Artificial intelligence (AI) is animating the world that electricity illuminated and that the internet connected. In the last decade, we witnessed AI-fueled technology become inexorably embedded into our daily lives, work, entertainment, society, and culture.
For AI’s successes and positive impact on virtually every sector of human enterprise—in finance, science, manufacturing, transportation, healthcare, environment, and energy—there are glaring faults with these systems. The thing with AI is fallacies beget fallacies, and the longer this goes on the more out of control this will become. Sometimes the missteps are gravely obvious, like discriminatory hospital algorithms that were less likely to refer black people than white people who were equally sick with complex medical conditions to improved care units. Other instances, where mortgage approvals are being offered on a biased basis against women, may be harder to decipher but are none-the-less ostracising entire classes of people. There are at least 1200 documented cases here, and this is just the tip of the iceberg.
Today, it is unimaginable to release software without rigorous testing and verification. Quality assurance (QA), devops, and automated test harness are well established practices that ensure errors are caught and mitigated early. Our founding team at Armilla AI built many robust models that are deployed globally. While we built them to positively impact the world, we know that manual testing processes and lack of automated tools to test and verify these systems cannot keep up. We know that algorithmic diversity presents a multifactorial problem for any tester to standardise and scale the validation process for machine learning (ML).
- Model types: from simple linear regressions to complex neural networks
- Data types: like images, text, tabular, etc.
- Prediction types: from classification to regression predictions
- Domain specific use cases
These complexities mean that every moment, consumers experience AI shortcomings: spam filters block important emails, GPS providing faulty directions, machine translations corrupt the meaning of phrases, mortgage models racially and economically discriminate, biometric systems mis-recognize people, healthcare systems misdiagnose along racial lines, and so much more. Overall, it is harder to find examples of AI that doesn’t fail. And when consumers fail to trust our systems, the systems will fail us as well.
As the technology becomes more powerful, we need to proactively invest in the verification, validation, control, and monitoring of these systems. There is too much at stake to not do so.
New laws will soon shape how companies use AI. The five largest federal financial regulators in the United States released a request for information on how banks use AI, signaling that new guidance is coming for the finance sector. The U.S. Federal Trade Commission (FTC) released a set of guidelines on “truth, fairness, and equity” in AI. Here they defined unfairness, and therefore the illegal use of AI, broadly as any act that “causes more harm than good.” The European Commission followed suit on April 21, 2021 by releasing its own proposal for AI regulation, which includes fines of up to 6% of a company’s annual revenues for noncompliance. Investing in creating transparency and accountability in AI systems will not remain a choice for much longer.
With the great potential of AI, there is also the potential downfall. That is why we created Armilla One, the first all-in-one QA for ML platform. We are adding the missing link to critical engineering practices in developing and maintaining ML systems. We want to empower businesses with the platform and provide governance tools to continuously validate and monitor ML systems. If there ever was a time to get in front of a problem, now is it. Together, we can eliminate faulty AI and its consequences.
Talk with us at Armilla AI today and power your journey to effective, responsible AI.
– Dan, Karthik, Rahm