Google & Responsible AI: 7 Years of Blunders, and the Way Back to Trust

Google launched its multi-modal AI model Gemini with a beautiful campaign outlining the technology’s potential. The technology can ingest inputs in multiple forms, and return outputs in multiple forms, too. This next wave of AI is, in short, incredible. Just three weeks after its launch, users of the model realized that requests for images with humans yielded only those of people with darker skin tones. A social media outcry backed by “receipts” led to Google halting the image generation portion of the service not long after its launch. This is just the latest in a series of responsible AI challenges experienced by the tech giant, which continues to be a target for AI critics and watchdogs.

Since the dawn of the “AI Era” as we know it, we’ve seen Google struggle to gain its balance as other major companies pull ahead, release models, form partnerships and, in general, enjoy success while also claiming to prioritize responsible AI. Google has had at least one major responsible AI blunder every year since 2018. The company’s inability to gain its balance appears to be largely due to a disjointed approach to, and execution of, responsible AI priorities and processes, compounded by challenges with other inadequate internal HR and communications processes. This deeper dive looks at Google’s significant responsible AI (RAI) challenges over time, examines how Google ended up as less-than-dependable in the world of RAI (and AI in general), and what Google can do to get back on track.

Note from the author: It’s important to disclose that I work for GitHub, which is owned by Microsoft - one of Google’s competitors. However, I am an AI ethicist, and I am on the side of responsible AI. I take no joy or pleasure in Google (or any of Microsoft’s competitors) failing to launch AI responsibly. If Google comes back from these challenges with a stronger stance that moves responsible AI forward (and I think that they will) that’s a win for humanity and I would welcome that success with open arms. The opinions expressed in this article and in my newsletter are my own. Now back to the article.

It should be clear that Google’s contribution to AI hasn’t been all bad, and we also don’t have a look into the innerworkings of Google’s current responsible AI work, which means we can only evaluate what’s publicly available to us. A peer of mine recently reminded me that large companies can be doing a host of massive, important, positive things and never make the news for it. It’s rare to see the internet praise a social media company for its commitment to expensive and time-consuming content moderation processes, for example. So, in the midst of the blunders, there have been positive advancements that Google can be proud of, like turning Bard into Gemini, and forming critical partnerships with opensource leaders like Hugging Face - and we should all keep these successes top-of-mind.

As of March 2024, Google’s approach to responsible AI, as shared with the public, appears to be two-pronged. The first prong includes a set of seven “objectives” for the technology, and the company’s commitment to them, which debuted after in 2018 in a document called “Artificial Intelligence at Google: our principles” after a controversy pushed the company to create principles and standard for creation, deployment and use of AI. The second prog provides guidance across four responsible AI ‘practices’ : fairness, interpretability, privacy, and safety and security. For each of these practices, Google offers a series of recommended actions that others can take when building, deploying, or using AI. As we know, words and statements can only go so far, and Google has struggled swimming up the responsible AI stream since at least 2018.

A timeline of Google’s Responsible AI Challenges:

Controversy 1:

US Government Drone Surveillance AI Controversy - May 2018

In May of 2018, Google employees raised their voices against Google’s partnership with the Pentagon , in which they would develop AI tools to analyze drone surveillance footage. Conflicting messaging further drove confusion, and employees were upset about the lack of a clear, principled stand. Google released a statement saying that the technology in question “flags images for human review” and is used for “non-offensive uses only.” This is in stark contrast to the email then head of Google Cloud Fei-Fei Li sent out stating that Google colleagues should “avoid at ALL COSTS any mention or implication of AI…weaponized AI is probably one of the most sensitized topics of AI - if not THE most. This is red meat to the media to find all ways to damage Google.”

Where it went wrong: A lack of clear guiding principles and boundaries on sensitive work (like military work), coupled with a lack of transparency and an attempt to hide the nature of the work being done (see the email from Fei-Fei Li) led to confusion for employees, causing distrust between the Google C-Suite and employees, and Google and consumers. Google created responsible AI principles only AFTER this blunder, raising questions about how Google was being guided up until the forcing-function of conflict led to the creation of principles.

Controversy 2:

External Tech Advisory Board is Launched and Immediately Dissolved - April 2019

After its challenges with the Pentagon contract, Google created and shared a list of responsible AI principles. A year after the release of the principles, Google formed the Advanced Technology External Advisory Council (ATEAC), which included a diverse group of members across business, academics, and policy. The board was dissolved shortly after it launched, due to company and public backlash at the inclusion of Heritage Foundation president Kay Coles James. According to The Verge, James is a “noted conservative figure who has openly espoused anti-LGBTQ rhetoric and, through the Heritage Foundation, fought efforts to extend rights to transgender individuals and to combat climate change.” Outcry from employees and consumers led to the dissolution of the board. A Google spokesperson stated “It’s become clear that in the current environment, ATEAC can’t function as we wanted. So we’re ending the council and going back to the drawing board.”

Where it went wrong: Creating a technology advisory council was a great idea, and incorporating diverse voices and perspectives is incredibly important. I believe this was Google’s intention. However, when selecting someone for technology council navigating complex issues around morals, ethics, and rights in technology, it is important those people not be actively working against human rights for some. While the board had not yet met or made decisions, the appointment of James signaled to the public that Google was willing to entertain the ideas and perspectives from someone who actively believes that transgender individuals should not have the same rights as everyone else. While this doesn’t violate the ethical principles set out by Google per se, it is in direct contrast to statements made by Google’s CEO about respecting human rights. This blunder significantly weakened people’s trust in Google’s ability to make meaningful moves to further the cause of responsible AI.

Controversy 3:

AI Ethics Researcher Forced Out - December 2020

In December of 2020, Timnit Gebru , then a co-lead of Google’s ethical AI team, co-authored a paper called “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” which asks, among other questions, “whether enough thought has been put into the potential risks associated with developing them and strategies to mitigate these risks.” Jeff Dean, then the head of Google AI, said the paper “didn’t meet our bar for publication” and therefore Google would not move to publish. In response to Dean’s decision not to publish or respond to the concerns the paper presented, Gebru said she would resign if Google did not meet a series of conditions for ethical and responsible AI, which Google declined to do. Gebru then asked to negotiate her final day of employment after returning from vacation, but was cut off from her Google email before she returned. The decision was met with huge protest from Googlers, 1400 of whom signed a letter of protest, alongside 1900 individuals outside of Google. The company was criticized for infringing on academic freedom and creating a chilling effect around research, due to their apparent decision to turn a blind eye to inconvenient truths. In response to Gebru’s unceremonious ouster, Jeff Dean posted an email he wrote, in which he said the paper ignored current research. The paper was leaked, which revealed that it was worked on by no less than six decorated researchers and cited 128 articles and papers - a notably long list - which calls into question Dean’s claims.

Google CEO Sundar Pichai responded to the firing saying , “I’ve heard the reaction to Dr. Gebru’s departure loud and clear: it seeded doubts and led some in our community to question their place at Google. I want to say how sorry I am for that, and I accept the responsibility of working to restore trust.” The email was not met with a positive response.

Where it went wrong: Dr. Gebru’s firing highlighted fractures in Google’s projected image of alignment on responsible AI. Google’s unceremonious firing of a researcher because of philosophical disagreements gives the impression that Google is looking for researchers to tell the story they want to project, rather than present the results of the research and potential solutions to the challenges the company faces in AI development. Sundar Pichai did not apologize directly for the way Gebru was fired, but apologized for how it made other Googlers feel. The entire saga throws in stark relief the differences between Google’s desired brand image, and the lived experience of employees working to make the world safer.

Controversy 4:

AI Ethics Researcher Fired (again) - February 2021 Shortly after Gebru’s firing, Margaret Mitchell, Gebru’s co-lead of the ethical AI team at Google, was locked out of her corporate account “pending investigation.” In February of 2021 Google shared in a statement “we confirmed that there were multiple violations of our code of conduct, as well as of our security policies, which included exfiltration of confidential business-sensitive documents and private data of other employees.” Mitchell is suspected to have attempted to find evidence related to Gebru’s ouster. On February 19th, 2021, Margaret Mitchell posted a tweet that simply said, “I’m fired.” It was only after Mitchell’s firing that Google shared that an internal investigation after Gebru’s departure would result in changed policies but mentioned nothing of the handling of Mitchell’s firing.

Where it went wrong: Whether there were violations or not, the firing of Margaret Mitchell gave the impression that Google had something to hide. While they shared that Mitchell had violated policy, the way Google managed the situation left people wondering what was happening in the C-suite, and why the firing of ethical AI team leads was so disastrous. The statement from Google indicating they had flubbed Gebru’s firing underscores Google’s need for ethical perspectives on process and did not instill confidence in Google’s ability to consistently deliver on promises for responsible and ethical AI, given that their non-AI policies led to such a poorly handled firing for to high-profile employees.

Controversy 5:

Blake Lemoine and “Sentient” AI - June 2022

Blake Lemoine was an engineer working in Google’s Responsible AI Organization when he began chatting with Google’s LaMDA large language model and felt that the model was displaying signs of sentience, which he escalated within Google and eventually shared in an interview with the Washington Post. Lemoine’s interactions with the LLM, escalations and interview took place less than two years after Gebru and Mitchell warned Google that they were creating AI which possessed, among other risks, a high chance of fooling individuals into thinking they were speaking with a sentient being. What makes this case compelling is not that someone was fooled, but that someone who knows the inner workings of the technology was fooled. Blake Lemoine was put on administrative leave shortly before the Washington Post published his story and was let go from the company in July of 2022.

Where it went wrong: The details of this case and its occurrence so soon after the dismissal of the ethicists who warned Google about this very problem are what makes it so damning. While this appears to be a case of a single rogue employee taking their own feelings and perspectives to the public, this story did resonate with people and surfaced some critical questions such as, “if an engineer on the RAI team could be fooled, can we all be fooled by AI?” People didn’t want to write Lemoine off as a quirky Google employee - they wanted answers from Google to help them understand how this could happen and what Google was doing to fix it. That level of clarity and resolution never came.

Controversy 6:

Bard Launch Failure - February 2023

Less than a year after the Blake Lemoine challenges, and less than 4 months after the release of OpenAI’s ChatGPT, Google released Bard, an AI chatbot meant to converse, answer questions, and assist users. The release of Bard to the world in the form of a demo didn’t instill confidence in Google’s new AI, because the demo showcasing Bard’s power also showed Bard sharing an answer that was factually incorrect. Google was hastily corrected by the internet at large and seemed to fumble their opportunity for a comeback by releasing the following statement: “This highlights the importance of a rigorous testing process, something that we’re kicking off this week with our Trusted Tester program. We’ll combine external feedback with our own internal testing to make sure Bard’s responses meet a high bar for quality, safety and groundedness in real-world information.”

Where it went wrong: The public response to this blunder wasn’t one of shock that AI could be incorrect - people generally know that AI is not perfect. The tone of surprise from the public was that, if the demo was pre-recorded, Google did not fact-check it, and if it was a live demo, they did not state up-front that AI can be incorrect and should always be checked. The damage control on the back end seemed to offload the blame to AI for the incorrect answer, but did not state that Google was responsible for providing the correct warnings up top. Ultimately, this was a hasty coverup to try and make the situation seem like it wasn’t a big deal, and once again Google was caught putting together checks and balances (enter the Trusted Tester program) after an ethical or responsible AI blunder had occurred, rather than in anticipation of ethical and responsible AI challenges.

Controversy 7:

Gemini Generative AI Photo Controversy - February 2024

Shortly after Google released its multi-modal AI Gemini, people started noticing that the photos being generated weren’t quite correct. According to Vox , this started with a tweet from a user on X who had asked for photos of the Founding Fathers of America, Vikings, and a pope. The AI returned images of all dark-skinned individuals, both men and women. The photos were clearly and alarmingly inaccurate given the historical information available to us. The tweet went viral, and people tried the AI for themselves. Once again, a Google AI move was going viral for the wrong reasons. Google released a blog post and response within two days of the viral tweet, sharing clearly what the intention was, what went wrong, and where they’ll go from here.

Where it went wrong: Google shared up-front that the AI might be inaccurate, but it did so in a somewhat long “opt-in” page which included privacy and configuration settings “above the fold” and “things to know” (which included the statement of experimental technology, etc.) below the fold. However, Google did follow up quickly with more than a statement - they posted an entire blog post which shared in plain, understandable language what had happened, and what they were going to do next. While the Gemini launch wasn’t a full win for Google, their response to the controversy was a win, and is signaling that Google is getting back on track with responsible AI.

Where Google can go from here:

The follow-up after Gemini’s unfair photo controversy shows a side of Google that is willing to acknowledge the blunder, take responsibility, explain what happened, and commit to a solution - more of this, please. Below I’ve suggested an incomplete list of practices Google can consider implementing both externally and internally to emerge from these challenges stronger. Everyone loves a comeback story, so if Google played their cards right, they could turn the tides in their favor.

Externally:

• Make the notice about AI inaccuracy the number one thing people read when engaging with generative AI products (i.e. Gemini). It could be moved to the top of the page, or you could require a second page with an attestation to ensure that people understand that AI can be inaccurate and should be checked. It should be easy to find, easy to read and easy to understand.
• Establish “RAI voices” from within the company. At present, for every individual blunder a new voice has provided a statement. HR leads, CEOs, leaders of the Google Cloud, leaders in Engineering. This fractured approach to communication probably makes sense to the internal workings of Google - but the people with whom you are communicating are not in Google. We’re on the outside of your verticals and orgs. Establishing two or three trusted voices from engineering, responsible AI and the C-Suite to communicate about RAI - good or bad - could help increase trust. These people should be listed or acknowledged publicly somewhere as the voice-pieces for a larger effort in RAI.
• Release honest, thorough, and thoughtful statements to the public. In some cases, a canned, legal-approved statement is required by Google - especially for very tricky situations related to hiring/firing/investigations, etc. However, in many of the RAI cases listed above, a canned legal response wasn’t necessary. Google could have been more honest, more thorough, and more thoughtful about what it was communicating to a waiting and worried public. Especially in this sea change moment, Google should be communicating clearly and often, in terms everyone can understand. “I take responsibility and I will work harder to ensure we do better” is not thoughtful or thorough. What are you taking responsibility for? What will you do to ensure things are better? What does better look like? All these questions and more should be addressed.
• Re-establish your tech board with a wider variety of voices who are committed to human rights. The tech board was a great idea, but it was poorly executed because it did not appear that the members were vetted, which gave the impression that the board selection process was politically swayed. (Whether this is true or not, we don’t know. But there was enough skepticism that people felt something was amiss, leading to the demise of the board.) What people worry about in the age of AI is not limited to job-loss or algorithmically reinforced systemic unfairness (although these are massive and justified concerns) but a fear more broadly of a loss of power. In the previous boards’ construction, a group of already-powerful, on-the-record and well-known people were assembled to help determine the fate of technology for Google. While having credentialed and knowledgeable people at the table is important, those people should represent a wide variety of experiences, cultures, fields, and perspectives. Every board member should be committed to human rights for all. Period. Google could also allow an open forum where employees and the public can share their fears/perspectives/hopes for AI with the board, and the board can review them as part of their decision-making process. This will help re-establish the lost trust with the public - a lesson we’ve learned from Leslie Knope.
• Review your responsible AI principles on a regular basis and share with the public your updates. Responsible AI principles are not carved in stone. They are living, breathing guides for the type of AI we want to build and the way we want to build it. They should be reviewed every two years at least - especially since Google’s RAI principles were formed in 30 days amidst controversy. Edits should be made in the form of updates to current principles, removal of some principles and addition of new principles. The edits and updates should be shared, with commentary and justification for the change, with the public.
• Establish guidelines for sharing AI research and share research that meets the established criteria - even if it doesn’t lend to Google’s image. We know a few things: Google’s leadership has a responsibility to shareholders. We also know that Google, a company behind some of the most critical and world-altering internet-based innovations has a duty to the very public that made the company a global success. Publishing research as a company is tricky because the CEO of the company has a duty to both the shareholders and the public, and research has been known to impact both groups positively and negatively. Why do companies need to publish research in the first place? They don’t have to, but because they have the resources and access to the systems in question, it behooves companies to do research to understand risk, explore new innovations, etc. The decision to publish or not publish should not be decided by PR professionals, because they could stifle important research from getting out into the world for fear of the risk associated with it. Research publishing should be decided by a pre-determined set of standards, such as how much this research will positively impact the world.

Internally:

• Update your policies to ensure that people are not punished for taking a stand. Google has made statements about updating policies to ensure that people can raise concerns, and the difference in Gebru and Lemoine’s firings seem to indicate a change in approach. They should continue to evaluate and update these policies.
• Provide a forum for people to voice concerns about AI. If Google does not have one already, they should establish a forum for people to share their AI concerns. The forum should allow for opt-in anonymity, and the concerns shared should be captured, catalogued, and reviewed regularly to understand internal sentiment, whether it’s growing and what might be influencing the concerns. This data should then be used to make RAI decisions. When people feel they are not being heard or taken seriously, they take their story public. Providing an open forum for AI concerns will serve to protect AI efforts and innovation, as well as the company and, by extension, humanity.
• Train all employees on the dangers of AI. If Google has not already prioritized this, it should create an ever-evolving training that alerts employees to the dangers of AI - bias, hallucinations/confabulations, etc. This training could be an attaché to the security training that all employees must take annually which will reduce cost and secure the training as a critical part of Google’s trust culture.
• Establish an AI/RAI incident response team. When a crisis arises related to AI and therefore RAI, a specific, pre-determined group of individuals should be convened to navigate the crisis. There should be a framework followed for identifying the challenge, the threat level (to the public, to the product, to institutions, to the company), the impact, the steps to mitigate and documenting the process. It should be a simple format to follow, but one that is thorough. A post-mortem should always be done with a broader group than the original contributors so that the process can be updated to be more efficient and thorough. Parts of this post-mortem can be shared with the public as a way of establishing trust. “This is what happened, this is how we approached it, this is what we did to resolve, this is what we will change moving forward.”

Responsible AI is still very much in its nascent stages as a field. Frameworks are proliferating and everyone has an opinion on what should have, should or could happen to make AI safer, better or more positively impactful. My suggestion to Google, or any company building or deploying AI, is that you view your duty to responsible AI as a holistic one. If AI establishes, impacts or is impacted by a process or procedure, it needs a responsible AI consideration. (Hint: in this AI era, almost every process or procedure is impacted by AI.) To Google I would say that the journey of growth the company has been on these last 7 years has no-doubt been challenging, but it’s been fascinating to review the growth as I’ve written this analysis.

In my opinion, Google isn't out of the running for leadership in the AI era. These blunders have set the stage for responsible AI leadership if the company can make a series of changes that will allow them to capitalize on the chance to lead. Leading in AI without leading in responsible AI is a fools errand, and I would advise Google to view AI leadership and responsible AI leadership as one in the same.

As I mentioned above, this is an incomplete list and I would welcome other suggestions in the comments or via email - [email protected]