The nasal test for Covid-19 requires a nurse to insert a 6-inch long swab deep into your nasal passages. The nurse inserts this long-handled swab into both of your nostrils and moves it around for 15 seconds.
Now, imagine that your nurse is a robot.
A few months ago, a nasal swab robot was developed by Brain Navi, a Taiwanese startup. The company’s intent was to minimize the spread of infection by reducing staff-patient contact. So, here we have a robot autonomously navigating the probe down into your throat, and carefully avoiding channels that lead up to the eyes.
The robot is supposed to be safe. But many patients would, understandably, be terrified.
Unfortunately, enterprise applications of artificial intelligence (AI) are often no less misguided. Today, AI has picked up remarkable capabilities. It’s better than humans in tasks such as voice and image recognition, across disciplines from audio transcription to games.
But does this mean we should simply hand over the reins to machines and sit back? Not quite.
Your business needs augmented intelligence
You need humans to make your AI solutions more effective, acceptable, and humane for your users. That’s when they will be adopted and deliver ROI for your organization. When AI and humans combine forces, the whole can be greater than the sum of its parts.
A large computer manufacturer wanted to find out what made its customers happy. Gramener, a company providing data science solutions analyzed tens of thousands of comments from the client’s bi-annual voice of customer (VoC) survey. A key step in this text analytics process was to find what the customers were talking about. Were they worried about billing or after-sales service?
The team used AI language models to classify comments into the right categories. The algorithm delivered an average accuracy of over 90%, but the business users weren’t happy. While the algorithm aced at most categories, there were a few where it stumbled, at around 60% accuracy. This led to poor decisions in those areas.
Algorithms perform best when they are trained on large volumes of data, with a representative variety of scenarios. The low-accuracy categories in this project had neither. The project team experimented by bringing in humans to handle those categories where the model’s confidence was low.
At low manual effort, the overall solution accuracy shot up. This delivered an improvement of 2 percentage points in the client’s Net Promoter Score.
Algorithms detect online fraud by studying factors such as consumer behavior and historical shopping patterns. They learn from past examples to identify what’s normal and what’s not. With the onset of the pandemic, these algorithms started failing.
In today’s ‘new normal’, consumers have gone remote. They spend more time online, and the spending patterns have shifted in unexpected ways. Suddenly, everything these algorithms have learned has become irrelevant. Covid-19 threw them a curveball.
Algorithms work well only in scenarios that they are trained for. In completely new situations, humans must step in. Organizations that have kept humans in the loop can quickly transition control to them in such situations. Humans can keep systems running smoothly by ensuring that they are resilient in the face of change.
Meanwhile algorithms can go back to the classroom to unlearn, relearn, and come back a little smarter. For example, a recent NIST study found that the use of face masks is breaking facial recognition algorithms, such as the ones used in border crossings. Most systems had error rates up to 50%, calling for manual intervention. The algorithms are being retrained to use areas visible around the eyes.
On March 18, 2018, Elaine Herzberg was walking her bike across Mill Avenue. It was around 10 p.m in Tempe, Arizona. She crossed several lanes of traffic, before being struck by a Volvo.
But this wasn’t any Volvo. It was a self-driving car, being tested by Uber.
The car was trained to detect jaywalkers at crosswalks. But, Herzberg had been crossing in the middle of the road, so the AI failed to detect her.
This tragic incident was the first pedestrian death caused by a self-driving car. It raised several questions. When AI makes a mistake, who should be held responsible? Is it the carmaker (Volvo), the AI system maker (Uber), the car driver (Rafaela Vasquez), or the pedestrian (Elaine Herzberg)?
Occasionally, high-precision algorithms will falter, even in familiar scenarios. Rather than roll back the advances made in automation, we must make efforts to improve accountability. Last month, the European Commission published recommendations from an independent expert report for self-driving cars.
The experts call for identifying ownership of all parties and for devising ways to attribute responsibility across scenarios. The report recommends an improvement of human-machine interactions so that AI and drivers can communicate better and understand each other’s limitations.
Will Siri, Alexa or Google Assistant discriminate against you? Earlier this year, researchers at Stanford University attempted to answer this question by studying the top voice recognition systems in the world. They found that these popular devices had more trouble understanding Black people than white people. They misidentified 35 percent of words spoken by Black users, but only 19 percent for white users.
Bias is a thorny issue in AI. But we must remember that algorithms are only as good as the data used to train them. Our world is anything but perfect. When algorithms learn from our data, they mimic these imperfections and magnify the bias. There is ongoing research in AI to improve fairness and ethics. However, no amount of model engineering will make algorithms perfect.
In the real world, if we are serious about fighting bias, we use our judgement. We make rules more inclusive and adopt measures to amplify suppressed voices. The same approach is needed in AI solutions. Design human intervention to check and address potential scenarios of discrimination. Use human judgment to fight a machine’s learned bias.
Augmented intelligence needs thoughtful design
We often measure progress in AI by comparing AI’s abilities to that of humans.
While that’s a useful benchmarking exercise, it’s a mistake to use this approach while designing AI solutions. Organizations often pit AI against humans. This doesn’t do justice to either one. It leads to suboptimal performance, brittle solutions, untrustworthy applications and unfair decisions.
Augmented intelligence combines the strengths of humans with those of AI. It combines the speed, logic and consistency of machines with the common sense, emotional intelligence and empathy of humans.
To achieve augmented intelligence, you need humans in the loop. This must be planned upfront. Merely adding new processes or responsibilities to an existing technology solution leads to poor results. You must (re)design the solution workflow, and decide which areas are best handled by algorithms. You should define whether humans must make decisions or review decisions made by a machine.
Building augmented intelligence is an ongoing journey. With evolution in machine capabilities and changes in user’s comfort and trust levels, you must continuously improve the design.
This will make AI-driven systems that do invasive medical procedures or that make high-stakes financial decisions more compassionate and trustworthy for your users.