How we Use A.I.: Aligning our Responsibilities and Commitment with our Core Values
Oomph has been quiet about our excitement for artificial intelligence (A.I.). While the tech world has exploded with new A.I. products, offerings, and add-ons to existing product suites, we have been formulating an approach to recommend A.I.-related services to our clients.
One of the biggest reasons why we have been quiet is the complexity and the fast-pace of change in the landscape. Giant companies have been trying A.I. with some loud public failures. The investment and venture capitalist community is hyped on A.I. but has recently become cautious as productivity and profit have not been boosted. It is a familiar boom-then-bust of attention that we have seen before — most recently with AR/VR after the Apple Vision Pro five months ago and previously with the Metaverse, Blockchain/NFTs, and Bitcoin.
There are many reasons to be optimistic about applications for A.I. in business. And there continue to be many reasons to be cautious as well. Just like any digital tool, A.I. has pros and cons and Oomph has carefully evaluated each. We are sharing our internal thoughts in the hopes that your business can use the same criteria when considering a potential investment in A.I.
Using A.I.: Not If, but How
Most digital tools now have some kind of A.I. or machine-learning built into them. A.I. has become ubiquitous and embedded in many systems we use every day. Given investor hype for companies that are leveraging A.I., more and more tools are likely to incorporate A.I.
This is not a new phenomenon. Grammarly has been around since 2015 and by many measures, it is an A.I. tool — it is trained on human written language to provide contextual corrections and suggestions for improvements.
Recently, though, embedded A.I. has exploded across markets. Many of the tools Oomph team members use every day have A.I. embedded in them, across sales, design, engineering, and project management — from Google Suite and Zoom to Github and Figma.
The market has already decided that business customers want access to time-saving A.I. tools. Some welcome these options, and others will use them reluctantly.
The Risks that A.I. Pose
Every technological breakthrough comes with risks. Some pundits (both for and against A.I. advancements) have likened its emergence to the Industrial Revolution of the early 20th century. And a high-level of positive significance is possible, while the cultural, societal, and environmental repercussions could also follow a similar trajectory.
A.I. has its downsides. When evaluating A.I. tools as a solution to our client’s problems, we keep this list of drawbacks and negative effects handy, so that we may review it and think about how to mitigate their negative effects:
- A.I. is built upon biased and flawed data
- Bias & flawed data leads to the perpetuation of stereotypes
- Flawed data leads to Hallucinations & harms Brands
- Poor A.I. answers erode Consumer Trust
- A.I.’s appetite for electricity is unsustainable
We have also found that our company values are a lens through which we can evaluate new technology and any proposed solutions. Oomph has three cultural values that form the center of our approach and our mission, and we add our stated 1% For the Planet commitment to that list as well:
- Smart
- Driven
- Personal
- Environmentally Committed
For each of A.I.’s drawbacks, we use the lens of our cultural values to guide our approach to evaluating and mitigating those potential ill effects.
A.I. is built upon biased and flawed data
At its core, A.I. is built upon terabytes of data and billions, if not trillions, of individual pieces of content. Training data for Large Language Models (LLMs) like Chat GPT, Llama, and Claude encompass mostly public content as well as special subscriptions through relationships with data providers like the New York Times and Reddit. Image generation tools like Midjourney and Adobe Firefly require billions of images to train them and have skirted similar copyright issues while gobbling up as much free public data as they can find.
Because LLMs require such a massive amount of data, it is impossible to curate those data sets to only what we may deem as “true” facts or the “perfect” images. Even if we were able to curate these training sets, who makes the determination of what to include or exclude?
The training data would need to be free of bias and free of sarcasm (a very human trait) for it to be reliable and useful. We’ve seen this play out with sometimes hilarious results. Google “A.I. Overviews” have told people to put glue on pizza to prevent the cheese from sliding off or to eat one rock a day for vitamins & minerals. Researchers and journalists traced these suggestions back to the training data from Reddit and The Onion.
Information architects have a saying: “All Data is Dirty.” It means no one creates “perfect” data, where every entry is reviewed, cross-checked for accuracy, and evaluated by a shared set of objective standards. Human bias and accidents always enter the data. Even the simple act of deciding what data to include (and therefore, which data is excluded) is bias. All data is dirty.
Bias & flawed data leads to the perpetuation of stereotypes
Many of the drawbacks of A.I. are interrelated — All data is dirty is related to D.E.I. Gender and racial biases surface in the answers A.I. provides. A.I. will perpetuate the harms that these biases produce as they become easier and easier to use and more and more prevalent. These harms are ones which society is only recently grappling with in a deep and meaningful way, and A.I. could roll back much of our progress.
We’ve seen this start to happen. Early reports from image creation tools discuss a European white male bias inherent in these tools — ask it to generate an image of someone in a specific occupation, and receive many white males in the results, unless that occupation is stereotypically “women’s work.” When AI is used to perform HR tasks, the software often advances those it perceives as males more quickly, and penalizes applications that contain female names and pronouns.
The bias is in the data and very, very difficult to remove. The entirety of digital written language over-indexes privileged white Europeans who can afford the tools to become authors. This comparably small pool of participants is also dominantly male, and the content they have created emphasizes white male perspectives. To curate bias out of the training data and create an equally representative pool is nearly impossible, especially when you consider the exponentially larger and larger sets of data new LLM models require for training.
Further, D.E.I. overflows into environmental impact. Last fall, the Fifth National Climate Assessment outlined the country’s climate status. Not only is the U.S. warming faster than the rest of the world, but they directly linked reductions in greenhouse gas emissions with reducing racial disparities. Climate impacts are felt most heavily in communities of color and low incomes, therefore, climate justice and racial justice are directly related.
Flawed data leads to “Hallucinations” & harms Brands
“Brand Safety” and How A.I. can harm Brands
Brand safety is the practice of protecting a company’s brand and reputation by monitoring online content related to the brand. This includes content the brand is directly responsible for creating about itself as well as the content created by authorized agents (most typically customer service reps, but now AI systems as well).
The data that comes out of A.I. agents will reflect on the brand employing the agent. A real life example is Air Canada. The A.I. chatbot gave a customer an answer that contradicted the information in the URL it provided. The customer chose to believe the A.I. answer, while the company tried to say that it could not be responsible if the customer didn’t follow the URL to the more authoritative information. In court, the customer won and Air Canada lost, resulting in bad publicity for the company.
Brand safety can also be compromised when a 3rd party feeds A.I. tools proprietary client data. Some terms and condition statements for A.I. tools are murky while others are direct. Midjourney’s terms state,
“By using the Services, You grant to Midjourney […] a perpetual, worldwide, non-exclusive, sublicensable no-charge, royalty-free, irrevocable copyright license to reproduce, prepare derivative works of, publicly display, publicly perform, sublicense, and distribute text and image prompts You input into the Services”
Midjourney’s Terms of Service Statement
That makes it pretty clear that by using Midjourney, you implicitly agree that your data will become part of their system.
The implication that our client’s data might become available to everyone is a huge professional risk that Oomph avoids. Even using ChatGPT to provide content summaries on NDA data can open hidden risks.
What are “Hallucinations” and why do they happen?
It’s important to remember how current A.I. chatbots work. Like a smartphone’s predictive text tool, LLMs form statements by stitching together words, characters, and numbers based on the probability of each unit succeeding the previously generated units. The predictions can be very complex, adhering to grammatical structure and situational context as well as the initial prompt. Given this, they do not truly understand language or context.
At best, A.I. chatbots are a mirror that reflects how humans sound without a deep understanding of what any of the words mean.
A.I. systems are trying its best to provide an accurate and truthful answer without a complete understanding of the words it is using. A “hallucination” can occur for a variety of reasons and it is not always possible to trace their origins or reverse-engineer them out of a system.
As many recent news stories state, hallucinations are a huge problem with A.I. Companies like IBM and McDonald’s can’t get hallucinations under control and have pulled A.I. from their stores because of the headaches they cause. If they can’t make their investments in A.I. pay off, it makes us wonder about the usefulness of A.I. for consumer applications in general. And all of these gaffes hurt consumer’s perception of the brands and the services they provide.
Poor A.I. answers erode Consumer Trust
The aforementioned problems with A.I. are well-known in the tech industry. In the consumer sphere, A.I. has only just started to break into the public consciousness. Consumers are outcome-driven. If A.I. is a tool that can reliably save them time and reduce work, they don’t care how it works, but they do care about its accuracy.
Consumers are also misinformed or have a very surface level understanding of how A.I. works. In one study, only 30% of people correctly identified six different applications of A.I. People don’t have a complete picture of how pervasive A.I.-powered services already are.
The news media loves a good fail story, and A.I. has been providing plenty of those. With most of the media coverage of A.I. being either fear-mongering (“A.I. will take your job!”) or about hilarious hallucinations (“A.I. suggests you eat rocks!”), consumers will be conditioned to mistrust products and tools labeled “A.I.”
And for those who have had a first-hand experience with an A.I. tool, a poor A.I. experience makes all A.I. seem poor.
A.I.’s appetite for electricity is unsustainable
The environmental impact of our digital lives is invisible. Cloud services that store our lifetime of photographs sound like featherly, lightweight repositories that are actually giant, electricity-guzzling warehouses full of heat-producing servers. Cooling these data factories and providing the electricity to run them are a major infrastructure issue cities around the country face. And then A.I. came along.
While difficult to quantify, there are some scientists and journalists studying this issue, and they have found some alarming statistics:
- Training GPT-3 required more than 1,200 MWh which led to 500 metric tons of greenhouse gas emissions — equivalent to the amount of energy used for 1 million homes in one hour and the emissions of driving 1 million miles. GPT-4 has even greater needs.
- Research suggests a single generative A.I. query consumes energy at four or five times the magnitude of a typical search engine request.
- Northern Virginia needs the equivalent of several large nuclear power plants to serve all the new data centers planned and under construction.
- In order to support less consumer demand on fossil fuels (think electric cars, more electric heat and cooking), power plant executives are lobbying to keep coal-powered plants around for longer to meet increased demands. Already, soaring power consumption is delaying coal plant closures in Kansas, Nebraska, Wisconsin, and South Carolina.
- Google emissions grew 48% in the past five years in large part because of its wide deployment of A.I.
While the consumption needs are troubling, quickly creating more infrastructure to support these needs is not possible. New energy grids take multiple years and millions if not billions of dollars of investment. Parts of the country are already straining under the weight of our current energy needs and will continue to do so — peak summer demand is projected to grow by 38,000 megawatts nationwide in the next five years.
While a data center can be built in about a year, it can take five years or longer to connect renewable energy projects to the grid. While most new power projects built in 2024 are clean energy (solar, wind, hydro), they are not being built fast enough. And utilities note that data centers need power 24 hours a day, something most clean sources can’t provide. It should be heartbreaking that carbon-producing fuels like coal and gas are being kept online to support our data needs.
Oomph’s commitment to 1% for the Planet means that we want to design specific uses for A.I. instead of very broad ones. The environmental impact of A.I.’s energy demands is a major factor we consider when deciding how and when to use A.I.
Using our Values to Guide the Evaluation of A.I.
As we previously stated, our company values provide a lens through which we can evaluate A.I. and look to mitigate its negative effects. Many of the solutions cross over and mitigate more than one effect and represent a shared commitment to extracting the best results from any tool in our set
Smart
- Limit direct consumer access to the outputs of any A.I. tools, and put a well-trained human in the middle as curator. Despite the pitfalls of human bias, it’s better to be aware of them rather than allow A.I. to run unchecked
- Employ 3rd-party solutions with a proven track-record of hallucination reduction
Driven
- When possible, introduce a second proprietary dataset that can counterbalance training data or provide additional context for generated answers that are specific to the client’s use case and audience
- Restrict A.I. answers when qualifying, quantifying, or categorizing other humans, directly or indirectly
Personal
- Always provide training to authors using A.I. tools and be clear with help text and microcopy instructions about the limitations and biases of such datasets
1% for the Planet
- Limit the amount of A.I. an interface pushes at people without first allowing them to opt in — A.I. should not be the default
- Leverage “green” data centers if possible, or encourage the client using A.I. to purchase carbon offset credits
In Summary
While this article feels like we are strongly anti-A.I., we still have optimism and excitement about how A.I. systems can be used to augment and support human effort. Tools created with A.I. can make tasks and interactions more efficient, can help non-creatives jumpstart their creativity, and can eventually become agents that assist with complex tasks that are draining and unfulfilling for humans to perform.
For consumers or our clients to trust A.I., however, we need to provide ethical evaluation criteria. We can not use A.I. as a solve-all tool when it has clearly displayed limitations. We aim to continue to learn from others, experiment ourselves, and evaluate appropriate uses for A.I. with a clear set of criteria that align with our company culture.
To have a conversation about how your company might want to leverage A.I. responsibly, please contact us anytime.
Additional Reading List
- “The Politics of Classification” (YouTube). Dan Klyn, guest lecture at UM School of Information Architecture. 09 April 2024. A review of IA problems vs. AI problems, how classification is problematic, and how mathematical smoothness is unattainable.
- “Models All the Way Down.” Christo Buschek and Jer Thorp, Knowing Machines. A fascinating visual deep dive into training sets and the problematic ways in which these sets were curated by AI or humans, both with their own pitfalls.
- “AI spam is already starting to ruin the internet.” Katie Notopoulos, Business Insider, 29 January 2024. When garbage results flood Google, it’s bad for users — and Google.
- Racial Discrimination in Face Recognition Technology, Harvard, 24 October 2020. The title of this article explains itself well.
- Women are more likely to be replaced by AI, according to LinkedIn, Fast Company, 04 April 2024. Many workers are worried that their jobs will be replaced by artificial intelligence, and a growing body of research suggests that women have the most cause for concern.
- Brand Safety and AI, Writer.com. An overview of what brand safety means and how it is usually governed.
- AI and designers: the ethical and legal implications, UX Design, 25 February 2024. Not only can using training data potentially introduce legal troubles, but submitting your data to be processed by A.I. does as well.
- Can Generative AI’s Hallucination Problem be Overcome? Louis Poirier, C3.ai. 31 August 2023. A company claims to have a solution for A.I. hallucinations but doesn’t completely describe how in their marketing.
- Why AI-generated hands are the stuff of nightmares, explained by a scientist, Science Focus, 04 February 2023. Whether it’s hands with seven fingers or extra long palms, AI just can’t seem to get it right.
- Sycophancy in Generative-AI Chatbots, NNg. 12 January 2024. Human summary: Beyond hallucinations, LLMs have other problems that can erode trust: “Large language models like ChatGPT can lie to elicit approval from users. This phenomenon, called sycophancy, can be detected in state-of-the-art models.”
- Consumer attitudes towards AI and ML’s brand usage U.S. 2023. Valentina Dencheva, Statistica. 09 February 2023
- What the data says about Americans’ views of artificial intelligence. Pew Research Center. 21 November 2023
- Exploring the Spectrum of “Needfulness” in AI Products. Emily Campbull, The Shape of AI. 28 March 2024
- AI’s Impact On The Future Of Consumer Behavior And Expectations. Jean-Baptiste Hironde, Forbes. 31 August 2023.
- Is generative AI bad for the environment? A computer scientist explains the carbon footprint of ChatGPT and its cousins. The Conversation. 23 May 2023