not   using algorithms and that's been quite clear in  some of the messaging we will be going back into   a normal situation where algorithms and modeling  are used but they're used behind the scenes and   will that have changed the public perception of  fairness and and will there be different processes   in place i'm not sure so for us definitely as a  regulator you know bringing these bringing these   issues out and having these kind of discussions  is is you know it's it's i it's excellent for us it's a very good point you know i i do wonder if  there's a distinction to draw between fairness and   justice um where when you think of fairness  we sort of think of a particular outcome   uh or a picture of rubric or metric in the system  but where we think of justice perhaps a bit of a   slightly different outcome that we're looking at  and i'm curious if you had any thoughts on that   sort of that that notion of of where fairness  ends and and questions of justice uh pick up yeah it's a really interesting question so um i  i i really don't want to do a crash course in in   political philosophy but but tennis is it concept  that has been fatigued for a few different reasons   i think that that that's a valid point um being  made by by ed um i think in terms of the this   question of justice but the sort of who is it uh  the way we think about it is who is it that is   um benefiting from the user of data driven systems  and technology and thinking about this as not just   making assumptions about the beneficial regroup  but but recognizing that there might be potential   levels of benefiting from the way a technology  has been designed and i'm being aware of these   asymmetries of power that might exist when we're  designing or developing an algorithmic system so   with the exam model obviously um i  mean that seemed quite stark in a way   the people who were out on the street  and um versus the the people who are   involved in designing evolving systems  and what we're really interested in   is how do you uh begin that uh deliberation or  that dialogue uh what that looks like in a system   in order to enable a clear sense of what societal  consensus would look like in in in in designing   these sorts of systems why we use approaches  like the citizens biometrics council or lockdown   debate that notion of when you create a context in  which you bring the whole system together in the   in the room you're much more likely to design  um ai systems that work for people and society   and and design around different standpoints and  perspective and so that's a really long answer but   so that's how we think about just unfair but  it might be quite different to the way other   people think about justin there yeah i think it's  a very good point man i can see the uh the comment   as well in the chat about accountability being  another consideration here really getting to   that notion of not just how you hold these systems  to account but also how do you prevent harm from   occurring in the first place making the more um  uh thinking of impact prior to the harm occurring   um i want to take another question the from the  from the audience which was from hannah spiro   uh and reba i think we'll start this with this  with you i think it touches up on the edelman   survey that you mentioned in your presentation  to what extent do you think we are actually able   to measure and track changes in public trust  towards the public sector use of algorithms   are surveys and focus groups um a good way to do  that or are there other methods we should consider   it's a great question um so it's it's  quite difficult to track and measure   um trust i think but i think that there's  something really interesting about making sort of   making a range of mixed methods approaches to  understanding this this approach so uh some of the   longitudinal surveys are really useful and helpful  and some of the more qualitative pieces of work   are really useful and helpful as well um the  aidan wouldn't trust monsters is incredibly   useful because it looks at the trend across the  international context in the landscape and um   aims to understand uh global trends and i think  that's quite helpful when you're thinking about   um levels of trust uh the the the the other  dimension around trust that's really important   to think about is trusting what and trust in who  so what i find quite interesting through that um   survey is the focus on technology or the focus on  the sector focused on the institution those sorts   of things through that yeah i think you can become  much more granular much more specific about what   it is that you're trying to measure and have if i  try to pull people and ask them about their levels   of trust i think that would give us a very limited  sense of um information about that but if i were   to um be much more contextual about that and in  this case users and applications of data i think   there's more to be said and i'm also much  more around that mixed methods approach so   the public liberation work that we do aims  to um really understand not just the what's   uh so what is it that people are feeling  or thinking but the why what's contributing   about what what what the conditions that are  leading to that um and and get into a much   more advisory space so um how would you you know  suggest or think about what you'd expect or uh   uh uh policymakers or decision makers to do their  their four principles or conditions um which is a   very different thing to the sort of thing that  you get from a car or an attitude server but   i'm very happy to to talk more about that um uh  outside of the scope of this discussion as well   i think it's a very good point i think it  raises this tension we're talking about between   quantitative data that feeds a model and the kind  of rich contextual information that helps explain   the gaps in that data um i want to take uh one  last question before we we might have to wrap up   and this is one for uh everyone on the panel  um i'd love to just go around just gonna do   rapid fire um last few minutes um and that's that  question is kind of going us back to the original   report i'm curious if we could um reflecting on  the discussion today and the report's findings   what measures do you think are necessary to put  in place to make sure that these systems are more   accountable transparent and just going forward  what would you say are your main learnings and   and one key thing you'd like to see organizations  like offcoal put in place in the future should i go first andrew yeah so uh in in a  you know kind of one minute quickfire version   i'd say um and it's not just organizations  like ofqual there are three other regulators   in other parts of the uk and of course many other  public sector organizations they should think   trustworthiness quality and value they should  apply that code of practice anything they're   doing which is relating to the way something is  going to impact on um a public public discourse   or or public uh acceptability think tkv don't  just think technical don't just think quality   think trustworthiness quality and value that's  my one sentence remedy thank you so much ed gail   to go next i would thanks very much andrew and  for me i just wanted to draw one of the the   kind of key things that we raised at the end was  around support so there's lots of people in this   space there's lots of great people in the space  there's an emerging body of of pr of um you know   best practice and actually it's worth saying that  we did have an internal discussion about you know   what are we going to write a best practice report  for this this exam review we're like no we're not   that's for others to do and to kind of um i  think you know if we were an analyst sitting   in the space we want it to be easy and clear and  that support structure to be working for them to   to make that system work thanks gail michael for  me it's um i think we've we've got a lot better   in recent years about collection of data and data  usage and being transparent about that with gdpr   we've been good at the outputs of statistics  recently and showing that end of the spectrum but   when we're missing that middle bit and i think  that was apparent with this work was we didn't   really get to see the inner workings of the  algorithm the model we we couldn't we couldn't   scrutinize it we couldn't learn from it as much  as we probably should have been able to so i   think being more transparent about kind of sharing  our practices in our coding in our algorithms etc   it will allow us to be better  producers of algorithms going forward   thanks so much michonne rima i was also about  to build on the point about transparency but   also transparency for what and for what purpose  so um not just that the algorithm needs to be   um open but also to sort of think about what level  of openness and then what is it that um a subject   to whom the decision it relates to could do about  a decision um would there be an accountability   process or procedure there that's that seems like  a really live and interesting thing to think about   in in in the development so so that's certainly  about transparency but there's some really   interesting work on what does transparency look  like and we were just about to publish an event   um very so i published a report and and and  convened an event very shortly on participatory   data government that said that transparency is  that fundamental building block that enables   more inclusive and a dialogic approach  to the design of these data systems   amazing thank you so much rima i want to  say thank you again to all of our guests   today and all of our attendees for joining a  mass thank you to ed humphrison michael hodge   gail rankin and raymond patel for joining us  if you'd like to continue the conversation you   can on twitter we are at ada lovelace inst  and at statsregulation we'll all see you   there have a lovely rest of your thursday  take care everyone thank you thank you you

2021-05-19

Show video