Accueil - Portfolio d'Emmanuelle ABOAF

Transcription de la conférence AI and Disability: Inclusion or Exclusion?

Transcript

Slide 1

Emmanuelle

Hello everyone ! Welcome to our talk “AI and Disability: inclusion or exclusion?”.

Thanh Lan and I are delighted to welcome you and to talk to you about a subject that is close to our hearts. You can scan the transcript or we'll give you the link during the live session.

Slide 2

Emmanuelle

My name is Emmanuelle ABOAF and I'm happy to co-present this talk with Thanh Lan.

I was born deaf with two cochlear implants. I like to say that I'm bionic. I've been a developer since twelve years and I work at Shodo in Paris. Shodo is an IT services company specialized in development, coaching and conferences and committed to social justice. Shodo advocates for greater inclusion in tech.

I am very committed to digital accessibility. I am not an expert in AI but I use automatic tools on a daily basis, which allowed me to analyze the impact of AI on my daily life.

I'm a member of the Duchess France association, which represents women in tech in France.

I'm also a member of the CNCF Deaf and Hard of Hearing initiative, which represents deaf and hard of hearing people in tech around the world.

Thanh Lan

Hi, my name est Thanh Lan DOUBLIER. I am very proud to co-present this conference with Emmanuelle. I am a machine learning engineer, formerly with AXA France. I am volunteer for several French NGOs related to data science and technology ‘Data for Good’ and ‘Lattitudes.’

As a conference speaker, I cover topics such as artificial intelligence (including its legal framework), MLOps, and inclusion.

I am part of the organising team for a conference called ‘Cloud Nord’, which takes place in Lille north of France. Additionally, I am part of a collective women developers called ‘Chtite Dev’.

Due to a medical condition affecting various parts of my body, I am an ambulatory wheelchair user, and I became deaf when I was a teenager.

Slide 3

Thanh Lan

The media and science fiction have given a highly distorted image of artificial intelligence. I believe that if you follow the conf42 conferences, you are already quite familiar with the reality, but we prefer to provide a brief theory recap.

Slide 4

Thanh Lan

For explain what artificial intelligence is, I like to use the example of baking a cake from a cookbook. Imagine you’re following a recipe—this recipe represents your ‘classic’ software.

Starting with the ingredients, you execute several instructions, and in the end, you’ll have a perfect strawberries pie that looks like the one in the photo from your recipe book.

Slide 5

Thanh Lan

But an application with artificial intelligence will work differently: We’ll still start from your recipe book, but in your pantry, you no longer have strawberries; instead, you have apples. Unfortunately, your cookbook doesn’t have any apple pie recipes. However, it’s full of other fruit pie recipes: pear, plum, etc. So, you’ll rely on the various recipes to try to make your apple pie. An AI application is based on mathematics, particularly statistics and probabilities, but the reasoning is the same: we provide it with data—here, all the fruit pie recipes from your recipe book—and it searches for what’s most likely. It’s more likely that the recipe includes sugar, butter, and flour, as found in all the fruit pie recipes in your cookbook, rather than finding pickles.

Slide 6

Thanh Lan

Ethics is related to morality and subjectivity. just because it’s legal doesn’t mean it’s ethical. As developers and solution designers, we have a moral responsibility toward our users. It’s not because your country doesn’t legislate against discriminating against minority individuals, especially those with disabilities, that it is ethically correct.

Law takes time to be created and modified, but it often aligns with the morality of a given society or where it is meant to be applied. Therefore, what is legal in one country is not necessarily moral for a citizen of another country. Artificial intelligence is highly sensitive to the cultural environment in which its models were created. For example, the use of AI in video surveillance is widely employed in certain countries but prohibited by law or heavily regulated in others. There are existing regulations around AI, and more are on the horizon, such as the AI Act in the European Union.

But some dispositions of the GDPR (General Data Protection Regulation) already have an impact on AI-integrated applications. We have previously seen that data is the foundation of all AI applications.

Slide 7

Emmanuelle

Let me give you an overview of disability.

Slide 8

Emmanuelle

Did you know that one billion people in the world are disabled? It is estimated that about 15% of the world's population has a disability. This figure is very difficult to estimate for several reasons:

  • a person may not declare their disability or does not know they have a disability
  • a disability can occur during one's life and being diagnosed can be an obstacle course to have one's disability recognized

Slide 9

Emmanuelle

Contrary to popular belief, disability is not just a problem for people with wheelchairs. A disability is often not very visible. Did you know that 80% of disabilities are not immediately visible? Nevertheless, we have a fact-check on this figure. It is often said that 80% of disabilities are invisible, but we don't really know the exact figures. What is certain is that the majority of disabilities are invisible.

For example, We are talking to you, but before you introduced us, you didn't know we was deaf. Our deafness are invisible. Also, as we're live, you can't guess that my partner Thanh Lan is in a wheelchair.

According to the sources in France, there are 5 main families of disabilities:

  • Physical disability
  • Sensory impairment
  • Intellectual disability
  • Mental disability
  • Disabling diseases (such as endometriosis, cancer or Charcot's disease for example). These are disabling diseases that can be disabling on a daily basis in the short or medium term.

We can have multiple disabilities.

Around you, you probably know someone who is affected by disability. Maybe you're concerned yourself. In the course of his life, we are not immune to being affected by a disability, which I do not wish on you, of course.

Slide 10

Thanh Lan

Thank Emmanuelle for this reminder. Now we have some theoretical knowledge, we can move on to practical examples.

Slide 11

Thanh Lan

If we will generate some images of person with disabilities people for use in advertising, for example. We need data and we choose to scrap images on Google images. There are the first results on google image when we search “people with disabilities” in this small video. We can see there are lot of wheelchair.

Slide 12

Thanh Lan

It’s ok for me, because I have disabilities and I am a wheelchair user.

In France and many other Western countries, we're used to seeing this wheelchair logo for people with disabilities.

In Greta Gerwig's Barbie film, we have a protagonist in a pink wheelchair. The film is very nice, I liked it a lot, but for a film seen as an ode to diversity, disability was reduced to a single character in a wheelchair with no dialogue.

This vision of disability summed up just by the wheelchair is a cognitive bias. It seems logical to us, but it's wrong.

This has two negative impacts: it's going to be in the very definition of disability. If I'm not in my wheelchair, that doesn't mean I'm no longer disabled, or that I've been cured, and above all….

Slide 13

Thanh Lan

it will exclude people or erase certain disabilities. For example, if the only criterion is the use of a wheelchair, it excludes many people like Emmanuelle who are still living with disabilities.

And if we create our model base on google images or a another biased database, we have this result : this is some images generate with microsoft designer. All this image represent a caucasian women (no racialized people, no men) and all in a wheelchair when wheelchair users are a minority of people with disabilities.

Slide 14

Emmanuelle

Jeremy Andrew Davis is autistic and tested the generation of autistic people with Midjourney.

You can see it through this video, on a sample of a hundred images, that all these images look the same. The AI-generated autistic person is commonly sad, depressed, always has the same weird face. In terms of diversity, he's still a white man. For AI, autistic people all look the same way. However, this is not the reality. Why does an autistic person have to be sad and depressed? Doesn't a disabled person have the right to feel good about themselves, to be happy?

It can be said very clearly that artificial intelligence has biases.

Slide 15

Thanh Lan

I would like you to imagine that we are in a team developing an AI project. And we will focus on the moments when we can create some difficulties for people with disabilities. The first step is to think about what is the need.

The first step is to brainstorm on the need you want to fill. Choose suitable metrics to evaluate and monitor our model.

These will vary according to our problem.

For example: in disease detection, we'll judge that it's less serious to have a false positive than a false negative, and potentially miss a patient. In the case of a false positive, the doctor can always check the tests manually or perform other analyses before treating the patient.

On the other hand, for a targeted commercial offer, false negatives may be considered less serious: if Mr X didn't specifically receive the email about the wheelchair promotion, this has little impact.

Slide 16

Thanh Lan

EDA : exploratory data analysis is a very important phase in all data science project.

We analyze the data at our disposal: their quantity and quality, for example, whether there are many missing values. Since data science relies on statistics and probabilities, we want to handle data with slightly aberrant values because they introduce noise into our model and make it less performant. Imagine that these cars in the parking lot represent your dataset; humans are somewhat like the cars in a parking lot—diverse and different. Now consider that we’re talking about all the software engineers in a typical company. In most countries, the majority of software engineers are men. However, if we were part of a dataset of software engineers, Emmanuelle and I would be considered outliers: not only are we women, but we also have disabilities. We don’t fit the typical profile, and we wouldn’t want to be completely erased from the tech industry because our profiles are unique. Diversity matters.

Slide 17

Thanh Lan

It’s like creating a prototype to eventually build a real car. The goal is to find the model that achieves the best results and therefore the highest scores on the metrics you determined in the initial stage. A common mistake would be to rely solely on these results to claim that your model is performing well. For example, if you use datasets related to American Sign Language, you might achieve very good results on your metrics… but…

Slide 18

Thanh Lan

But if you put this model into production in France, for example, it won’t work well because people use another sign language: French sign language. As a result, you’ll receive very poor user feedback. This also presents challenges common to all software: maintainability, scalability, response time, etc. Additionally, you need to monitor performance and retrain the models when there’s a decrease in performance. We can also discuss drift: changes in context that make our model less effective.

Slide 19

Thanh Lan

Here’s also a risk during retraining, as it often requires an exploratory phase using data collected in production. Some models can be negatively influenced by interactions with users (for example, models that become more biased).

Slide 20

Emmanuelle

When Midjourney came out, I was curious and I asked Midjouney to generate a disabled person. It generated men in wheelchairs for me. But I'm not in a wheelchair not all.

Thanh Lan, are these images representative of you?

Thanh Lan

I use most of the time a wheelchair... but I 'm not a man, I hope I'm not sad all the time and I am Asian. I know Midjourney doesn't know I am a asian girl but the majority of people with disabilities I know are not sad.

Emmanuelle

So, I asked to generate a deaf person for me. Midjourney has generated only very old women. Do I look like an old woman? Deafness can affect people at any age and is not limited to the elderly.

Thank Lan, are you old too?

Thanh Lan

No, I don’t... And this representation is very problematics because For young deaf people or young people who become deaf, it's very hard to accept deafness if the society just recognizes older people as the only legitimate deaf people.

Emmanuelle

Finally, I tried one last time to generate a deaf developer like me because representation matters. It generated operators.

Thanh Lan, do you know why?

Thanh Lan

Women and the deaf are not statistically represented. It's complicated for the models to associate these two terms, so they rely on what they think is correct: a female operator. In an industry where we're trying to do so much to promote inclusivity, this is very disappointing.

Emmanuelle

We are definitely not an operator. Personally, I don't use a headset. Well, not anymore. When I listen to music or when I make a phone call, my cochlear implants have Bluetooth, they basically become AirPods. This means that I can listen to music that is neither seen nor known.

Slide 21

Emmanuelle

I'm going to talk to you about innovations that are having an impact the daily lives of disabled people.

Slide 22

Emmanuelle

Let's start with automatic caption and transcription. It is one of the most well-known tools. I mentioned this earlier. We are seeing more and more automatic caption and transcription in video and video platforms. Automatic caption is:

  • Increasingly efficient
  • Increasingly used in everyday life
  • Available in multiple languages
  • Can be used for machine translation
  • Easy to use. This is because it is easy to integrate automatic captions into the tools.

Slide 23

Emmanuelle

But can we really rely on it fully?

When you are in a video conference or when you are watching a live video, there are often automatic captioning errors. We have to deal with it by using mental replacement since we can't go back or pause to tell the AI that it made a mistake.

This means that we are forced to read lips (when the image is good), listen (when we have hearing aids) and analyze the context when there are automatic captions errors. It's exhausting.

When the video is not live and the video is uploaded to video platforms such as YouTube for example, there are automatic captions. As I said, automatic captions are not yet 100% reliable and, therefore, require human intervention to correct errors.

So don't hesitate to use automatic tools to create captions because they do all the work of syncing, but don't forget to check that they are correct. If not, correct. By correcting, you're showing the AI that it's made mistakes. And we know that she learns from her mistakes.

I gave a talk in French on automatic captions at Paris Web. If you are interested, I invite you to watch it to better understand the captions.

Slide 24

Seeing AI is an application developed by Microsoft that automatically describes the environment around us. Among other things, it allows you to:

  • Read text aloud as soon as it appears in front of the camera
  • Scan a document and read it aloud
  • Beep to locate barcodes and then analyze them to identify products
  • Recognize the people around you and decipher their emotions
  • Describe scenes and recognize images
  • Identify banknotes

Or Be My Eyes. It's an app that connects blind and partially sighted people with volunteers. Volunteers provide visual assistance to blind and visually impaired users via video call.

With the arrival of GPT and recently GPT 4o, Be My Eyes has created a new Virtual Volunteer tool. AI would be able to analyze the context and give the answer just like a human volunteer would.

But one question remains, can the blind or partially sighted person blindly trust AI?

This raises an ethical and moral question. If the AI makes a mistake, it can have more or less serious consequences.

Slide 23

Emmanuelle

There are plenty of innovations in progress or in beta that can be useful. There are tremendous opportunities to improve the lives of disabled people.

Like Signer.ai or Signapse.ai offer automatic American Sign Language translations on videos, texts and audios.

Emoface, an AI that can recognize emotions to help autistic people.

Wiseone rephrases complicated texts.

Oticon reduces ambient noise in hearing aids.

Glaaster transforms texts for dyslexic people.

Otter.ai Voice takes notes and writes summaries.

Speechify reads texts aloud.

SymboTalk is used to communicate using images and symbols.

Sesame Enable turns smartphones and tables into hands-free devices.

Waymap makes travelling easier by providing detailed audio instructions.

Mintt detects falls in real time and alerts emergency services.

And Rango is a smart cane that detects obstacles and helps blind people to find their way around.

There are a lot of possibilities and it's very exciting!

Slide 26

Emmanuelle

In addition to the biases that are present in AI, unfortunately, there have been dramas with AI. A person stressed about global warming saw his mental health change as he conversed with Eliza, an AI, and confided his feelings in her. This person has found in the AI a confidant and has forgotten that Eliza is devoid of feelings, of empathy.

So the person said, I want to die, do you think I should? Eliza replied, "I'd like to see you dead!"

Sadly, the person committed suicide.

As a result of this, the startup that built Eliza put safeguards in place to prevent it from happening again. When there are obvious signs of suicide or depression, there are numbers available.

Our biases have a strong impact and can sometimes have a dramatic impact on disabled people. That's why it's important to work with disabled people to prevent this from happening again.

Slide 27

Emmanuelle

Let me remind you – and we tend to forget them – artificial intelligence is a tool.

Sound recognition, for example, can cause false positives. I had used this system and had so many false positives that I ended up not using it anymore. An intercom or doorbell ringing when there's no one behind my door. I also had alerts telling me that police sirens were sounding when they were coming from the television. I didn't know what was real and what wasn't, so I turned it off.

Terms or contexts that don't mean anything. I see it regularly with automatic captions. As I said, mistakes exist and must be corrected. Tim Cook once gave a speech at Gallaudet University, a university for deaf and hard-of-hearing students, saying "AI is good but is not that good". This means that we cannot rely totally on AI and that we still need human intelligence to correct errors.

Thanh Lan

About mistakes, this picture is a little robot in Estonia, they delivers your parcels like food. It's something that works quite well in the countries where it's deployed. During our research, we saw projects to develop autonomous wheelchairs a bit like these robots, but when I see this photo I can only be worried. The same goes for blind people, a mistake can have serious consequences, even more so than if you're delivering a burger.

Like Emmanuelle, I tested and abandoned sound recognition because it wasn't reliable enough, and projects that are too expensive for disabled people, even if they are technologically interesting, are also of little interest. Our needs are often different from what uninvolved people can imagine: by translating a signed language like French sign language and very different from the image people have of it: deaf people sign very quickly, body posture and facial expressions are very important. What more, this won't make the content accessible to all deaf people: I'm deaf and I can't sign.

Slide 28

Emmanuelle

Can AI improve website accessibility?

No. But you can use automated testing to detect accessibility issues. I've already asked ChatGPT to incorporate accessibility into the code, it didn't work well.

And also, no overlay tool can make a website accessible.

The only way to make a website accessible to everyone is to get your hands on the code.

Slide 29

Emmanuelle

AI has already changed our lives every day. I regularly use automatic captions, even if it's not perfect. I use an automatic tool to translate my content or reformulate it differently because my syntax isn't very good.

I'm just sorry that the AI doesn't understand me very well because of my deaf voice, but progress is being made.

And for you, Thanh Lan?

AI has enabled me to be more autonomous and safer. As a teenager, I never thought I'd be able to do so much on my own, and tools like fall detection mean I'm safe even when I'm home alone. I've been deaf for 20 years now, and the progress I've made in reducing ambient sounds has made my hearing aids more comfortable.

Slide 30

“Nothing about us without us”. It’s a mantra from USA. It is important to design tools with disabled people, to co-create together so as not to bias.

There is a real need to collaborate with disabled people to reduce risks and biases and to communicate with them to build the useful tools and make them effective.

Better yet, we need to hire disabled people in the tech industry. To do this, of course, they need to be trained and therefore made accessible to them.

Your products are making an impact in the lives of disabled people.

Thanh Lan, would you like to add something?

Thanh Lan

Today we're all talking about disability, but we're all someone else's other. It's important that there's more diversity in tech to combat bias in the design of models and products. Diversity is not just about disability, but also about gender, ethnicity, religion, etc. Your users are varied, so it's important that this diversity also exists in your teams.

Slide 31

Emmanuelle

Thank you so much for listening to us. You can scan the barcode to find our presentation, transcript and resources.