Five AI Myths That Should Not Hold You Back
At Radix, we view AI as the most transformative technology of our time. We are convinced that AI is crucial to the growth and prosperity of businesses, individuals, and society. However, the many myths floating around about AI have led to some companies deciding not to use it.
In this blog, we’re diving into five of the most common myths about AI — and we reached out to one of the leading authorities in the field of AI, Seppe vanden Broucke, to get his input. Seppe is currently Assistant Professor at UGent and Lecturer at KU Leuven.
Myth #1: AI projects are disruptive and come with drastic changes
Wrong. The idea that every AI project will disrupt existing processes or dramatically change company culture does not hold true.
While AI is a key enabler for disruptive innovation, most companies use AI as part of an overall process or workflow. Few companies use AI to the point that it causes widespread disruption across the business.
Take the AI project we completed with the team at GSK. The team needed a solution that would help reduce the time it takes to develop vaccines. We worked with the team to identify where AI would positively impact the vaccine development process. Our solution uses AI and machine learning to optimize colony-forming unit (CFU) counting. Thanks to this solution, lab technicians now spend up to 6x less time on CFU counting and reporting.
Real-world use cases like this show that you shouldn’t think of AI as something that will completely change a company’s strategy. Instead, you should think of AI as something that helps you achieve specific strategic or tactical goals. The best use of AI today is integrating it into critical processes or solving the pain points of your business.
“Most AI projects are mainly about automation and often boil down as a mechanism to pursue process improvements,” Seppe told us. “In fully realized solutions, the AI model is typically only a small part of the full process.” Seppe also explained that only companies which are very mature in their analytics management can pursue truly disruptive projects, though AI projects don’t need to be disruptive. For example, a small successful AI project that improves a process and leads to a significant cost reduction is a win when it comes to the bottom line.
Myth #2: You need massive amounts of data to train AI models
No. Training AI models does not require massive amounts of data.
The amount of data needed for AI models depends on many factors, such as the complexity of the model and the desired accuracy. In general, you should always try to achieve the best results with the minimum amount of data required. Wanting to include too much data is often a major problem and should be avoided.
We’ve learned a lot about AI training data over the years. So, based on our experience, we recommend taking these 4 concerns into account:
- Don’t wait until you think your data is perfect to train your AI models.
- Use open-source data when possible. You can benefit significantly from open-source data and pre-trained AI models that you can fine-tune (transfer learning).
- Make an initial proof of concept (POC), starting with your current data. Then afterward, capture the added value and improve the data step by step.
- Improve your data with data augmentation, data synthesis, active learning, and discriminative methods. Essentially enable your AI model to guide you towards which kind of data you need to gather.
“AI is increasingly being offered as an off-the-shelf API,” Seppe explained. “This can be a better alternative than doing it from scratch, especially for small teams and organizations." Professor vanden Broucke also added that “open data sets, data augmentation, and especially transfer learning and pre-trained models have lessened the data burden somewhat.”
Myth #3: Data needs to be of perfect quality for AI to succeed
Another misconception about AI. You need good quality data for AI models to bring positive results, but the data doesn’t need to be perfect.
How you prepare your data greatly impacts how your AI solutions work. You need to always prepare the data well so that the AI can handle outliers, missing values, and data redundancies. You should also integrate external data and build an exemplary process for gathering and evaluating data. In many cases, imperfect data can work. For example, you sometimes need to depersonalize data for privacy reasons.
“While it is true that better data leads to better outcomes faster, every organization suffers from data quality issues,” Seppe told us. “Try to get a small set of ‘golden’ data right and do something with that first.”
Professor vanden Broucke further explained that “in my opinion, data quality is often used as a scapegoat for failed projects, though more likely projects fail due to ill-defined goals, bad management, resource and time constraints. Complaints about data quality should be an opportunity for decision-makers to look further and not only question data quality but also process management.”
Myth #4: Only major tech organizations have the resources to leverage AI
False. You don’t need a lot of resources to use AI.
In the past, only companies with teams of AI experts could fully leverage AI — think Google, Facebook, Microsoft, and Amazon. But today, you can find many open-source projects that let you use AI without needing teams of AI experts. Plus, there are companies like ours working on democratizing AI so that all companies can benefit from it. You don’t have to build every AI model from scratch or reinvent the wheel. Use open-source projects like TensorFlow and PyTorch to take advantage of "Foundation Models" already published by large companies, or use custom AI solutions built by Radix to achieve your business goals.
“This is indeed a potentially risky evolution as AI models become larger, take longer to train, and require huge amounts of computing,” Seppe vanden Broucke explained. “That said, most organizations do not need to be Google, Amazon, or Facebook. An organization’s first concern should not be ‘how can we build models which are just as large?’ but ‘how can we leverage existing solutions and techniques for what matters for us?’”
Myth #5: AI comes with ethical challenges making it inappropriate for use
True and False. Humans bring ethical challenges to AI, but that doesn’t mean companies shouldn’t ever use AI.
Removing human bias when using AI is extremely difficult, if not impossible, to do. You need to train AI algorithms with good quality data. However, people and their ethics still dominate training data and AI systems overall. Driving fairness in AI requires making difficult choices based on your own ethical rules.
You can build ethical AI solutions — in-house from scratch or with an AI technology partner — if you focus on the next things:
- Explainability: Make sure you find a way to explain your AI system’s decisions.
- Oversight: Set up a framework for effective human oversight, ensuring accountability when using AI.
- Supervision: Let humans supervise the AI system and use your own ethical framework when implementing and using AI.
- Trust: Question the results of your AI solutions based on fairness, privacy, and explainability.
“Technology is always neutral, but this unethical AI is indeed a risk on which much has been written already,” Seppe said. “There is an increased focus on making AI more interpretable, though more work is needed in that area.”
Seppe also added that “mostly, it is a question of culture. AI should not be injected into the organization as an all-knowing machine that should always be trusted (it will lead to an adoption failure), but where end-users are co-creators and are free to ask questions and question the model.”
Don’t let myths about AI prevent you from achieving great results with the technology. Talk to us about how you can use AI to benefit your business.
About Seppe vanden Broucke
Seppe vanden Broucke is an Assistant Professor at the department of Business Informatics at UGent (Belgium) and is a guest lecturer at KU Leuven (Belgium). Seppe's research interests include business data mining and analytics, machine learning, process management and process mining and has co-authored more than 50 scientific papers and several books. He also regularly tutors, advises and provides consulting support to firms and has extensively collaborated with industry partners in various sectors to help develop and roll out analytical solutions. Seppe is also the academic co-coordinator of the Postgraduate Studies in Big Data & Analytics at KU Leuven, where he has taught analytics to more than 200 industry participants over the past five years.