Accessibility Washing Won't Fix Broken Systems

Ethics and what you're doing relates to disability , because obviously we talk a lot on Access Chat about disability , inclusion and accessibility etc . So welcome to the show .

What I can say about my journey ?

So my journey into AI ethics began with a background in philosophy , where I developed a critical thinking and an ethical lens that continues to guide my work today , and later I pursued a master's degree in public health focused on bioethics , which gave me a very practical understanding of how ethical dilemmas play out in science , medicine and technology .

Eventually I moved into consulting , where I encountered more agile environments and real challenges that organizations face , particularly around AI governance , accessibility

and bias mitigation . That is what led me to found ATIC to help companies and institutions bring ethics into action , not just a branding exercise or a legal requirement , you know . So to me , ai ethics is not just about complying with the AI Act or the Accessibility Act .

With the AI Act or the Accessibility Act , it is about ensuring that technology serves people fairly , inclusively and responsibly . But I think that this starts by recognizing that technology is not neutral .

It reflects zealous choices and even the absence of those who design it , and I have tried to bring this perspective into projects I have worked on , or from conferences like Collision and Web Summit to talk about co-creating inclusive policies with a range of stakeholders .

You know , because , at the end of the day , ethics it's not about doing what is right , it's about making it work , it's on practice , you know . But one thing is this issue of disability and diversity has always been part of my life . I grew up with an education , embrace , a difference , but also with a new creative awareness of the barriers that surround us .

My grandmother , for instance , part of her vision gradually due to glaucoma , and my wife has keratoconus degenerative high condition , and these and other experiences gave me a very direct understanding of challenges that go far behind physical barriers , incompeting accessibility in healthcare , transport , social integration and , of course , fundamental rights .

By coincidence or not , I started my career in research , always working at the intersection of health , technology and social sciences , and one of the most transformative projects I was involved in was called Intimacy and Disability .

It was there that I began exploring the social model of disability from Michael Oliver and came to face with the structural problems that we have , from the lack of research on hate crimes against disabled people to the absence of medical training tailored to the specific

needs . You know , and one of the main recommendations of that project over a decade ago remains relevant today the need of the expand of rehabilitation services , assistive technologies and support for independent living . And in 2021, . For instance , I have worked with an NPO entirely run by people with multiple disabilities promoting the independent living philosophy .

The everyday reality in Portugal was extremely difficult , from the lack of access to personal assistance and often an affordable to absence of basic , accessible tools . You know .

So I think that you know , like with ai , right , um , there are structural issues with ai , but but you can apply it for either good purposes or ill intent , right ? So I I guess what ? so where in your statement ? So how ? How is it that that you think there is this lack of neutrality in technology itself , or is it in the application ?

That is where ethics and governance frameworks become essential , not to slow things down , but to make sure we are not leaving people behind . Ethics , governance frameworks , is a system of policies , procedures and structures that promote ethical behavior in an organization . Technology is not natural in my perspective , not in action .

And also , when you were mentioning about the difference between ethics and compliance , you know it's usually the ethical thinking that gets you to the point of the formulation of the guidelines and the regulations and so on . So I know that Debra is itching to ask a question , but maybe the last little thing is around . You were talking about bias .

What are some of the things that you've been working on that are addressing those biases ? And then I'll hand over to Deborah .

But what I don't know and when I think about it too much , Lia , it makes me nervous have we gotten better because of everything we've tried to do to raise the awareness about it ? I know we haven't done enough , but have we made progress ? I'd like to think we have , because I'm internally optimistic , but I would just add that to Neil's question too .

We must always keep in mind that multiple institutions operate in this space , from national market of surveillance authorities to the European Commission , to the AI office and the European Data Protection Supervisory , for instance .

On top of that , each country brings its own regulatory and implemented context , which can lead to inconsistencies across the countries , you know .

So , even acknowledging the efforts of the regulators , I still find Disability Forum , for instance , and other advocacy groups , of course , repeatedly emphasized the importance of a human-centered approach , one that truly reflects the motto nothing about us without us .

And , as I learned early in my career , there is a big difference between making policies with people and making policies for people , like technology . Now we look at the intersection between AI , hack and Accessibility Act . We see ethical challenges on operational , legal and social forms .

Although both aim to safeguard , of course , fundamental rights , I think they are not fully aligned . I think this disconnect compromises digital inclusion . But one key operational issue is that AI doesn't explicitly define disability as a risk factor .

Even systems that have a major impact on disabled users , like employment or education , might not be flagged as a high risk , and that is not just a legal oversight , it's an ethical failure , in my view .

I mean without clear incentives or harmonized standards , accessibility can fail off the priority list , not necessarily out of bad intent I'm not saying this but because it's not structurally required . This is how we end up with interfaces like virtual assistants or facial recognition tools that exclude people with disabilities .

But of course , there are also legal complexity , if you usually do these kind of questions . If an AI system causes harm , who is the responsible ? The developer , the provider , the operator ? So this fragmentation can make accountability more difficult .

And let's not forget the Accessibility Act , while robust in many areas , doesn't fully cover emerging AI systems like adaptive machine learning or voice-based platforms . Even AI is used to assess accessibility , for instance scanning websites for WACA , waca compliance . We must consider privacy , transparency and false positives . These are real risks .

It reminds me of accessibility overlays tools that promise to fix everything , but often why deeper , systematic accessibility failures . It's used to call this accessibility washing . I don't know if I answered your question , but what I can say is accessibility can be a last minute fix . You know .

It needs to be embedded from the start , with the direct involve , involvement of people with disabilities across the entire development life cycle . Ethics and accessibility are not two separate concerns if we want ai to serve society fairly . Accessibility must become a core compliance and design principle , not the decorative layer . You know , antonio .

It's not just they don't know where the data is coming from , it's also that the data has an history that can go back 20 years , 30 years , sometimes even more , and it's not really possible to make historical data ethical . It's basically impossible , and a good start is just to admit that .

No , you can't really make data that was collected in the 50s , in the 60s , in the 70s or in the 80s ethical , because at the source there was large bias .

If I could just go to the example of historical data from the United States , from the human rights movement from those days , if we dump that data into the web , of course it's completely unethical and there's no way to fix it . The only way is not to use it . So the dilemma is what are we going to do ? How are we going to accept this ?

Are we going to say this is going to delay technology advancements ? to say is this is going to delay technology advancements . The issue is , how are we going to accommodate all this ?

Because in the end , probably there's a developer looking at the data no idea what the data is about , but also on the line there's an executive say deliver me results , I want to see results . So it's quite a difficult issue because technology is now being completely embedded in our social fabric . This is not about a machine talking with a machine .

It's a machine talking with humans , with effects that are unknown .

I have some suggestions from a ethical point of view , because these systems , like I already said , are not neutral . They operate with the data that reflects historical patterns . Like I already said , are not neutral .

They operate with data that reflects historical patterns , like you are saying , of discrimination and inequality and for this reason , we need ongoing ethical strategies to mitigate these kind of failures .

So , from a theoretical perspective , we could approach this through different philosophical traditions , which would make for an efficient conversation , for sure , but in practical terms , it starts with something very simple including people with disabilities and other underrepresented communities in the design projects , in technology development and also in policymaking At least .

Then , of course , we must talk about data , systems must be trained with diverse and representative data sets across gender , ethnicity and other frequently marginalized factors . But don't get me wrong , but diversity alone is not enough . But don't get me wrong , but diversity alone is not enough .

So we must also actively detect offensive or discriminatory language and prevent the model from amplifying harmful conduct . Sometimes we forget that the scale of the Internet does not reflect the scales of real life , and that is a serious trap , since many communities are underrepresented in online data and therefore risk being misunderstood , misrepresented or ignored .

And so what people are proposing , often when there is this lack of diversity data , is to create synthetic data . Right , but there is also , as one of our former access chat guests , utah troger , and has said , synthetic data is just fake data , right . So how do now ?

Now , there may be use cases where you can create stuff that that can help you shape things , but how do we deal with the ethical issues of making stuff up ? You know , because the synthetic data is just making stuff up .

So what are the ethical questions around making up data in order to try and address some of these historical underrepresentations in the data that's already prevalent on the internet ?

I know that this does not respond to your question , because I can give you an option , an ethical option , that is , stress testing for ethical risks such as bias , exclusion or misclassification , and safeguards to ensure minority perspectives are not diluted in statistical noise . But this is not enough and I know this idea .

I know and you probably also know this is not enough . When it's not possible to fully remove bias from training , data models usually must undergo rigorous audits and regular evaluations and real-world testing with diverse users real users .

We also need tools like customized instructions , the biasing methods such as counterfactual data argumentation and fairness-aware algorithms to detect and fix structural flaws . But all of this depends on one key condition continuous human oversight , and AI systems should not be treated as autonomous or uncustomable entities and this is

a problem and they must remain open to human intervention and contestation . We must ensure that systems learn from human feedback , adapt to the different contexts and evolve over the time to improve inclusiveness . This brings us to the core of AI ethics . It's not enough to recognize the problems .

Of course , we must ensure this transparency , auditability and accountability , and that means explaining how decisions are made , allowing human review and making sure people can challenge automated outcomes Lia , I think that Neil and I must be on the same page today , because I was well , I guess all three of us are but I had a question .

I just was thinking in my head Because once again , I get when it takes , ai takes and it puts things together to make a picture or a graphic . Anyway , I just think we have a lot to explore as humans over the next few times . And Neil said in the chat how can we have human intervention and oversight when the speed of decision is making it so fast ?

And so just to bring that in too , Lia , you definitely need to come back on and talk about this again . But also , I was thinking the question I was going to ask was but at the same time , ai it seems like AI can also AI for good can help us get our hands around the ethics , but we as human beings aren't always ethical .

So it's just such an important conversation , but I'm just so glad we have brilliant people like you working on it , Lia , because it feels a little over my head still .

When we talk about diversity , equity , inclusion and accessibility , AI plays a paradoxical role . In fact , it can either promote accessibility or reinforce exclusion , depending how it's designed and implemented . Personally , I find it difficult to accept that digital accessibility is still seen as a technical requirement when it is , in fact , a matter of social justice .

But this is my personal perspective . I mean , fundamental rights are not negotiable . This is what I think , but I know that I have other people that don't think like me , and this is the problem . This is why an ethical approach to AI must go far behind

regulatory or technical compliance , Because I must foster a permanent conversation about responsibility , social impact , because it is in these gray areas that most ethical failures emerge For me .

For some reason they have a family related who has a ? Disability or if they themselves have disabilities , the lack of visibility . Don't put these topics on people's heads . The other is we still have many people , many events around disability who are basically very close on themselves , people with disabilities talking with people .

Now , everybody knows the problem , that's it . Very few you know the Web Summit technology events . They very rarely approach the topic of disability and inclusion . So it's almost like you don't really see it , so you don't really think why should I take care of a problem that is not really part of something that is in ?

front of me I think this also relates with some of the conversation that we're having here as well . You don't see it ? Why should AI ethics care about visibility ?

But people prefer to say I don't have a lot of data , but this is not a problem . And this is very curious because data goes hand in hand with the issue of accessibility in companies , the business side . We know , according to the World Health Organization , more than 1.3 billion people worldwide with some form of significant disability .

In Portugal , our data from the last 2021 census shows that 10.9% , I think , of resident population aged 5 or over .

We must face these numbers and act with them , not these numbers that we have , because the numbers are a lot bigger and we know this . This is the business side . Just wanted to .

I don't want to say this in a not appropriate way , but the question is innovation whom you know , at the end of the day , entrepreneurs , if technology are not designed with accessibility in mind . We already know risk , reproduction or even number fine exclusion , but when we know about , when we

talk about antonio , about everyone , what we are saying ? When we talk about everyone , I don't mean an abstract or idealized public . I mean real , diverse people with different bodies , with terraces , geographies , lived experiences .

So this is the problem People with disabilities , racialized communities , lgbtq , high A+ individuals , older adults , people with low digital access or digital literacy and those living in rural or under-connected areas . Counting heads .

It's about ensuring that the historic calendar represented in data sets , testing panels and policy rooms are not just visible but actively shaping the outcomes . So for me , inclusion means asking who is missing from the data , who is excluded by design , who is not in the room when AI systems are not being developed , regulated or deployed .

So I think we can agree that representativeness is not a checkbox . It's a kind of a commitment to intersectionality and how we share knowledge , distribute responsibility and build the future .

I need to thank Amazon and MyClearText for keeping us on air , keeping us captioned , and thank you , Lia , for a fascinating conversation , Really thought-provoking stuff here . I look forward to this conversation continuing on social media . And please , just one last thing Tell people , tell our guests how they can find you .

So thank you so much .