Please flip your device.

Let’s begin!

This year’s Internet Health Report is about the systems of power and human decisions that define how artificial intelligence is used, and whom it impacts. We set the scene here with an accessible compilation of research and data visuals about the current state of AI worldwide.

When we say AI, it’s shorthand for a wide range of automation and algorithmic processes, including machine learning, computer vision, natural language processing, and more.

Who has power?

From your social media feed to fast food restaurants, companies in every sector are turning to AI to unlock new ways to collect and analyze data to tailor their offerings.

But the benefits — and the harms — are not evenly distributed.

$ 15.7 trillion

This is how much AI is predicted to contribute to the global economy by 2030.

Sizing the prize, PricewaterhouseCoopers, 2017

The companies with resources to invest are carving out competitive advantages. And the countries with access to engineers, large amounts of data, and computational power, are consolidating their dominance of software and hardware in ways that impact how AI is deployed worldwide.

The United States and China are far ahead when it comes to private investments in AI. But that is just one indicator of how differently the rise of AI is experienced worldwide.

This visual shows which countries have the highest amount of private investment in AI in 2021 represented in stacks of dollar bills. The United States leads by a large margin with 52.88 billion USD. Second, also by a larger margin, is China with 17.21 billion USD. The remaining countries are: United Kingdom: 4.6 billion, Israel: 2.41 billion, Germany: 1.98 billion, Canada: 1.87 billion, France: 1.55 billion, India: 1.35 billion, Australia: 1.25 billion, South Korea: 1.10 billion.”

Another way power is reflected is at the very heart of AI research and development.

The cost of training machine learning systems has decreased, and the availability of data is greater. But even as more of the world delves into AI, a major imbalance is reflected in the landscape of AI research papers.

In thousands of papers, it is the same datasets from just a few countries that are used most often to evaluate the performance of machine learning models everywhere.

This visual shows a distorted world map, where the size of countries reflects how often datasets provided by organizations based in that country have been used for benchmarking AI. Datasets provided by organizations based in the United States are by far the most frequently used with 26,910 usages. Second is Germany (with 2,633 usages) and third is Hong Kong (with 1,567). The remaining countries and frequency of dataset usages: China 1,248, United Kingdom 1,119, Canada 835, Singapore 484, Australia 348, Israel 296, Belgium 289, Italy 260, Switzerland 243, Sweden 234, Colombia 223, Japan 209, Czechia 167, South Korea 157, Netherlands 149, Finland 114, Poland 96, Spain 91, France 84, India 73, Saudi Arabia 71, Pakistan 43, Austria 40, Taiwan 38, Greece 26, Turkey 23, Egypt 12, Bulgaria 12, Denmark 11, United Arab Emirates 11, Brazil 11, Ireland 10, Portugal 10, Iran 8, Russia 8, Qatar 8, Jordan 6, Norway 6, Bangladesh 6, Slovenia 5, Puerto Rico 4, Philippines 4, Croatia 4, Vietnam 4, Romania 3, Macao 3, Indonesia 2, Cyprus 2, Malta 1, Kazakhstan 1. No other countries have any dataset usages.

This doesn’t mean that datasets or machine learning models aren’t being developed in the rest of the world. They are!

But the discourse about how AI should be used — and who should benefit from it — is currently heavily weighted toward people and institutions who already wield tremendous power over the internet (and the world).

In fact, more than half of the datasets used for AI performance benchmarking across more than 26,000 research papers were from just 12 elite institutions and tech companies in the United States, Germany, and Hong Kong (China).

This is a bar chart that shows twelve institutions and how many times their datasets were used for benchmarking: Stanford University 3,354, Microsoft 3,274, Princeton University 2,524, Max Planck Society 2,114, Google 1,506, The Chinese University of Hong Kong 1,325, AT&T 1,159, Toyota Technological Institute at Chicago 1,124, New York University 906, Georgia Institute of Technology 756, University of California, Berkeley 736, Facebook 657.

A large and frequently reused dataset does not guarantee better machine learning than a smaller one designed for a specific purpose.

On the contrary, many of the most popular datasets are made up of content scraped from the internet, which overwhelmingly reflects words and images that skew English, American, white, and for the male gaze.

This interactive visual shows an illustration of a paper folder containing five datasets frequently used to benchmark AI. ImageNet is a dataset with more than 14 million annotated images sourced from web searches and the photo-sharing platform, Flickr. It’s a popular dataset for benchmarking in optical recognition. Laion400 is a dataset with 400 million image and text pairs that were automatically filtered from a gigantic dataset of internet content, Common Crawl. The data is used to train large scale models. CelebA is a dataset with more than 200,000 photos of celebrity faces gathered from the internet, annotated with facial attributes like smiling, bearded, etc. It is used for training and testing of computer vision. COCO (which stands for ‘Common Objects in Context’) is a dataset with 328,000 images selected via targeted searches on the photo-sharing platform Flickr. The images were annotated by crowd workers. It’s used for object detection. Finally, ‘Stanford Natural Language Inference’ is a data with 570,000 sentence pairs based on image captions in the Flickr30k dataset. Crowd workers then labeled how sentences interrelate. It’s used for training models to make text inferences.

For instance, the machine learning models most credited for advancing the field of automated language generation, like GPT-3, frequently reproduce racist and sexist stereotypes, in large part due to their training data.

Why not use other datasets? Machine learning models and datasets reflect both the biases of their makers and power dynamics that are deeply rooted in societies and online, but this is not widely acknowledged. More datasets should be created specifically to diversify machine learning methods for equity.

Raesetje Sefala
Computer vision researcher, DAIR Institute

This is the one major thing that we should all be thinking about. The people behind the data points, and the people creating these algorithms.

Raesetje Sefala

Who is accountable?

There is no question that the companies who stand to gain the most from AI are the world’s biggest tech companies.

Their revenues have skyrocketed throughout the global COVID-19 pandemic, and several are among the highest earning companies of all time.

This chart shows how the combined revenue of Amazon, Apple, Google/Alphabet, Microsoft, Facebook/Meta, Alibaba, Tencent and Baidu has grown continuously since 2017, with a sharp increase since the start of the pandemic in 2020. The combined revenue was more than 1.5T USD in 2021.

Each of these companies make money in different ways, but AI is core to the business operations of all of them.

Let’s consider Amazon, the highest earning tech company in 2021. AI is key to every major revenue category reported by Amazon last year.

This chart shows the seven major revenue categories of Amazon from 2021. For each category, it shows how many US dollars Amazon made in 2021 and how AI is relevant to it. First, Online Stores (222.08 billion USD): The online goods are not only recommended by algorithms, they are handled by workers in warehouses managed by highly automated systems. Second, third Party Seller Services (103.37 billion USD): Goods from other providers are seamlessly integrated into search engines of online stores for which Amazon takes commissions and fees. Third, cloud Computing (62.2 billion USD): Amazon Web Services has the biggest market share of all cloud service providers and handles data storage and computing for AI worldwide. Fourth, subscription Services (31.77 billion USD): Most Amazon subscription services are powered by AI. From Alexa voice assistants, to Ring camera doorbells, and countless AI enterprise services. Fifth, advertising (31.16 billion USD): Amazon sells digital ads that are automatically distributed on Amazon websites as well as via their text, audio, and video content products worldwide. Sixth, physical Stores (17.08 billion USD): Amazon’s futuristic supermarkets and stores track people, products, and payments with AI sensors such as optical recognition and biometrics. Finally, other (2.18 billion USD): Data insights gleaned from Amazon’s many services lead to technical improvements and commercial advantages that are instrumental for their AI development.

Big tech companies play an outsized role in shaping our experience of the internet, and life itself. What we see, what we buy, even what we believe is nudged along by them daily.

Yet, according to Ranking Digital Rights, there is little to no transparency into how companies test and deploy algorithmic systems and use our personal data, including when it comes to curation, recommendation, and ranking.

The datavisual shows how nine companies scored from best (100) to worst (0). Their scores are: 1. Microsoft: 55 out of 100, 2. Twitter: 50 out of 100, 3. Alibaba: 45 out of 100, 4. Tencent: 40 out of 100, 5. Facebook/Meta: 35 out of 100, 6. Baidu: 33 out of 100, 7. Google/Alphabet: 23 out of 100, 8. Amazon: 20 out of 100, 9. Apple: 0 out of 100.

Recommender systems determine what people see in social media feeds and what content is rewarded with ad revenue. False and harmful content that drives up engagement is frequently recommended by platforms, even when it may later be found to violate platform rules.

In a crowdsourced study of YouTube by Mozilla, 71% of the videos people said they “regretted” watching were algorithmically recommended to them.

This bar chart shows that 71.06% discovered videos they regret because YouTube recommended them. Only 7.47% discovered them by searching on YouTube. Another 21.48% discovered them otherwise, e.g. by clicking a URL on a website.

Because there is so little transparency into how systems work, researchers often need to recruit users for data donations to help study platforms and seek answers.

Justin Arenstein
Founder, Code for Africa

Platforms should be doing more sharing and more joint problem solving, to solve a problem that, ultimately they’ve created.

Justin Arenstein

The boom in AI is accelerated by astronomical levels of data collection, by big tech and others. Almost everything we do is tracked and analyzed.

It’s hard to fathom just how much data is collected, kept, and sold.

Every 5 minutes

This is how often your location may be logged by mobile apps in the $12 billion location data industry.

Who Is Policing the Location Data Industry?, Alfred Ng, Jon Keegan. The Markup, 2022

Why does this matter?

Because AI offers companies with intimate knowledge about us new ways to predict what we’ll do — and new ways to influence how we behave.

It can be as simple as offering two people a different price for the exact same service without telling you.

This visual allows you to click on six icons of people along an axis that shows dollar amounts. It displays: A subscriber in South Korea that is 18-29 years old and is a man seeking women in South Korea needs to pay 5.16$. Another subscriber from South Korea that is 50+ years old and is a woman seeking men needs to pay 22.36$. A subscriber from the US that is 30-49 years old and is a woman seeking women in the US needs to pay 7.99$. Another subscriber from the US that is 30-49 years old and is a man seeking men needs to pay 26.99$. A subscriber from India that is 18-29 years old and is a man seeking women in India needs to pay 4.52$. Another subscriber from India that is 50+ years old and is a woman seeking women needs to pay 10.25$.

Digital ads and social media networks are also weaponized to spread disinformation. In the absence of greater transparency and collaboration with researchers, a global industry of companies and organizations engaged in covert messaging is thriving.

81 countries

This was the number of countries where social media was used for “computational propaganda” in 2020.

Industrialized Disinformation: 2020 Global Inventory of Organized Social Media Manipulation, Samantha Bradshaw, Hannah Bailey & Philip N. Howard. Oxford Internet Institute, 2021

Is it fair?

The problems of misinformation and hate speech are felt worldwide, but platforms do not respond to them with urgency everywhere. Platforms develop AI to moderate content at scale, but do not resource it equally in all languages.

For instance, although 90% of Facebook’s users live outside the United States, only 13% of moderation hours were allocated to labeling and deleting misinformation in other countries in 2020. India alone has more users of Facebook, by far.

This barchart shows the names of ten countries and the number of users there: India: 329.65 million Facebook users, United States: 179.65 million users, Indonesia: 129.85 million users, Brazil: 116 million users, Mexico: 89.7 million users, Philippines: 82.85 million users, Vietnam: 70.4 million users, Thailand: 50.05 million users, Egypt: 44.7 million users, Bangladesh: 44.7 million users
Sahar Massachi
Cofounder, The Integrity Institute

At the very least, you need to have some people who understand the language and are paying attention to it as their job.

Sahar Massachi

Automated systems are frequently trained on data that is manually annotated by humans. For example, it can take hundreds of hours of labeling to prepare just one hour of video for the computer vision of a self-driving car.

This labor that is invisible to end-users is often exported to countries with low wages or completed via global crowd work platforms like Amazon Mechanical Turk.

$2.83​ per ​hour

This was the median hourly wage for Amazon Mechanical Turk workers when accounting for invisible labor in 2021.

Quantifying the Invisible Labor in Crowd Work, Carlos Toxtli, Siddharth Suri, Saiph Savage, 2021

Often, when the algorithmic management of labor is central to a business, fairness takes a backseat to maximizing productivity. This is true of hundreds of “gig work” platforms around the world. An estimated 40% of gig workers earn below the minimum hourly wage in their own countries.

This visual displays on a graph how gig worker earnings in 14 different countries compare to the local minimum wage in labor categories of ‘care provider’, ‘delivery worker’, ‘domestic worker’, and ‘driver’. Most are concentrated in the range just above the local minimum wage or below. In Argentina, all four categories score above the minimum wage. In Indonesia, only domestic workers earn above the minimum wage.

This imbalance is amplified by the lack of data privacy regulations in many countries. And even in countries that do have them, AI is increasingly used by authorities to expand surveillance and control with facial recognition, license plate readers, and more.

There are an estimated upto 1 billion surveillance cameras worldwide. This means one camera for every eight people on the planet.

The appetite for applying AI to policing and carceral systems worldwide is huge. For instance, via predictive policing and pre-trial risk assessments.

However, AI is sending people to jail and getting it wrong.

Yeshimabeit Milner
Executive Director, Data for Black Lives

It takes a lot of imagination and creativity to move out of the rigid definition of how data should be used, to think differently about data and to reclaim it to make tools that are not oppressive.

Yeshimabeit Milner

In real life, over and over, the harms of AI disproportionately affect people who are not advantaged by global systems of power.

News headlines about algorithmic biases are glaring signals of how technology can be used to oppress rather than uplift people.

The headlines displayed in this visual are: How an Algorithm Blocked Kidney Transplants to Black Patients, Wrongfully Accused by an Algorithm, How French Welfare Services Are Creating ‘Robo-Debt’, Meet the Secret Algorithm That’s Keeping Students Out of College, Personal Voice Assistants Struggle with Black Voices, New Study Shows, Court Rules Deliveroo Used ‘Discriminatory’ Algorithm.

What can be done?

What values are advanced by researchers? Frequently, they are commercial ones.

Today, nearly half of the most influential machine learning research papers — and many top AI university faculties — are funded by big tech. More research from a broader set of people and institutions could help shift industry norms.

This visual shows a sharp rise of authors affiliated with Big Tech represented graphically between two pencils. In detail: Big Tech Affiliation: From 12.77% in 2008-2009 to 47.17% in 2018-2019. Other Corporate Affiliation: From 10.64% in 2008-2009 to 7.55% in 2018-2019. No Corporate Affiliation: From 76.60% in 2008-2009 to 45.28% in 2018-2019.

By some measures, ethical considerations in the research field are on the rise, but more interdisciplinary understanding of risk and harms is still needed.

The vast majority of leading research papers are centered on technical performance, not social needs or risks. And few of the most cited papers discuss ethical principles or user rights.

This graphic representation of stacked papers in different quantities shows that the large majority of values uplifted by the 100 most highly cited papers are related to technical performance: 97,02%. Only 1.78% of the values are user rights and 1.78% are ethical principles.

The barriers and costs for building AI are lower thanks in part to a new generation of open source tools and independent and grassroots AI developer communities worldwide.

But many of the same harms of big tech and big data AI development will be repeated if more trustworthy research, data, and development practices are not adopted.

Michael Running Wolf
Founder, Indigenous in AI

If you go into a community with a mindset that data is just a resource, a monetary valuable thing, you’re fundamentally harming the community and you’re also diminishing the value of this data.

Michael Running Wolf

Communities guided by values of fairness and human rights are challenging us to rethink not just how AI is built but for whom. Such questions are not being asked or answered by big tech.

Regulation can help set guardrails for innovation that diminish harm and enforce data privacy, user rights, and accountability. Many laws already apply to AI, but policies that are specific to AI (or sometimes bans) are also surfacing in different regions, countries and cities. In the public sector, there are many novel approaches to AI accountability.

This visual allows for selection of 8 different principles and shows their descriptions. 1) Principles and Guidelines: Guidelines on how AI should be implemented (binding as well as non-binding) can help identify high-level policy goals, principles, and values and provide normative standards against which organizations, and in some cases the public, can evaluate practices. 2) Prohibitions and Moratoria: Bans or prohibitions on the use of particular “high risk” algorithmic systems (as in the use of facial recognition by law enforcement) can be issued permanently or temporarily when safeguards and accountability mechanisms are lacking. 3) Public Transparency: Public registries of algorithmic systems, source-code transparency requirements, and mandatory explanations of algorithmic logics, are all examples of transparency mechanisms that can provide information about algorithmic systems for accountability. 4) Impact Assessments: Algorithmic impact assessments can help organizations understand, categorize and respond to the potential risks or harms of algorithmic systems, in the public sector and beyond. Impacted communities can also be directly involved in the evaluation of algorithmic systems. 5) Audits and Inspections: Subjecting algorithmic systems to independent audits can mean either looking at how systems function technically to account for flaws or biases, investigating their compliance with regulations, or to understand the financial or social consequences of a system. 6) Independent Oversight Bodies: An external or independent oversight mechanism can monitor the algorithmic systems of public bodies or companies and make recommendations or decisions about their use. Oversight can be done via legislation, executive offices, or various advisory functions. 7) Rights to Hearing and Appeal: Decisions aided by algorithmic systems can be required to provide a forum for affected individuals or groups to contest biased or erroneous decisions. A fair process could involve explanations for decisions, a hearing, and the ability to appeal decisions. 8) Procurement Conditions: When governments procure algorithmic systems it can be required that they be transparent and non-discriminatory. Governments could then only acquire systems that comply with certain standards of fairness and vendors would have to comply.

This is a rapidly evolving field. No one has all of the answers.

How do we build trust? New techniques and methods for preserving privacy, for logging the origins of data, for operationalizing ethics, for auditing algorithms and giving users more power, are among many paths explored.

Mozilla is building knowledge through convenings and collaborations with people around the world who are rethinking the use of data and AI for more equitable outcomes. Our core purpose is for the internet to be healthy. We fight for an internet with more privacy and security, openness, inclusivity, and where power is decentralized in ways that benefit humanity. With AI, we often see the worst health issues of the internet amplified. This is why we are researching, campaigning, and grantmaking for trustworthy AI with such urgency.

Collaboration across multiple sectors is necessary for solutions. With this report, the podcast, and stories that accompany it, we especially call on tech builders and policy folks to engage in the conversation and act.

References

10 Projects Rethinking Data Stewardship: Announcing Mozilla’s Latest Creative Media Awards, Mozilla, 2022

Algorithmic content curation, recommendation, and/or ranking systems (F12), Ranking Digital Rights, 2022

A consumer investigation into personalised pricing, Consumers International, Mozilla Foundation, 2022

A World With a Billion Cameras Watching You Is Just Around the Corner, Liza Lin, Newley Purnell. Wall Street Journal, 2019

AI Audit Challenge, Stanford University Human-Centered Artificial Intelligence Institute, Stanford Cyber Policy Center, 2022

AI Data Labelling. Roundtable Readback, aapti institute, 2020

AI is sending people to jail—and getting it wrong, Karen Hao. MIT Technology Review, 2019

Algorithmic accountability for the public sector. Learning from the first wave of policy implementation, Divij Joshi, Tonu Basu, Jenny Brennan, and Amba Kak. Ada Lovelace Institute, AI Now Institute, Open Government Partnership, 2021

Amazon Mechanical Turk, Amazon, 2022

Annual Report 2021, Amazon, 2022

AWS AI services, Amazon

Annual reports and press releases of big tech companies from 2017-2021

Artificial Intelligence Incident Database, Responsible AI Collaborative, 2022

Artificial Intelligence Index Report 2022, Stanford University Human-Centered Artificial Intelligence Institute, 2022

CelebA, The Chinese University of Hong Kong

COCO: Common Objects in Context, Microsoft

Creating Trustworthy AI, Becca Ricks, Mark Surman. Mozilla, 2020

Data Futures Lab, Mozilla

Datasheets for Datasets, Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, Kate Crawford, 2021

Ethical AI Ecosystem, Abhinav Raghunathan. The Ethical AI Database, 2022

Ethical and social risks of harm from Language Models, Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins, Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks, William Isaac, Sean Legassick, Geoffrey Irving, Iason Gabrie. DeepMind, 2021

Facebook Employees Flag Drug Cartels and Human Traffickers. The Company’s Response Is Weak, Documents Show. Justin Scheck, Newley Purnell, Jeff Horwitz. Wall Street Journal, 2021

How do the biggest internet companies make money?, Mozilla, 2019

ImageNet, Stanford University, Princeton University

Industrialized Disinformation: 2020 Global Inventory of Organized Social Media Manipulation, Samantha Bradshaw, Hannah Bailey & Philip N. Howard. Oxford Internet Institute, 2021

LAION-400M, Laion

Leading countries based on Facebook audience size as of January 2022, Statista, 2022

Liberty at Risk: Pre-trial Risk Assessment Tools in the U.S., epic.org, 2020

Male gaze, Wikipedia, 2022

Mozilla Festival

Mozilla Foundation

Mozilla’s vision for the evolution of the Web, Mozilla, 2022

Multimodal datasets: misogyny, pornography, and malignant stereotypes, Abeba Birhane, Vinay Uday Prabhu, Emmanuel Kahembwe, 2021

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell, 2021

Quantifying the Invisible Labor in Crowd Work, Carlos Toxtli, Siddharth Suri, Saiph Savage, 2021

Real Change How?, Mozilla, 2021

Realizing the Potential of AI Localism, Stefaan G. Verhulst , Mona Sloane. Project Syndicate, 2020

Reclaim Your Face

Reduced, Reused and Recycled: The Life of a Dataset in Machine Learning Research, Bernard Koch, Emily Denton, Alex Hanna, Jacob G. Foster, 2021

Responsible AI in Africa: Challenges and Opportunities, Damian O Eke, 2022

Responsible Computer Science Challenge, Mozilla

Rising Through the Ranks, Spandana Singh. New America, 2019

Self-driving cars prove to be labour-intensive for humans, Tim Bradshaw. Financial Times, 2017

Sizing the prize, PricewaterhouseCoopers, 2017

Small Data’s Big AI Potential, Husanjot Chahal, Helen Toner, Ilya Rahkovsky. Center for Security and Emerging Technology, 2021

Stanford Natural Language Inference, Stanford University

State of AI Report 2021: More money, more influence, Nathan Benaich, Ian Hogarth, 2021

Street-Level Surveillance, EFF, 2017

The 2022 BTS Executive Summary, Ranking Digital Rights, 2022

The Biggest Data Breach, Irish Council for Civil Liberties, 2022

The Efforts to Make Text-Based AI Less Racist and Terrible, Khari Johnson. Wired, 2021

The gig workers index: Mixed emotions, dim prospects, Peter Guest, Youyou Zhou. Rest of World, 2021

The Mozilla Manifesto, Mozilla

The Values Encoded in Machine Learning Research, Abeba Birhane, Pratyusha Kalluri, Dallas Card, William Agnew, Ravit Dotan, Michelle Bao, 2021

Toward User-Driven Algorithm Auditing: Investigating users’ strategies for uncovering harmful algorithmic behavior, Alicia DeVos, Aditi Dhabalia, Hong Shen, Kenneth Holstein, Motahhare Eslami, 2022

Trustworthy AI Working Groups, Mozilla Festival

Who Is Policing the Location Data Industry?, Alfred Ng, Jon Keegan. The Markup, 2022

YouTube Regrets, Jesse McCrosky and Brandi Geurkink. Mozilla, 2021