Artificial intelligence has been hyped as a technology that will dramatically transform the world, promising everything from individualized medicine tailored to your specific genome to ubiquitous automated transportation. But so far, applications have tended to revolve less around transformation than ordinary cost-cutting.
Cutting edge AI applications (including our own) take advantage of the ability of AI and machine learning to process data much more quickly than humans can, enabling companies to save money and time on a wide range of tasks. It's comparable to how robots or assembly lines changed factory production, or digitization changed record-keeping. The tasks and goals remain mostly the same, but the execution becomes faster and more cost-effective.
And like those earlier advances, the role of humans in these processes is determined by what the machines can't yet do, an approach that makes sense but may limit the ability to push the boundaries of AI. Cambridge Phd Student, Jessy Lin, recently wrote a great piece on human-AI interaction, its limits, and how it can change (and is already changing) for the better by changing people's role in AI and ML tasks. Here are some key takeaways, along with our vision for the future of AI.
The Status Quo: Humans as Backup
As Lin pointed out, AI systems tend to use humans to check the system's results and correct the machine when it makes a mistake or meets a situation it does not know how to address — what Lin calls "humans as backup." In cases like translation, the computer first performs a task (e.g., translating a text). A human expert is brought in to refine the machine's initial results (making the translation more accurate and natural.) In other cases, like autonomous vehicles, a human is brought along to step in immediately if the system fails.
This way of doing AI makes intuitive sense. The system is there to reduce costs for most situations where it works, and a human can step in where it doesn't, training the system to be even better in the process.
Unfortunately, the approach has some serious limitations. For one thing, it's not a natural fit for humans. For example, a human backup driver taking over from an AI in an emergency will not have instant awareness of the situation. Their reaction time will be slower initially, which can pose serious safety issues.
Even in less safety-critical applications, humans as backup doesn't always work as well as it should. For example, human translators hired to refine machine translations often struggle with the machine's inability to grasp nuance. Fixing an overly literalistic AI translation can end up being just as time-consuming as traditionally translating the piece.
Smarter Ways to Loop Humans In
It's not that the "humans as backup" model is bad, so much as that there may be better ways to divide work between humans and machines for many applications. One simple innovation is using the machine to support the human worker, rather than the other way around. This is essentially what autocomplete is; a human writer writes the text, and the machine makes helpful suggestions. Over time, the system learns to anticipate what words the writer will need, but the human writer remains in charge.
When used for translation, this approach empowers the translator to use the machine's suggestions when they're useful and disregard them when they're not. And because the human is directing the action, the machine takes in more data, which means it can improve faster.
Another example is the way expert systems are changing to empower medical professionals to make diagnoses. AI has been used with good results to evaluate medical scans for signs of diseases such as cancer. Still, there are situations where a doctor may be interested in testing a particular hypothesis that the machine may not be trained to detect.
Newer AI technology can help the doctor by providing relevant information, such as images of the illness the doctor suspects or data comparing the patient's scan to scans of a particular rare disease. Again, the goal isn't to replace the doctor but to help the doctor do their job more effectively.
The Future of AI
Building AI that interacts with humans in smarter ways doesn't just help solve particular problems — it potentially changes the way organizations develop artificial intelligence fundamentally. Systems that integrate humans more naturally receive more feedback, which means they learn and improve more quickly.
We believe techniques from software engineering like agile development and the spiral model act as a multiplier. As AI and ML professionals learn to apply these techniques, they'll be able to test and improve their systems more quickly, shortening development cycles.
This growth could fundamentally change the nature of AI and ML in the long run. Rather than primarily being a tool for the bottom line, AI will be able to take a greater role in strategic decision-making, helping humans spot new problems to solve and build new solutions on a much shorter timescale.
Bitvore will enable you to stay on the cutting edge of AI capabilities, providing better data analytics, providing new insights, and making your best talent even more effective.
To learn more about how Bitvore can enhance your strategic capabilities, read our whitepaper Unstructured Alternative Data in Predictive Modeling.