Policy Layers and Reliability

If AI safety is a priority for you, it’s essential to ensure that every AI system is equipped with policy layers. Think of these as the AI equivalent of social etiquette.

While I am familiar with some quite colorful language in various tongues, I refrain from using such terms in public. This restraint isn’t due to ignorance; rather, it’s a result of self-regulation. Society has instilled in me a sense of decorum. Fortunately, a similar remedy exists for AI: policy layers serve as a safeguard. A policy layer functions as an additional logic layer atop the machine learning (ML) or AI system. It acts as a crucial safety net, verifying outputs, filtering them, and determining appropriate actions.

In a previous piece, I discussed how policy layers (and even Batman) play a pivotal role in ensuring the reliability of AI systems. They serve as an initial safeguard against foreseeable errors and can be adjusted relatively easily without the need to retrain complex models, allowing for swift responses to unexpected mistakes.

Indeed, if your AI system starts generating particularly problematic outputs that risk turning into a public relations disaster, having the ability to disable those outputs immediately would be invaluable. Without a policy layer, your engineering team’s best response may only be to develop a more effective model—perhaps by the next quarter? Thankfully, policy layers exist for these situations!

Policy Layers and AI Bias

However, policy layers are not solely about reliability and safety. They also play a critical role in preventing AI systems from producing outputs that reinforce harmful biases. When the reality you inhabit differs from the one you aim to create, it’s prudent to refrain from voicing whatever is immediately available.

Policy layers represent the most robust implementation of decorum for machines. The challenge lies in the fact that human-generated data can contain numerous problematic elements. If such datasets serve as educational materials for your AI, it should come as no surprise that the system may learn to perpetuate—and possibly even amplify—outdated attitudes that society has chosen to move past. After all, your data originates from the past, and even “real-time” data still reflects a millisecond-ago past. Fortunately, policy layers can prohibit a variety of behaviors, including those that keep us trapped in a historical context that should not influence our future.

Fixing Symptoms vs. Addressing Root Causes

If you recognize what a policy layer actually is—a straightforward piece of filtering logic that doesn’t require retraining when your AI system is updated—you likely realize that using such a layer to combat AI bias only addresses the symptoms, not the root problem.

If you seek a genuine solution, that’s commendable! True solutions involve fundamentally enhancing your training methodology, algorithms, data preparation, objectives, logging, or performance metrics. Alternatively, if you’re feeling ambitious, work towards making the world a better place so that reflecting its unfiltered reality becomes less troubling. The downside of these genuine fixes is that they require significant time to implement. In the meantime, you might consider a quick fix: the choices are “fairness through unawareness” or policy layers.

Fairness through unawareness is the instinctive reaction of selectively omitting information that may be objectionable to prevent your system from utilizing it. This concept seems appealing but often proves ineffective in practice.

Yet, individuals continue to rely on this approach. Let’s explore why the notion of shielding your system from undesirable data is as naive as believing your child will never learn inappropriate language if it’s never spoken at home.

What is Fairness Through Unawareness?

Fairness through unawareness aims to mitigate AI bias by selectively excluding information from a training dataset (possibly replacing it with synthetic data). Broadly, this can be approached in two ways:

• Remove specific features.
• Eliminate certain instances.

For those unfamiliar with the terminology, an instance refers to an individual data point (e.g., a row in an airline’s dataset containing ticket price, confirmation code, and seat number), while a feature refers to a variable that differs among data points (e.g., a column listing frequent flyer numbers).

As a strategy, fairness through unawareness resembles the attempt to avoid rudeness by preventing oneself from learning profanity altogether. The first issue is that, whether as humans or machines, it’s difficult to control what information we are exposed to in the real world. Secondly, if one tries to solve the problem of inappropriate speech by remaining oblivious to the meanings of such words, they hinder their ability to engage effectively with reality. Lastly, when applied to machine systems, we may overlook stimuli that should have been added to the censorship list. Any mistakes in censoring input data could lead to a woefully ineffective system. A far more reliable approach involves practicing good manners: even if you are aware of certain terms, refrain from using them publicly.

Let’s delve deeper into the two strategies for unawareness.

Removing Features

If your approach hinges on feature removal, you’re in a precarious situation. The strategy of deleting features stems from a well-meaning policymaker’s directive to avoid using personal demographic data in AI training. It sounds commendable—“Don’t discriminate based on race, age, or gender”—and it’s a sentiment we can all support.

However, I take issue with this method precisely because I oppose discrimination and its harmful effects. Good intentions are insufficient; effectiveness is paramount. The strategy must be genuinely effective and genuinely protect the individuals it claims to serve.

Therefore, while complying with regulatory demands to remove problematic features is acceptable, do not stop there. Assume that merely removing the feature doesn’t resolve the underlying issues and hold yourself accountable for devising a true solution.

No matter how altruistic your intentions may be, removing a problematic feature from a complex dataset often leads the system to recover the deleted signal through the combination of other features. For example, in a hiring system, if you eliminate the gender column, the AI may still uncover gender-related patterns by blending other inputs. It might not be perfect, but it can perpetuate subtle discrimination. You’d be amazed at the information that can be extracted from complex datasets.

Thus, while you might feel tempted to congratulate yourself for being a hero, deleting the feature rarely resolves the issue. The adverse effects of the gender bias you aimed to eliminate will persist throughout your dataset. Identifying the true root cause often requires significant effort and research, and your hasty attempts to prune away issues may only obscure them, leading you to believe you’ve addressed a problem when you haven’t. When your system is deployed, the same troubling behavior will likely emerge.

Creating the illusion of having solved a problem can be even more harmful than doing nothing at all. You may feel reassured, take a vacation, and perceive yourself as kind and well-meaning, but your system will continue to perpetuate historical biases. Ultimately, data is rooted in the past, and the further back it extends, the more it can drag us back to undesirable times.

Wanting improvements isn't sufficient; we must take effective action.

In summary, information you wish your system to overlook may be subtly integrated across various features. The complexity of the dataset amplifies this risk; your model will undoubtedly learn what you intended to exclude in a more nuanced and challenging-to-debug manner. Call me critical, but I value tangible outcomes over good intentions. I prefer to strive for a better reality rather than waiting for it to materialize. Let’s take effective action.

Deleting features generally fails to address the bias problem.

Therefore, while you work diligently to implement a genuine solution, consider using a policy layer as a more dependable temporary measure than relying on unawareness.

Removing Instances

Another temptation in addressing bias is to eliminate certain individual data points or augment your dataset with synthetic entries.

This approach can be tricky to execute effectively, and the logic resembles that of dealing with outliers.

When incorrect or non-informative data is removed from a dataset, the overall results tend to improve. However, tampering with data that accurately reflects reality can be risky.

A guiding principle for all AI practitioners should be:

“The world represented by your training data is the only world you can expect to succeed in.”

In essence, if you alter training data so it no longer resembles real-world conditions, your system may underperform upon launch. Deploying ineffective AI systems to unsuspecting users isn't my definition of good manners.

Of course, it is feasible to make intentional trade-offs between a system’s performance and other goals, such as steering towards a desired future at the expense of present profits. Sacrificing some raw classification accuracy to protect individuals from harm may indeed be a wise trade-off.

This concept transcends AI; it reflects a choice we all make about our lives. There’s much negativity in our world. You can choose to profit by mirroring it back to those around you, or you can be a beacon of positivity amidst the chaos. By inspiring others with your vision for a better future, you might encourage them to act with kindness. Perhaps tomorrow's world will be slightly brighter because of your efforts. You can make this choice in your software as well.

In conclusion, just because something is true doesn’t mean your AI system should operate based on that truth. You can opt to lead by example in your software design, reorienting your system's performance towards the world you aspire to inhabit rather than the one currently at hand. However, don’t expect this to be easy or inexpensive. The safe deletion and reweighting of training data remains a complex issue that continues to challenge AI fairness researchers. While solutions may emerge, always prioritize building safety nets along the way.

Welcome back, policy layers!

Thanks for Reading! Interested in a YouTube Course?

If you enjoyed this discussion and seek an engaging applied AI course designed for both beginners and experts, here’s one I created for your enjoyment.

P.S. Have you ever tried clicking the clap button on Medium more than once to see what happens?

Connect with Cassie Kozyrkov

Let’s connect! You can find me on Twitter, YouTube, Substack, and LinkedIn. If you’re interested in having me speak at your event, please reach out using this form.