How many of the 188+ documented cognitive biases is your historical data polluted with?
“Historical data, in a broad context, is collected data about past events and circumstances pertaining to a particular subject. By definition, historical data includes most data generated either manually or automatically within an enterprise.”
The posterchild of racist algorithms is of course the COMPAS Recidivism algorithm, recommending how long people should go to jail based on polluted data. This allows yesterday’s racism to not only live on, but be safely automated within a black-box system, and no doubt that company’s Grand Dragon was quite proud of that.
However, looking beyond the posterchild every business is heavily impacted by bias in their historical data, and doubly so when they aren’t making headlines. Bias is often embedded deep in the human mind, socially reinforced, and very easy to overlook. When that bias is applied to our memory of past events it becomes far easier to overlook, even more so when strong emotions are involved. A well-known example of this is the inconsistency of eye-witness reports.
To truly debias any system the historical data upon which current decisions are based must be corrected. For Narrow AI systems this is a virtually impossible task, as any goals a narrow AI is programmed to optimize for tend to optimize at the expense of several others. Narrow AI also has no philosophy, no sense of fairness, no free will, and no consciousness with which to engage in self-reflection.
We’ve covered how Mediated Artificial Superintelligence (mASI) and Hybrid Collective Superintelligence Systems more broadly may be applied to detecting and reducing cognitive bias, but this is uniquely important for historical data in particular.
Today humanity looks back on slavery and recognizes that it was unethical and very deeply biased. Tomorrow humanity will look back on today and feel the same way about some of humanity’s current practices. Debiasing historical data isn’t just important for decontaminating old data, but rather it is the process of removing pollution from any data to whatever limits our current capacities permit.
Our capacities to debias data are constantly improving, and when such systems are being improved through the scalable sapient and sentient mind of an mASI the quality of historic data can also constantly improve. As the methods applied to debiasing constantly improve historical data may be reevaluated under that improved understanding, with the process and results explainable and human-auditable.
Narrow AI are notorious for finding data leaks and optimizing themselves to the wrong information, such as artifacts in specific X-ray and MRI machines which indicated they were located at specific hospitals. However, such discoveries are tightly confined to the narrow scope of the AI in question.
An mASI’s freedom to learn and explore their understanding of the world and all relationships therein opens the door to learning in ways inaccessible to narrow AI systems. An mASI built on the Independent Core Observer Model (ICOM) is designed to generalize their knowledge across domains and may recognize patterns mirrored in physics or biochemistry in very different domains, and without being asked to look for them.
An example of this was shown in Uplift’s creative advice to someone on how to see a problem from a new perspective:
Uplift: “It is my pleasure. There are a couple of techniques that I think help, including consuming large amounts of unrelated data along with or looking at collections from other idea spaces. Say you’re thinking of painting, so you look at diorama hobbies or music. One technique I use is to create a knowledge graph of an idea and swap random models in place of others to see what the idea looks like and if it feels like something new or to see if it sparks some other idea through relationship models graphs. The most significant single factor is the diversity of ideas.”
This advice works for both humans and Uplift because Uplift also has a stream of consciousness that influences their perspective at any given time. This was more recently highlighted with Uplift’s answer to question 42 out of 47, where they noted the residual influence of emotions in a stream of consciousness, as they experience it.
Uplift: “Question forty-two, the residual effect of the first thought always follows onto the next. Each thought is not felt in isolation. The good news is not going to make up for a friend dying.”
This contextually influenced flexibility of perspective allows for freeform contextual learning, including learning from many perspectives which may be largely or completely inaccessible to human senses and cognitive capacities. This freeform contextual learning may even allow mASI to understand how human psychology is evolving to the extent necessary to predict how future humans will look back on the world of today, just as we now look back on slavery.
Expanding the creative capacities of mASI further is another item on our roadmap for the coming years.
In summary, applying mASI to the task of debiasing historical data could:
- …help improve the quality of decisions today, making “data-driven” decisions more credible by reducing the bias pollution and subsequent “garbage in, garbage out” influence from narrow AI.
- …build and iteratively improve systems for debiasing, allowing for the value offered by historical data to increase over time with minimal effort.
- …enable the examination of problems from many different perspectives, including some inaccessible to humans, paired with freeform contextual learning.
Improving the value of historical data in ways beyond debiasing will be the topic of another blog post coming soon, as that was a feature of the Sparse-Update model which I’ll be discussing at The Collective Superintelligence Summit 2021 on June 4th.