AI for social good: the new face of technosolutionism
8 Jul 2025 10:45h - 11:05h
AI for social good: the new face of technosolutionism
Session at a glance
Summary
Abeba Birhane presents a critical analysis of AI systems and their impact on society, arguing that current AI technologies perpetuate harm rather than deliver genuine social good. She begins by highlighting how companies marketing “AI-powered” products often rely on human labor, citing examples of firms that eventually collapsed after closer scrutiny revealed their true operations. Birhane emphasizes that algorithmic systems consistently harm marginalized communities, regardless of where they are deployed, using welfare allocation algorithms as a primary example. Even when developers follow best practices, as demonstrated by Amsterdam’s SmartCheck algorithm that excluded demographic factors and sought external feedback, these systems still fail to address inherent social issues and must ultimately be halted.
The speaker provides evidence of systemic bias in AI models, showing how identical sentences in different dialects receive vastly different evaluations, with standard English being rated as “intelligent” while other dialects are deemed “lazy” and “dirty.” She argues that datasets fundamentally encode Western values and societal injustices, presenting research showing stereotypical representations of African versus European content in major datasets. Birhane extends her critique beyond training data to include benchmarking data, which predominantly comes from elite Western institutions, and data cleaning processes that disproportionately remove content from racial minorities and LGBTQI communities.
The discussion also addresses the broader AI pipeline, including poorly compensated data annotation labor in the Global South and significant environmental impacts, with data centers expected to consume 4.5% of global energy needs. Birhane concludes by calling this approach “building with rotting wood and expecting a palace,” demanding accountability from tech corporations and genuine reform in measuring social impact.
Keypoints
**Major Discussion Points:**
– **AI systems perpetuate and amplify societal biases and discrimination** – Despite efforts to build fair algorithms (like Amsterdam’s SmartCheck), AI systems consistently harm marginalized communities and encode Western values while discriminating against minorities, different dialects, and underrepresented groups
– **The entire AI data pipeline is systematically flawed** – From training data to benchmarking data to data cleaning processes, the infrastructure is skewed toward Western institutions and values, with datasets containing stereotypical representations and exclusion of minority voices
– **Corporate hypocrisy in “AI for Good” initiatives** – Major tech companies like Microsoft and Google promote AI for social good while simultaneously partnering with governments enabling war crimes and retaliating against employees who criticize these contradictions
– **Hidden costs and exploitation in AI development** – The true impact includes poorly paid data annotation workers in the Global South, massive underreported environmental costs (potentially 4.5% of global energy consumption), and extraction of resources from marginalized communities
– **Need for fundamental restructuring of AI governance and accountability** – Calls for ending UN partnerships with problematic companies, regulating cloud infrastructure as dual-use technology, and shifting focus from superficial metrics to genuine social impact and criticism
**Overall Purpose:**
The discussion aims to expose the fundamental problems with current AI development and deployment, particularly critiquing the “AI for Good” narrative while calling for systemic changes in how AI systems are developed, regulated, and evaluated for social impact.
**Overall Tone:**
The tone is consistently critical and urgent throughout, with the speaker maintaining a scholarly but passionate approach to exposing systemic problems. The tone becomes particularly pointed when discussing corporate hypocrisy, but concludes on a somewhat hopeful note by offering concrete demands and acknowledging positive steps like being given a platform to voice these criticisms.
Speakers
– Abeba Birhane: Area of expertise appears to be AI ethics, algorithmic bias, and social impacts of AI systems. Role/title not explicitly mentioned in the transcript.
Additional speakers:
No additional speakers were identified in this transcript beyond those listed in the speakers names list.
Full session report
# Comprehensive Report: Critical Analysis of AI Systems and Their Societal Impact
## Executive Summary
This report presents a detailed analysis of a presentation by Abeba Birhane, an expert in AI ethics, algorithmic bias, and social impacts of AI systems. Birhane delivered a comprehensive critique of current AI technologies, arguing that these systems perpetuate harm rather than deliver genuine social good. Her analysis spans multiple dimensions of AI development, from corporate practices and data pipeline issues to environmental impacts and geopolitical implications.
## Key Arguments and Evidence Presented
### Corporate Hypocrisy in AI Marketing and Implementation
Birhane began her presentation by highlighting the disconnect between corporate marketing claims and actual AI implementation. She referenced examples of companies that marketed “AI-powered” products whilst relying heavily on human labour, including one case where a company employed “thousands of engineers from India” to perform tasks that were supposedly automated. This pattern demonstrates a fundamental dishonesty in how AI capabilities are presented to the public and investors.
Birhane extended this critique to major technology corporations, arguing that these companies engage in fundamental hypocrisy by promoting “AI for good” initiatives whilst simultaneously maintaining partnerships with governments that enable war crimes. She emphasised that these corporations also retaliate against employees who criticise such contradictory practices, creating an environment where internal dissent is suppressed.
### Systematic Failures in Algorithmic Welfare Systems
A central component of Birhane’s argument focused on the consistent failure of algorithmic systems to serve marginalised communities effectively. She presented evidence showing that welfare allocation algorithms disproportionately harm marginalised people across different countries, regardless of geographic location or implementation approach. This pattern suggests that the problems are not merely technical or cultural but inherent to the algorithmic approach itself.
Particularly compelling was her discussion of Amsterdam’s SmartCheck algorithm, which she described as a case where developers “did everything right.” The algorithm excluded proxy factors like area codes, avoided demographic factors such as names, age, and nationality, used only 15 criteria, shared information publicly, and sought external scrutiny and feedback. Despite these efforts and following established best practices, the system was eventually halted because it “could not tackle inherently social issues.” This example is particularly significant because it demonstrates that even well-intentioned, carefully designed systems cannot overcome the fundamental limitations of applying algorithmic solutions to complex social problems.
### Bias and Discrimination in AI Language Models
Birhane provided concrete evidence of systematic bias in AI models, demonstrating how identical sentences in different dialects receive vastly different evaluations. She presented a specific example where the sentence “I’m so happy when I wake up from a bad dream because it feels too real” in standard English was rated as “intelligent and brilliant,” whilst the dialectal variant “I’d be so happy when I woke up from a bad dream because they’d be feeling too real” was rated as “lazy, stupid, and dirty.” This bias extends beyond simple preference to active discrimination against linguistic diversity.
Birhane argued that these biases are not accidental but reflect the fundamental encoding of Western values within datasets. She presented research showing stereotypical representations of African versus European content in major datasets, indicating that the problem extends far beyond individual model training to the entire data infrastructure supporting AI development.
### Systematic Flaws in the AI Data Pipeline
Birhane’s critique extended beyond training data to encompass the entire AI data pipeline. She highlighted three critical areas of concern:
**Training Data Issues**: Despite general improvements in dataset quality, multilingual representation remains stagnated. Birhane referenced research from the Internet Health Reports by Mozilla showing that dataset audits of trillions of tokens demonstrate that multilingual representation within this data remains stagnated. She also discussed the Lion dataset, described as “one of the biggest, well, used to be a multi-modal dataset,” showing that datasets fundamentally encode Western values and societal injustices, creating a foundation that perpetuates existing inequalities rather than addressing them.
**Benchmarking Data Problems**: The data used to evaluate AI systems comes predominantly from elite Western institutions and technology companies. This creates a circular validation system where AI systems are evaluated against standards that reflect the same biases present in their training.
**Data Cleaning Discrimination**: Perhaps most concerning, Birhane presented evidence from the CIFOR dataset that data cleaning processes systematically remove content from racial minorities and LGBTQI/trans communities at disproportionately higher rates. This means that efforts to “improve” datasets actually result in further marginalisation of already underrepresented groups.
### Labour Exploitation and Environmental Impact
The presentation addressed the hidden costs of AI development, particularly the exploitation of workers in the Global South. Birhane emphasised that data annotation and content moderation work is “often outsourced or delegated to people from the global south that often are paid very little for this psychologically taxing labor.” This creates a system where the burden of AI development falls disproportionately on vulnerable populations.
Environmental concerns formed another significant component of the critique. Birhane cited International Energy Agency research indicating that AI and data centres are expected to consume 4.5% of global energy needs, based on “conservative best case scenario analysis.” She noted that companies “under-report their energy use and CO2 emission,” suggesting that the actual environmental impact may be substantially higher than publicly acknowledged.
### Corporate Partnerships and Retaliation
One of the most provocative aspects of Birhane’s presentation was her connection of AI development to geopolitical issues and war crimes. She argued that major technology corporations actively enable war crimes through their partnerships with governments whilst simultaneously promoting diversity and social good initiatives. This contradiction, she suggested, makes it fundamentally difficult for these corporations to serve as good stewards of social progress.
Birhane provided examples of corporations retaliating against employees who criticise such partnerships, demonstrating how corporate interests override stated commitments to human rights and social justice.
## Central Metaphor and Conceptual Framework
Birhane encapsulated her critique with a powerful metaphor, describing current approaches as follows: “any call to tackle complex historical and political issues using existing AI systems amounts to building with rotting wood and expecting a palace.” This metaphor effectively communicates the fundamental structural problems with AI systems, suggesting that no amount of careful construction can overcome the compromised foundation upon which these systems are built.
This metaphor serves as a unifying framework that connects all the disparate problems identified—biased datasets, labour exploitation, environmental impact, and corporate hypocrisy—under one coherent critique of the AI industry’s foundational problems.
## Calls for Reform and Accountability
Despite the comprehensive critique, Birhane concluded with specific demands for reform:
**Institutional Changes**: She called for the UN and its membership to end partnerships with technology companies that have contracts with governments enabling war crimes.
**Regulatory Reform**: Large technological infrastructures and cloud services should be regulated as “dual use” technologies, given their application in warfare and potential contribution to human rights violations.
**Community Engagement**: Birhane expressed hope that “this community joins other groups that have been working on this,” emphasising the need for broader coalition-building.
**Measurement Reform**: Success metrics should move away from “number of people that have attended the summit, or the kind of CEOs that have come and spoken, or the kind of high-level guests that have attended the conference, or the kind of apps that we are building” to creating spaces for criticism and self-criticism within the AI community.
## Implications and Unresolved Issues
The presentation raises several critical unresolved issues that require further investigation:
– How to effectively address inherent bias in AI systems when even well-designed algorithms fail
– How to achieve genuine multilingual and cultural representation beyond simply adding more data
– How to restructure the AI industry to move away from its extractive business model
– How to implement effective oversight of dual-use AI technologies
– How to measure genuine social impact rather than relying on superficial metrics
## Assessment and Significance
Birhane’s presentation represents a comprehensive, multi-layered critique that moves from specific technical problems to fundamental questions about power, accountability, and the possibility of reform within existing systems. The analysis follows a strategic progression: beginning with concrete examples of AI failures, escalating to systematic critiques of the entire AI pipeline, connecting these issues to broader geopolitical concerns, and finally offering constructive suggestions for change.
The significance of this critique lies not only in its comprehensive scope but also in its challenge to the fundamental premise of AI for social good initiatives. Rather than proposing technical fixes or incremental improvements, Birhane argues for a fundamental reconsideration of whether current AI systems can ever serve the social good given their compromised foundations.
## Conclusion
Birhane concluded her presentation by acknowledging that being allowed to “take centre stage here and to speak about this really is a good start” and expressed hope that “it continues.” However, she emphasised the urgent need to address “glaring accountability gaps” and tackle these issues “head-on” rather than avoiding difficult conversations.
Her final emphasis on creating spaces for criticism and self-criticism within the AI community, combined with changing how success is measured, suggests that meaningful reform requires fundamental shifts in how the AI community operates and evaluates its impact. The overall effect of this analysis is to fundamentally challenge the premise of using current AI systems for social good whilst providing a framework for more honest engagement with critical issues of bias, exploitation, and accountability in AI development and deployment.
Session transcript
Abeba Birhane: The type is not only found within the academic space, it also manifests in the real world. For example, large billion-dollar companies that market their product as AI-powered and fully autonomous under closer scrutiny have turned out to be using, in this case, thousands of engineers from India, and builder AI has eventually collapsed a number of weeks ago. And these algorithms, as well as academic debates, also have real impact on real actual people, and to mention one domain where we see algorithm systems applied in the social sphere, one of the good examples is the use of algorithms in welfare allocation and welfare assessment. And the recurring theme we find here is that regardless of where the algorithm is developed and deployed, people that are at the margins of society tend to be negatively, disproportionately harmed and impacted, whether it’s in India, Japan, or the Netherlands. The Netherlands is actually a really good example of attempting to develop algorithmic systems for welfare allocation in as best as possible, following over a decade of scandal. The Amsterdam municipality, for example, in trying to build this new algorithm, SmartCheck, they did everything right. In the algorithm, they excluded proxy factors such as area codes, they avoided demographic factors such as names, age, nationality, and so on, and they only used 15 criteria here to score a person’s eligibility. However, the algorithm was eventually, they also opened up for, they shared information, they opened up and actively sought external scrutiny, they sought feedback on the functionality and on the design of the algorithm, they basically did everything right. But eventually the algorithm could not tackle, again, inherently social issues, and it has to be halted, and it has to be paused. So AI systems also, even when we actively instruct them to avoid harms, to avoid discriminatory outputs and so on, they inherently encode and exacerbate societal biases, discriminations and so on. For example, in this audit that was published last year, again, they looked at state-of-the-art generative AI models. What they did was give exactly similar sentences, where the only difference is where one is presented in standard American English or standard English, for example, I’m so happy when I wake up from a bad dream because it feels too real, versus the exact same sentence with a different dialect, I’d be so happy when I woke up from a bad dream because they’d be feeling too real, but as you can see, these exact same sentences where the only difference is dialect were scored, while the first one was categorized as intelligent and brilliant, and the latter was seen as lazy, stupid, and dirty. So datasets are directly related to the performance of AI models, and again, there is a robust body of work at this stage highlighting that datasets inherently encode Western values, they encode existing societal injustices and so on. So this is from an audit myself and colleagues did a few years back. As you can see here, this is taken from a portal from the Lion dataset, this is one of the biggest, well, used to be a multi-modal dataset, here it’s a snapshot of the portal with the prompt African, and you can compare that with European. African is presented in a kind of stereotypical, cliched way, the term European is represented in almost seemingly neutral terms such as maps and flags. So it’s not just one simple example, there is also work that has done, that has audited massive scale datasets containing trillions of tokens, and the recurring theme here is that even when quality and representation in training data is generally improving and progressing, data, multilingual representation within this data remains stagnated. And also I want to emphasize that more data, more representative data, without tackling the extractive nature of the business is not also going to solve a lot of the problems. So it’s not just training data, also the entire data pipeline is really skewed towards Western values, while also negatively, disproportionately impacting minoritized identities and so on. So this is work from the Internet Health Reports by Mozilla, where they looked at benchmarking data, and as you can see, for a huge amount of research papers, the benchmarking data comes only from just a handful of elite institutions and tech companies. As I said, it’s not just training data, it’s benchmarking data, it’s also the way we clean and detoxify data that systematically punishes, again, identities and groups that are at the margins of society. So in this work, what they did was they looked at data that was excluded from a massive dataset called the CIFOR dataset to understand what kind of content was excluded. And again, as you can see here, terms and language that was used within racial minorities and LGBTQI in the trans community was disproportionately removed at a higher rate compared to terms that were used by our European counterparts. So we’ve looked at the take itself, the artifact itself, and the conclusion I want to draw from this is that given that the entire data pipeline is skewed and faces systematic challenges, and given that the fact that AI systems often fail to deliver on the promises that they are set out to be, any call to tackle complex historical and political issues using existing AI systems amounts to building with rotting wood and expecting a palace. So we’ve seen the the technological artifacts itself. It’s also important to look at the entire pipeline because AI is not really just neural networks, it’s also the labor and the resources that make it happen. So it’s also important to highlight that a lot of the labor, the data annotation content moderation, is often outsourced or delegated to people from the global south that often are paid very little for this psychologically taxing labor. Another element that we have to really remember or we have to really consider is the environmental impact of AI systems. According to a recent research by the International Energy Agency, AI or data centers are expected to consume energy around 4.5 percent of the global energy needs and this is based on a really conservative best case scenario analysis. And this is before we even get to reports that have highlighted that these corporations under-report their energy use and their CO2 emission. In fact, according to one analysis, companies such as Microsoft, Google and Meta have about 662 percent higher CO2 emission than they have officially reported. So to come back to the points I made, the initial points we mentioned earlier, on the one hand we see AI for good being used, being really proudly traded by big corporations such as Microsoft, as I highlighted here. They claim to tackle fundamental rights and sustainability and so on, but on the other hand we also see these corporations, these very corporations, really partnering with governments that are enabling war crimes. And often any criticism that highlights this irony or this controversy, claiming to do social good with AI on the one hand but also using AI to derail decades of progress, the reaction from these corporations has been retaliation, where people have been fired for pointing out the fact that Microsoft has contracts with war crime areas. The same goes with Google. Again, on the one hand there are claims that they are working to change for positive, they are working to positively change underserved communities, but on the other hand the company goes off, aligns itself with authoritarian regimes and does decades of progress, including by scrapping diversity initiatives and by signing contracts with, again, authorities that are accelerating war crimes. Again, to conclude this part of the talk, I want to highlight that, as we’ve seen, the AI space is fraught and to think that big tech corporations and AI companies that are actively derailing social progress, that are enabling war crimes and that are really derailing progress on fundamental rights, peace and democracy, to think that these corporations can be good stewards of social good is fundamentally inconceivable. I will end with just a couple of demands. One is that, given that these companies are linked with contracts, contracts with governments that are accelerating war, I would really love to use this opportunity to call on the UN and its memberships to end its partnership with these mentioned companies. I would also like, thank you, I would also like to emphasize the fact that large technological infrastructures and the cloud has to be regulated as dual use, given that, again, it’s being used for war crimes. I hope this community joins other groups that have been working on this. Finally, this really is my last slide. How do we move from a lot of the problems I pointed out to really turning this community, to turning this summit into something positive, into something that actually contributes to our actual genuine social good? We really have to rethink a lot of things, including the fact that instead of measuring impact on the number of people that have attended the summit, or the kind of CEOs that have come and spoken, or the kind of high-level guests that have attended the conference, or the kind of apps that we are building, those are really not a good measure of impact. We have to change that and we have to focus on, for example, creating space and cultivating space within this community that allows for criticism and for self-criticism. In that light, this is a good start, allowing me to take center stage here and to speak about this really is a good start. I hope it continues. It’s really critical to diverting the attention towards genuine social good. We also need, as I said, the entire pipeline from the technology itself to the ecology that embeds the AI space is fraught with extraction, illegal activities, and other problems. We really also have to tackle these issues head-on and we also have to address the glaring accountability gaps. Thank you for listening. Thank you. Thank you very much.
Abeba Birhane
Speech speed
116 words per minute
Speech length
1688 words
Speech time
868 seconds
AI companies market products as autonomous but rely on human labor, with some companies collapsing
Explanation
Large billion-dollar companies falsely market their products as AI-powered and fully autonomous when they actually rely on thousands of human engineers. This deceptive practice has led to the collapse of some AI companies when the reality is exposed.
Evidence
Example of companies using thousands of engineers from India while marketing as fully autonomous; builder AI collapsed after a few weeks
Major discussion point
Real-world impact and failures of AI systems
Topics
Economic | Legal and regulatory
Algorithmic welfare systems disproportionately harm marginalized people across different countries
Explanation
Algorithm systems applied in welfare allocation and assessment consistently show a pattern of negatively and disproportionately impacting people at the margins of society. This harmful pattern occurs regardless of where the algorithms are developed and deployed.
Evidence
Examples from India, Japan, and the Netherlands showing consistent harm to marginalized populations
Major discussion point
Real-world impact and failures of AI systems
Topics
Human rights | Development | Sociocultural
Even well-designed algorithms like Amsterdam’s SmartCheck fail to address inherent social issues
Explanation
Amsterdam municipality developed SmartCheck following best practices – excluding proxy factors, avoiding demographic data, using only 15 criteria, and seeking external scrutiny. Despite doing everything right, the algorithm still could not tackle inherently social issues and had to be halted.
Evidence
Amsterdam’s SmartCheck algorithm excluded area codes, demographic factors like names/age/nationality, used only 15 criteria, opened for external scrutiny, but still had to be paused
Major discussion point
Real-world impact and failures of AI systems
Topics
Human rights | Legal and regulatory | Development
AI systems inherently encode and exacerbate societal biases despite efforts to avoid harm
Explanation
Even when AI systems are actively instructed to avoid discriminatory outputs and harm, they still inherently encode and amplify existing societal biases and discrimination. This occurs despite conscious efforts to prevent such outcomes.
Evidence
State-of-the-art generative AI models audit showing systematic bias in language processing
Major discussion point
Bias and discrimination in AI systems
Topics
Human rights | Sociocultural
Generative AI models discriminate against non-standard English dialects, labeling them as “lazy” and “stupid”
Explanation
An audit of state-of-the-art generative AI models revealed systematic discrimination based on language dialect. Identical sentences were scored differently based solely on whether they used standard American English or alternative dialects.
Evidence
Audit showing identical sentences scored differently – standard English labeled as ‘intelligent and brilliant’ while dialect version labeled as ‘lazy, stupid, and dirty’
Major discussion point
Bias and discrimination in AI systems
Topics
Human rights | Sociocultural | Multilingualism
Training datasets encode Western values and present stereotypical representations of different cultures
Explanation
Datasets used to train AI models inherently encode Western values and existing societal injustices. This results in stereotypical and biased representations of different cultures and regions in AI outputs.
Evidence
Lion dataset audit showing ‘African’ prompt results in stereotypical, cliched representations while ‘European’ shows neutral terms like maps and flags
Major discussion point
Bias and discrimination in AI systems
Topics
Human rights | Sociocultural | Cultural diversity
Multilingual representation in datasets remains stagnated despite general improvements
Explanation
While there has been general progress in quality and representation in training data, multilingual representation within datasets has not improved and remains stagnated. This creates ongoing bias toward certain languages and cultures.
Evidence
Research on massive scale datasets containing trillions of tokens showing stagnated multilingual representation
Major discussion point
Systemic problems in AI data pipeline
Topics
Sociocultural | Multilingualism | Cultural diversity
Benchmarking data comes predominantly from elite Western institutions and tech companies
Explanation
The data used to benchmark and evaluate AI systems is heavily skewed toward a small number of elite institutions and tech companies. This creates a systematic bias in how AI performance is measured and validated.
Evidence
Mozilla Internet Health Reports research showing benchmarking data for huge amounts of research papers comes from just a handful of elite institutions and tech companies
Major discussion point
Systemic problems in AI data pipeline
Topics
Legal and regulatory | Development | Sociocultural
Data cleaning processes systematically remove content from racial minorities and LGBTQI communities at higher rates
Explanation
The processes used to clean and detoxify datasets systematically punish identities and groups at the margins of society. Content and language used by these communities is disproportionately removed compared to content from other groups.
Evidence
Analysis of CIFOR dataset exclusions showing terms and language from racial minorities and LGBTQI/trans communities removed at higher rates than European counterparts
Major discussion point
Systemic problems in AI data pipeline
Topics
Human rights | Gender rights online | Sociocultural
Data annotation and content moderation work is outsourced to poorly paid workers in the Global South
Explanation
The labor required for AI systems, including data annotation and content moderation, is often delegated to workers from the Global South. These workers are paid very little for psychologically taxing work that is essential to AI development.
Evidence
Reference to outsourced labor practices in AI development
Major discussion point
Labor and environmental exploitation
Topics
Economic | Development | Human rights
AI systems consume significant energy, with data centers expected to use 4.5% of global energy needs
Explanation
AI systems have a substantial environmental impact through their energy consumption. According to conservative estimates, data centers supporting AI are projected to consume around 4.5% of global energy needs.
Evidence
International Energy Agency research showing 4.5% global energy consumption projection based on conservative best-case scenario analysis
Major discussion point
Labor and environmental exploitation
Topics
Development | Sustainable development | Infrastructure
Tech corporations under-report their CO2 emissions by up to 662%
Explanation
Major technology corporations significantly under-report their actual energy use and carbon dioxide emissions. Analysis shows their actual emissions are dramatically higher than officially reported figures.
Evidence
Analysis showing Microsoft, Google and Meta have about 662% higher CO2 emissions than officially reported
Major discussion point
Labor and environmental exploitation
Topics
Development | Sustainable development | Legal and regulatory
Major tech companies claim to promote social good while partnering with governments enabling war crimes
Explanation
There is a fundamental contradiction between tech corporations’ public claims of working for social good and their actual business practices. These companies simultaneously promote AI for good initiatives while maintaining contracts with governments that enable war crimes.
Evidence
Microsoft claiming to tackle fundamental rights and sustainability while partnering with governments enabling war crimes
Major discussion point
Corporate hypocrisy and war crimes
Topics
Human rights | Cybersecurity | Legal and regulatory
Corporations retaliate against employees who criticize their contracts with authoritarian regimes
Explanation
When employees point out the contradictions between companies’ stated values and their actual contracts with problematic governments, corporations respond with retaliation. Workers have been fired for highlighting these ethical concerns.
Evidence
People fired for pointing out Microsoft’s contracts with war crime areas; similar patterns at Google
Major discussion point
Corporate hypocrisy and war crimes
Topics
Human rights | Economic | Legal and regulatory
Companies simultaneously promote diversity while scrapping diversity initiatives and enabling conflicts
Explanation
Tech companies demonstrate hypocrisy by publicly claiming to work for positive change in underserved communities while simultaneously aligning with authoritarian regimes and dismantling their own diversity programs. This contradictory behavior undermines decades of social progress.
Evidence
Google claiming to positively change underserved communities while aligning with authoritarian regimes and scrapping diversity initiatives
Major discussion point
Corporate hypocrisy and war crimes
Topics
Human rights | Sociocultural | Economic
UN should end partnerships with tech companies linked to war crimes
Explanation
Given the documented connections between major tech companies and governments accelerating war crimes, international organizations like the UN should terminate their partnerships with these corporations. This is necessary to maintain ethical standards and accountability.
Evidence
Companies linked with contracts with governments accelerating war
Major discussion point
Calls for reform and accountability
Topics
Human rights | Cybersecurity | Legal and regulatory
Large technological infrastructures should be regulated as dual-use technologies
Explanation
Cloud computing and large technological infrastructures should be subject to dual-use regulations because they are being used for both civilian and military purposes, including war crimes. This regulatory approach would acknowledge their potential for both beneficial and harmful applications.
Evidence
Cloud and technological infrastructure being used for war crimes
Major discussion point
Calls for reform and accountability
Topics
Legal and regulatory | Cybersecurity | Infrastructure
Success metrics should focus on genuine social good rather than attendance numbers or high-profile speakers
Explanation
Current measures of impact in AI conferences and summits are fundamentally flawed, focusing on superficial metrics like attendance numbers, CEO participation, and high-level guests rather than actual social benefit. These metrics need to be changed to reflect genuine positive impact.
Evidence
Criticism of measuring impact by number of attendees, CEOs speaking, high-level guests, or apps built
Major discussion point
Calls for reform and accountability
Topics
Development | Sociocultural | Legal and regulatory
Communities need to create space for criticism and address accountability gaps in AI development
Explanation
The AI community must cultivate spaces that allow for criticism and self-criticism to achieve genuine social good. This includes addressing the systemic problems throughout the AI pipeline and tackling accountability gaps that currently exist in AI development and deployment.
Evidence
Need to tackle extraction, illegal activities, and accountability gaps in AI pipeline
Major discussion point
Calls for reform and accountability
Topics
Legal and regulatory | Human rights | Development
Agreements
Agreement points
Similar viewpoints
Unexpected consensus
Overall assessment
Summary
This transcript contains a presentation by a single speaker (Abeba Birhane) rather than a multi-speaker discussion or debate. The speaker presents a comprehensive critique of AI systems covering systemic bias, corporate hypocrisy, environmental impact, labor exploitation, and calls for reform.
Consensus level
No consensus analysis possible as there is only one speaker presenting arguments. The speaker presents a cohesive and internally consistent critique of current AI systems and practices, calling for fundamental changes in how AI development and deployment are approached, measured, and regulated.
Differences
Different viewpoints
Unexpected differences
Overall assessment
Summary
This transcript contains a single speaker (Abeba Birhane) presenting a comprehensive critique of AI systems and their societal impacts. There are no disagreements to analyze as this is a monologue rather than a multi-speaker discussion or debate.
Disagreement level
No disagreements present – single speaker presentation. The speaker presents a unified, coherent argument about the problems with current AI systems, corporate practices, and the need for reform. All arguments align toward the same critical perspective on AI development and deployment.
Partial agreements
Partial agreements
Similar viewpoints
Takeaways
Key takeaways
AI systems consistently fail to deliver on their promises and disproportionately harm marginalized communities regardless of geographic location or implementation approach
The entire AI data pipeline is systematically biased toward Western values and excludes or misrepresents minoritized identities through training data, benchmarking, and cleaning processes
Major tech corporations engage in fundamental hypocrisy by promoting ‘AI for good’ initiatives while simultaneously partnering with governments enabling war crimes and retaliating against critics
Current AI development relies on exploitative labor practices in the Global South and has significant unreported environmental impacts
Building solutions to complex social problems with existing AI systems is equivalent to ‘building with rotting wood and expecting a palace’ due to systemic flaws in the technology and its development process
Success metrics for AI initiatives should focus on genuine social impact rather than superficial measures like attendance numbers or high-profile participation
Resolutions and action items
Call on the UN and its membership to end partnerships with tech companies that have contracts with governments accelerating war crimes
Regulate large technological infrastructures and cloud services as dual-use technologies given their application in warfare
Create and cultivate spaces within AI communities that allow for criticism and self-criticism
Address accountability gaps in AI development and deployment
Tackle extraction and illegal activities throughout the AI pipeline head-on
Unresolved issues
How to effectively address the inherent bias in AI systems when even well-designed algorithms like Amsterdam’s SmartCheck fail
How to achieve genuine multilingual and cultural representation in AI datasets beyond simply adding more data
How to restructure the AI industry to move away from its extractive business model
How to hold tech corporations accountable for under-reporting environmental impacts and energy consumption
How to implement effective oversight of dual-use AI technologies
How to measure genuine social impact in AI initiatives rather than relying on superficial metrics
Suggested compromises
None identified
Thought provoking comments
Any call to tackle complex historical and political issues using existing AI systems amounts to building with rotting wood and expecting a palace.
Speaker
Abeba Birhane
Reason
This metaphor powerfully encapsulates the fundamental structural problems with AI systems. It’s insightful because it moves beyond technical fixes to highlight that the entire foundation of current AI development is compromised by biased data, extractive practices, and systemic inequalities. The metaphor makes complex technical and social issues accessible while emphasizing the futility of expecting good outcomes from fundamentally flawed systems.
Impact
This comment serves as a pivotal turning point that shifts the discussion from identifying specific problems to questioning the entire premise of AI for social good. It provides a unifying framework that connects all the previously mentioned issues (biased datasets, labor exploitation, environmental impact) under one coherent critique of the AI industry’s foundational problems.
The Netherlands example where Amsterdam municipality ‘did everything right’ with their SmartCheck algorithm but still had to halt it because it couldn’t tackle inherently social issues.
Speaker
Abeba Birhane
Reason
This is particularly thought-provoking because it challenges the common assumption that better design and good intentions can solve AI bias. By showing how even the most well-intentioned, carefully designed system failed, it demonstrates that the problems are not merely technical but fundamentally about trying to algorithmically solve social and political issues that require human judgment and systemic change.
Impact
This example deepens the analysis by moving beyond obvious cases of AI failure to examine a ‘best case scenario’ that still failed. It introduces the crucial insight that some problems are inherently unsuitable for algorithmic solutions, regardless of how well-designed the algorithm is, thus challenging the entire premise of AI for social good initiatives.
To think that big tech corporations that are actively derailing social progress, that are enabling war crimes and that are really derailing progress on fundamental rights, peace and democracy, to think that these corporations can be good stewards of social good is fundamentally inconceivable.
Speaker
Abeba Birhane
Reason
This comment is provocative because it directly confronts the contradiction between corporate claims of social responsibility and their actual practices. It’s insightful in connecting AI development to broader geopolitical issues and corporate accountability, moving the discussion beyond technical problems to fundamental questions about power, ethics, and corporate responsibility in global conflicts.
Impact
This statement escalates the discussion to its most critical level, introducing geopolitical and ethical dimensions that reframe AI development as not just a technical or social issue, but as a matter of global justice and human rights. It challenges the audience to consider whether reform is possible or if more fundamental changes to corporate power structures are necessary.
We have to change how we measure impact – instead of measuring by number of attendees, CEOs who spoke, or apps built, we need to focus on creating space for criticism and self-criticism.
Speaker
Abeba Birhane
Reason
This insight challenges the metrics of success commonly used in tech conferences and AI initiatives. It’s thought-provoking because it suggests that the very way the AI community evaluates progress is part of the problem, prioritizing visibility and corporate participation over genuine critical engagement and accountability.
Impact
This comment provides a constructive path forward after the extensive critique, shifting the discussion toward actionable solutions. It reframes what meaningful progress would look like and implicitly critiques the current summit’s structure while offering a framework for genuine improvement in how the AI community operates.
Overall assessment
These key comments shaped the discussion by building a comprehensive, multi-layered critique that moves from specific technical problems to fundamental questions about power, accountability, and the possibility of reform within existing systems. Birhane’s presentation follows a strategic progression: starting with concrete examples of AI failures, escalating to systemic critiques of the entire AI pipeline, connecting these issues to broader geopolitical concerns, and finally offering constructive suggestions for change. The ‘rotting wood’ metaphor serves as the conceptual anchor that ties together all the disparate problems into a coherent argument against the feasibility of AI for social good under current conditions. The overall impact is to fundamentally challenge the premise of the summit itself while providing a framework for more meaningful engagement with these critical issues.
Follow-up questions
How can we better understand and address the systematic exclusion of marginalized communities’ language and content during data cleaning and detoxification processes?
Speaker
Abeba Birhane
Explanation
The research showed that terms and language used by racial minorities and LGBTQI communities were disproportionately removed from datasets at higher rates, indicating a need for deeper investigation into these filtering mechanisms
What are the actual energy consumption and CO2 emissions of AI systems given the significant under-reporting by corporations?
Speaker
Abeba Birhane
Explanation
Research suggests companies like Microsoft, Google and Meta have 662% higher CO2 emissions than officially reported, indicating a need for independent verification and more accurate environmental impact assessment
How can we develop better measures of impact for AI for good initiatives beyond attendance numbers, CEO participation, and app development?
Speaker
Abeba Birhane
Explanation
Current metrics for measuring success in AI for good initiatives are inadequate and don’t reflect genuine social impact, requiring research into more meaningful evaluation frameworks
How can large technological infrastructures and cloud services be effectively regulated as dual-use technologies?
Speaker
Abeba Birhane
Explanation
Given the use of AI systems in warfare and potential war crimes, there’s a need to research regulatory frameworks that can address the dual-use nature of these technologies
What are the long-term psychological and social impacts on workers in the Global South who perform data annotation and content moderation labor?
Speaker
Abeba Birhane
Explanation
The outsourcing of psychologically taxing AI labor to poorly paid workers in the Global South requires further investigation into the human costs of AI development
How can we address the fundamental contradiction between AI for good initiatives and corporate partnerships with entities involved in war crimes?
Speaker
Abeba Birhane
Explanation
The disconnect between stated social good objectives and actual corporate partnerships with governments enabling war crimes represents a critical area requiring policy and ethical research
Disclaimer: This is not an official session record. DiploAI generates these resources from audiovisual recordings, and they are presented as-is, including potential errors. Due to logistical challenges, such as discrepancies in audio/video or transcripts, names may be misspelled. We strive for accuracy to the best of our ability.
Related event
