There are a lot of talks about artificial intelligence (AI) use, and since it is still relatively novel and unpredictable, the debates about the safest and most ethical ways that it can be applied are endless.
One of the things that comes up is the use of white box and black box AI and comparing them to decide which is better. And this isn’t just a matter of ideological debate; entire laws and policies have been created based on the perceived benefits and dangers of both. In this guide, we will delve into the white box and black box AI, explain both concepts, provide you with examples of their use, and finally find out which is better.
White Box AI
As most of us know, most AI programs function by coming to logical conclusions based on the data that they have access to. This conclusion could be a decision recommendation (e.g. approving or denying an application) or even the creation of a new piece of data such as writing an email. In essence, white box AI is AI that is transparent about how its algorithm functions and how it arrives at conclusions.
Let’s say an AI is tasked with analyzing loan applications and deciding whether or not a candidate is suitable. A white box AI is open about how it analyses each application, what factors are taken into account, and how it reaches its conclusion. As such, when the AI approves or denies such an application, those who use and design AI can tell what and how led to this decision. They are also able to predict what sorts of solutions such an AI would come up with.
Some of the machine-learning models that are used for white box AIs include linear regression, decision trees, and rule-based systems. Because they are so simple and their processes are so predictable, white box AIs are commonly used in industries like business where full transparency would be very valued.
Black Box AI
On the opposite end of the spectrum, we have black box AI. This type of AI is not transparent about its inner workings and how it comes up with its conclusions. You can provide it with input and get output, but you will not be able to examine the system’s code or the logic that produced the output. Like with the example of a loan application, the AI would either approve or deny such applications but wouldn’t say how it makes these decisions.
Black box AI is quite tricky in that it is known to be very accurate in making predictions. Examples of black box AI in action include deep learning algorithms – the mechanisms behind them are brilliant yet too complex for humans to comprehend.
Some of the biggest AI breakthroughs have come from black box AI, and once upon a time, they were the industry standard. But now, many have moved away from the use of black box AI because of its lack of transparency. This isn’t to say that black box AI has been completely abandoned. As we’ve said, it tends to produce brilliant results. However, many businesses choose to not employ such AI in day-to-day functions.
What Is Better?
Now that you understand how both white box AI and black box AI work, it is worth asking if either is better than the other. And the truth is that both are useful but either might be a better choice depending on the situation. For example, businesses that deal with sensitive information or offer services to others should have a full understanding of how their AI algorithms work. In such a case, a white box AI would be the better choice. This isn’t to discount black box AIs, however. Even though they are no longer the industry standard, they are still useful for further AI development and while we don’t yet know how they work entirely, we might get a clearer picture in the future. Plus, it’s been found that black box AIs operate at a very high efficiency level and the potential breakthroughs we can get from them should not be discounted.
Essentially, white box AI is better for industry practice while black box AI offers immense potential for future AI development.
Can White Box and Black Box AI Work Together?
White box and black box AI models can be used in conjunction with one another. Examples include post hoc interpretability models which are used to retrospectively determine how the AI model’s decisions came to be. So, even though the AI model itself does not explain how it came to a decision, these interpretability models help to investigate them. It is not perfect just yet but it does give us the best of both worlds when it comes to AI: we get the efficiency of black box AI while also having the transparency of white box AI.
Regulated industries that are not allowed to use black box AI for day-to-day tasks may also use it as a sort of confirmation tool. Let’s say a decision that needs to be made is worked on by a white box AI. The same organization can have a black box AI try to solve the same problem as a way of confirming whether the former was correct or not. In many ways, black box AI and white box AI can be used to make each other better and deliver more benefits for those who use them.
Opportunities and Challenges
Both white box and black box AI models have many opportunities.
For white box AI, some of its advantages include:
- Decision-making within the public sector. Since these AI models are transparent and efficient, they can help reduce bureaucratic delays and hasten decision-making while still being in line with official standards.
- Transparency in the legal sector. White box AI can be used to pass judgments within the legal sector while creating more transparency, which has been an issue in the legal sector for years.
- Decision-making in highly regulated settings. For example, schools where tasks such as grading student papers and other processes that are highly standardized would benefit from transparent AI while adhering to existing rules.
- Review of existing processes. The transparency that white box AI provides means that users can test and review their current processes. For example, an institution could change certain variables in its current processes and see how the AI’s results change based on this.
- AI exploration in sensitive industries. Certain industries that deal with sensitive matters such as security and the law have been reluctant to explore AI. But white box AI is more likely to be embraced by these industries and this could further AI adoption.
However, along with the mentioned benefits, there are a few things to consider. Firstly, white box AI does not lead to such groundbreaking developments as black box AIs. Because white box AI is not generating groundbreaking results, we tend to use them in less complex situations. This, in turn, means that such models don’t improve as fast as fast as we might want them to. Secondly, white box AI tends to work slower than black box AI as it not only generates results but also provides explanations for how it arrived at certain conclusions, which makes it unsuitable for extremely time-sensitive situations.
For black box AI, the advantages are as follows:
- Use in the financial industry. The high efficiency of black box AI means that it can analyze large amounts of financial data, make predictions regarding the future state of the market, and offer recommendations.
- Faster AI development. Black box AI requires less training from humans and can develop at a faster pace than white box AIs. In this lies the opportunity to develop AI at a faster speed and with more efficiency.
- Overcoming human shortcomings. Black box AI could be applied to situations and industries where human understanding falls short. This includes detecting fraud or patterns in non-human species’ behaviors that we don’t yet understand.
Meanwhile, among the challenges worth considering when dealing with black box AI, are the following:
- There is increasing mistrust of AIs without transparent algorithms and this makes some reluctant to use black box AIs. Even AIs whose inner workings are known to the public are only just being used in certain fields like academia. As you would imagine, those in these fields would be even more reluctant to use black box AI.
- Black box AI could be subject to biases. Because their inner workings are not made public, there is no way to know for certain whether or not the AI operates with prejudices or biases. This makes it especially tricky to apply to sectors like the legal field where prejudice is already an existing issue.
- The validity of the results can not be confirmed.
Black box AI and white box AI represent two approaches to AI development: complete transparency when it comes to the algorithms we rely on or acceptance of AI breakthroughs even if we don’t know how they work. For now, regulators seem to err on the side of caution but that doesn’t mean that black box AIs are out of the picture. Moving forward, we can expect that both will be used extensively and will drive us further into technological development, whether we fully understand them or not.