Ethics sweeping
I was going to title this post “Google and Women” but this would be a very limiting angle for what is a critical broader issue.
It’s too easy to put thorny topics under the feminist cover when there are multiple perspectives to be considered including ethics, data bias, social prejudice, minorities. Just focusing on gender can trigger an over-saturation and numbness in readers causing a diametrical reaction of aversion, not attention. Regardless, we must not stop raising flags of social relevance with the risk of becoming complacent.
This week the media wrapped its tentacles around the story of Dr. Timnit Gebru, a co-leader of Google’s Ethical Artificial Intelligence team who was suddenly dismissed under questionable circumstances. [Google workers reject company’s account of AI researcher’s exit as anger grows | Google | The Guardian]
Dr. Gebru is a black woman, which caused an uproar about Google’s sexist, prejudiced and toxic employment environment punishing diversity.
Fair to say, the details of the story are mixed - was she fired, did she resign under an ultimatum she posed to Google or, was she removed because of her exposing report on ethics in Google’s AI language modelling algorithms?
What was Google’s gripe?
The dispute was over an AI Ethics paper that Dr. Gebru co-wrote with 7 other researchers, titled “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”.
Google blocked the paper on grounds of it being incomplete, poorly referenced (with 120+ references) and, not meeting research standards. And Dr. Gebru was removed.
Luckily for us, the paper still circulated sufficiently around the media for its content to become known and discussed.
The main points the authors made were ([We read the paper that forced Timnit Gebru out of Google. Here’s what it says. | MIT Technology Review]:
1. Environment: Training massive language models consumes vast amounts of processing power, which has a significant environmental impact. The study found that training a language model with an advanced methodology called “neural architecture search” produces about the same carbon footprint as the lifespan of 5 average cars. Training a language model for Google’s search engine backbone has the carbon footprint of a return flight between San Fransisco and New York. Models are trained and retrained numerous times.
2. Transparency: Language models are becoming so large and incomprehensible by humans that transparency and explainability are out of reach and prone to incorporating dangerous language as it spreads through the media. AI experts and machines collect data from all over the internet to train models, data which includes racist, sexist and abusive language, all going into model training without discernment. Auditing for biases is becoming impossible when a model might have 200 million parameters and 200 millions of examples.
3. Misinterpretation and misinformation: There are increasing challenges in interpreting meaning and nuances, especially in non-English languages. It can lead to simple misunderstandings or significant danger for people (as in the famous story of the Palestinian whose peaceful message was mistranslated in Hebrew and he was subsequently arrested [Facebook translates ‘good morning’ into ‘attack them’, leading to arrest | Facebook | The Guardian] and polarising spread of misinformation.
The argument of the researchers also aimed to highlight the increased harm to marginalised socio-economic and ethnic groups and the direct benefit of Google’s success to the wealthy economies. In essence, continuing biases and prejudices through exponentially growing but inexplainable technology.
More than just a battle with diversity
We can look at this as a clear and unpleasant example of sexist and racist behaviour on Google’s part, but to me this incident is only a symptom of a more disturbing issue.
What becomes apparent is Google’s (and its parent company Alphabet) general refusal to admit and address ethical gaps in their business model and show their commitment to a better, and not divisive, society.
Luckily, if we can say that, the story gained media attention because the researcher in question was a black woman. If it had been a white man, it wouldn’t have been so interesting. It’s a double-edged sword.
Thousands of Google and tech employees have protested Google’s reaction, which brings hope that people within the industry are distancing themselves from the corporate giant and demonstrating their own value system.
Twitter is ripe with ex-Googlers that rebuff the argument over the research quality of the paper.
“This is such a lie. It was part of my job on the Google PR team to review these papers. Typically we got so many we didn’t review them in time or a researcher would just publish & we wouldn’t know until afterwards. We NEVER punished people for not doing proper process.” (William Fitzgerald)
Regardless of the specifics of the dismissal (he said, she said), the world is siding with David and not Goliath, and rightfully so. This is the only way to bring issues of power and ethics to the attention of the masses and trigger change.
In a paradox twist, Google shows its resistance to address ethical behaviour by behaving unethically to its employees. Google’s CEO Sundar Pichai has apologised about the handling of the case but has not admitted any mistake or willingness to review and reinstate the researcher.
Criticising the ethical and environmental impact of Google’s lifeline - the language models that fuel the search engine - is like leading a dictator coup.
Employment diversity and the Search Engine debacle
Of course, this isn’t the first time that Google’s ethical (or lack of) behaviour has been put under the microscope of sexism and racism.
Google’s image search engine results on gender and minority groups have also been under fire in the last few years.
Image search results and automated search terms completion has a huge impact on how people perceive gender and society - affirming stereotypes, inequity, identity, job prospects, family roles and, social behaviour.
It has been argued that Google’s failure to show fairness across gender and race within search results is directly related to Google’s lack of diversity among the company itself. It’s essentially a white boys’ club.
Women reportedly represent 32% of Google’s workforce (23% in Tech) where black women have a 1.2% share (among black men it’s only a fraction higher at 1.8%). For Latinx employees, the ratio is 1.7% and 3.5% for women and men respectively, and Asian employees are distributed at 12% and 25% for women and men (based on a 2018 report in the Washington Post).
If you enter “Woman” in Google’s Image search you will see some filter search suggestions at the top. I get beautiful, beach, most beautiful, pregnant, chest, natural, middle aged, face, strong, black.
For “Man” I get handsome, face, hair style, black, suit, old, beard, happy, beautiful, strong.
Granted, both results are very appearance oriented but women are still predominantly represented for their physical and reproductive attributes.
Social media researcher and UCLA professor Safiya Noble has been writing on search engine stereotyping and prejudice for about 10 years. She has strongly criticised Google and other search engines for misusing their power for controlling how society thinks, a topic that has gained momentum through her 2018 book “Algorithms of Oppression”, accompanying articles and Ted Talk [How biased are our algorithms?].
Her work has been the basis of many other articles and media research studies.
“As a professor and researcher of digital cultures, [I have found that] a lack of care and investment by tech companies towards users who are not white and male allows racism and sexism to creep into search engines, social networks and other algorithmic technologies. “ (Jonathan Cohn, The Conversation 2018)
I first read her article from 2012 [“Missed Connections: What Search Engines Say About Women” (Spring 2012) | Safiya Umoja Noble, Ph.D.] about 6 months ago, and to be honest, I had given this so little thought. I was just another user unknowingly participating in a global social experiment.
Among many things, she has exposed:
Google sexualisation of Black and Latino women image search results,
prejudiced autocomplete function for search terms (for example if you enter “Black women are…” you got angry, loud, annoying, mean, insecure, lazy),
search results focusing on physical appearances (in the image search results filter bar you are offered ‘useful’s such as attractive, skinny, pregnant)
favouring white men in successful, powerful stereotypes.
The fear is not only how we, as adults, interpret this, but what happens to the new generation of 10-12 year olds now learning how to use Google and use their results?
Computer says No
“Considering Google’s status as a monopoly … it is surprisingly muted when it comes to these vital areas it must improve on. The same is true for virtually every other major tech company…”
So what is Google doing about it?
Interestingly, in the last two years there has been a significant shift in how Google displays results, suspiciously coinciding with the publication of Noble’s book and subsequent public discussion.
However, it seems that rather than improving their algorithms and addressing the blaring bias, they have actually swept functionality under the carpet by:
Reducing the autocompletion results offered
Removing the image filter options for certain image search terms (e.g. “Black men” and “Black women” don’t show up any additional search filters)
After a scandal erupted in 2015 over Google’s image recognition algorithm which turned out to be classifying photos of black people as “gorillas", embarrassed Google took action and deleted all images of gorillas and primates from their training data. The result - they fixed the offensive behaviour but they were no longer able to identify gorillas in their image search. In 2018 this still hadn’t been resolved, so that black people and gorillas could coexist with reliable distinction within Google’s ecosystem. [When It Comes to Gorillas, Google Photos Remains Blind | WIRED]
Do no evil?
Without a doubt Artificial Intelligence involving language and images is an extremely difficult area in terms of sourcing vast quantities of data (unbiased, diverse and clean), developing highly accurate algorithms and having the frameworks in place to validate results.
The equally important objective (next to making money, can’t ignore that) of any globally adopted technology should be to minimise social, political and physical harm.
However, since their first motto of “don’t be evil” and growing into a digital monster, Google has laid its paws on the world’s data and sadly on the future of its citizens. It has become too comfortable with its lucrative business model and our dependency on its services to feel the need to make significant shifts towards a more ethical and transparent ideology.
It feels like Google is hoping that these issues will just go away or that we can’t do anything about it anyway. But they won’t and we can.
We just have to be willing to change.
As with any system, it is important not to just accept the status quo of dominance and convenience, but to be critical and take charge of our own ethical values.
There are great alternatives to Google search and Google products, and an increasing number of people sick of Google’s authoritarian regime are moving over.
Personally I am a huge fan (and user) of Protonmail, a secure Swiss email provider (soon with Calendar and Cloud storage functionality), and QWANT - "the search engine that respects your privacy".
I don’t use GoogleDrive for Photo storage or sensitive documents.
It is no secret (anymore) how influential Youtube is, through its recommendation and attention-grabbing algorithm, in creating polarisations in society with a detrimental impact.
With 4 billion searches executed every day by 4 billion people around the world, there is an obvious social responsibility that has to be demanded of Google, based on how search results can influence social psychology and global events.
Even if this means a slight drop in revenue. Alphabet's revenue in 2019 was reported at $162 billion, the bulk of which comes from search engine advertising (AdWords). They are not fighting for survival, but we are.
I can’t understand why, if you have such a wide grip on the social psyche you don’t use it to unite and engender tolerance, equity and acceptance, rather than to divide.
So why, oh why not use your amassed power to NOT do evil?
Interesting Links and References:
Google: employee gender distribution by department 2020 | Statista
Google’s algorithms discriminate against women, people of colour – a simple image search proves it
How men and women appear in online job image searches | Pew Research Center
Google’s Algorithm: History of Racism Against Black Women | Time
Comments