People make mistakes all of the time. To err is human, after all. Our organizations and institutions have evolved to manage those mistakes through training and processes. Brain surgery is supposed to be performed by well-trained brain surgeons. Office buildings go through extensive processes to ensure their safety. These systems reflect our understanding of human fallibility and our efforts to mitigate it.
However, this approach to human error does not directly translate to the realm of artificial intelligence technology. Too often, the amazing things AI can do dazzles people to the point of assuming that the machines are perfect, or nearly so. This is not the case with real-world applications. AI systems make a lot of mistakes, and they tend to make different mistakes than humans. As a result, our human-oriented organizations are frequently ill-prepared for the patterns of AI mistakes.
As exciting as the latest generative AI solutions can be in terms of creating prose (or graphics), they are still limited. Computers are still dumb as bricks: they don’t “understand” their output. They are simply able to do certain things very well and/or very fast based on the data on which they were trained.
Inside the tech industry, they are derisively described as “stochastic parrots,” “spicy autocomplete,” or “spellcheckers on steroids” (the last one is mine). We’ve all seen spellcheckers improve over time from spotting basic typos to commenting on punctuation, grammar and style. And, we are all used to hitting “ignore” when the spellcheckers often get something wrong.
The Cost of AI-Made Errors
The cost of AI-made errors varies widely. Ignoring a spellchecker recommending the wrong word costs a person a few seconds. A bad AI-generated text transcript of a voice recording might take more time to edit and correct than it would take to have someone simply type out the transcript. An AI tool missing a curable tumor in a medical image could result in someone losing their life (assuming a human radiologist would have spotted the tumor if they had looked at the image).
Beyond the direct costs, there can also be reputational costs experienced by an organization when an AI-caused error reflects badly on it. Open-ended AI applications, where any kinds of input can be entered, can often be manipulated to generate negative outcomes which are uncommon but still possible. The press has been full of cautionary tales about AI-caused errors.
- A car dealership which put a generative AI chatbot on its support website, and was quickly tricked into making (according to the chatbot) a legally binding promise to sell a high-end SUV for US$1.
- The U.S.-based National Eating Disorders Association made the unfortunate decision to lay off its human counselors (who were in the process of unionizing) and replace them with “Tessa,” an AI-powered chatbot. A week later Tessa was decommissioned after screenshots were captured of Tessa telling people to do the opposite things of what modern practice would recommend to someone with an eating disorder.
International technology users are familiar with the patterns of U.S.-centric technology companies when it comes to handling errors.
- The job of looking at horrific content to train AI to minimize errors (or unacceptable outputs) is outsourced to workers in Africa or Asia.
- Google’s AI platform was recently caught describing India’s current government as “fascist,” which did not go over well with the Modi administration.
- These technologies rarely work in the languages spoken by most people in the world (or make embarrassing errors).
Even with these negative stories in mind, I am still a big supporter of using AI for all types of organizations, from for-profit companies to NGOs and government agencies. However, real AI applications are rarely as miraculous as popularly portrayed. The end goal of almost all AI solutions should be in making human beings more efficient and effective. Who wouldn’t want a given group of people to be able to accomplish 10-20% more with a relatively modest expenditure?
4 Ways to Reduce AI Errors
So, how do you gain the benefits of AI without becoming a negative news story (or spending more on implementing a new AI solution that doesn’t work out)? Here are some simple recommendations.
1. Don’t Rush to Find Nails for the AI Hammer
Don’t start running around feeling like you must immediately apply AI to something, anything. This won’t work any better than forcing blockchain or metaverse technology on your organization worked out well for anybody over the past few years.
Starting with a tech solution and then looking for an application doesn’t work as well as starting with a problem and picking the right tech to help solve that problem.
2. Focus on Making Things Better for Humans
Look for repetitive or time-consuming tasks where an AI-based solution might save your team, or the people you serve, valuable time. This is where AI shines, if you have enough data to train the AI.
3. Choose Applications Where the Cost of Errors is Low
Avoid handing over life or death decisions, or decisions about serious amounts of money, to an unsupervised AI algorithm. Favor closed-ended applications, where the possible outcomes from the AI are constrained to a preset list of options.
A good example of this is helping users find the right tech support article from a fixed database of articles, where an error would be simply directing them to an article that doesn’t address their specific issue.
4. Keep Humans in the Loop
Use AI to support a human being do a task faster or better, but where the human is responsible for the outcome. One of the biggest benefits of maintaining human beings in the loop is to limit the impact of AI errors.
Generative AI is Spellchecker on Steroids
A simple rubric for generative AI (like ChatGPT) is word substitution. I used to joke that if you took a claim about blockchain, and replaced the word “blockchain” with “database,” and the claim made no sense, then it made no sense! For example, if the claim is “Blockchain will solve world hunger,” the revised implausible claim would read “A database will solve world hunger.”
If you take “generative AI” and replace it with “spellchecker on steroids,” and it makes no sense, be very careful! ChatGPT is a great tool for improving written materials common to so many organizations. And, if it’s five or ten times as useful as a spellchecker, that’s pretty exciting. But, sending out cover letters, legal briefs, customer communications and grant applications directly from tools like ChatGPT is not likely to go well.
The new wave of AI tools is an exciting step forward for information technology. Unlike blockchain, which has yet to revolutionize the social impact field, AI already has transformed software technology and will continue to have outsized impacts.
However, generative AI is not a panacea, it’s just a bigger than average incremental step. If you can successfully make human beings more powerful and effective while avoiding unaffordably expensive mistakes, including possibly damaging your reputation, then you are likely to be successful deploying this latest technology.
By Jim Fruchterman, the founder of Tech Matters
