In the last decade, we’ve seen the focus of AI design move away from research and development initiatives into real-world applied support.
Prior to applying AI in industries, data scientists and business leaders alike focused their attention on understanding the fundamentals of machine learning and data. Significant time was spent prototyping and iterating models that required mass amounts of power and training information.
In recent years, however, we’ve seen a shift towards AI literacy concerns: What sort of models should be built? How do we build them correctly?
These major eras have not only defined the evolution of AI design, but also indicate where we’re headed.
2010s: A History of Big Data and AI
The early 2010s saw the rise of Big Data.
There was more information than ever before, arriving in increased volumes and with unprecedented speed—but its limited accessibility meant it was unlikely that many AI projects would make it to the modeling stage.
Although AI startups were climbing in numbers, very loose definitions of AI and machine learning caused many to adopt an “anything goes” attitude. Business leaders struggled to understand where and how to begin designing AI for their organizations.
Large models, like the neural network Google built in 2012 to study early facial and image recognition capabilities, coincided with the growing popularity of deep learning. Over a decade ago, this network leveraged 16,000 computer processors with one billion total collectors to learn how to identify cats.
Although not as impressive as the 17-billion-parameter GPT-3 neural network more recently released, it was an enormous leap toward the advanced ML and AI we’re working with today.
In 2019, 90% of the world’s digital data had been created in the previous two years. This impacted companies in a number of ways, both positive and negative:
- Smaller organizations, previously restricted by limited amounts of data, were able to train and incorporate AI into their business strategies.
- The growing availability of data led to an increased interest in AI usage.
- Models fed with mass amounts of data were more difficult to audit for negative bias or inaccuracies.
This trend marked the rise of artificial intelligence as we know it: customer service bots, automated scheduling, voice assistants, and more.
2020s: The Right Data and Risk Mitigation
Today, data is much more accessible. The rise of cloud computing has led to a shift in the access versus ownership of information, resulting in significantly less barriers to adopting AI.
The challenge is no longer having enough data. It’s obtaining the right data. In fact, Andrew Ng has found that with the advances of transfer learning and large language models, we can leverage smaller, high-quality data sets to achieve even better benchmarks than in the past.
But even with the right data, companies must proactively manage the many types of risk associated with their solutions.
- Privacy of data. Is user and company data protected?
- Physical, financial, and emotional harm. The actual harm to any involved humans.
- Auditability. Can you demonstrate and reproduce the same conditions?
- Equity of use. Is there equal representation of race, gender, dialects, etc. in your solution?
It can be challenging to imagine what these risks look like, which is why we’ve seen a number of companies experience unintended consequences with AI. And while risk management is often easier among one or two models, it becomes exponentially more challenging to control when organizations grow.
So, how do you manage risk when your AI starts to scale?
- Map out every facet of your process. What areas of business, decision-making, and consumer journeys will your solution touch? Where will your solution impact humans?
- Great creative with scenarios. Unpack the possible interactions between your solution and its end users. What is the worst that could happen? What is a cost you would consider unacceptable in one of these interactions? Could your solution yield a sexist decision? Are there any underrepresented classes? What’s the cost of this going wrong?
- Build a diverse team. Not only should your team be multidisciplinary and diverse, but it should also reflect the humans your solution intends on serving.
- Set the right KPIs. Set accurate, reliable KPIs to measure the results of your AI—and have a contingency plan in case they are not met.
- Account for model drift. AI and data age much more rapidly than traditional software. Once you’ve identified the risks, put a plan in place to account for model change over time.
Mitigating risk and updating models accordingly is now the “heavy lift” of AI—but what’s next?
A Future of Foundation Models
Today, we’re seeing more and more companies benefit from transfer learning, where a foundation model is trained on a large amount of data before being used as a “starting point” to train other models for niche tasks.
Foundation models eliminate the need to start over from scratch with every new project, and increase both the efficiency and predictability of related models.
But while these models will lower barriers to designing and piloting AI, there are still limitations to keep in mind.
- Third-party platforms and the data they contain can be unpredictable and even subject to complete removal.
- The training data used for these foundational models is much less transparent compared to custom-built models.
- Consistent measurements of bias when monitoring AI aren’t always guaranteed. It’s critical to remember the original model was trained by an entirely different group of individuals with entirely different biases.
Like other technologies, AI has a fast-evolving history of developments, uses, and lessons learned. Its impact will only continue to grow across industries.
How do you plan on leveraging its power responsibly?
About the author on AI design: Cal Al-Dhubaib is a globally recognized data scientist and AI strategist in trustworthy artificial intelligence, as well as the founder and CEO of Cleveland-based AI consultancy firm, Pandata.