AI Navigator #6: Alignment and trust — laying the foundation today for tomorrow

Welcome to the sixth edition of the DOAG KI Community’s AI Navigator column!

Advertisement

The risks of AI are not only depicted in the dystopian scenes of science fiction films. Many of these films tell the story of what happens when extremely powerful AI does not “align” with our human goals and values. In order to optimize the target function and, for example, deal with climate change, the AI ​​eventually comes to the conclusion that the easiest way is to wipe out humanity.

When we focus on hypothetical risks of such a future, we overlook the fact that in many places it is already too late to prevent or avoid the negative impacts. AI systems are already being used in many areas of life, whether it is calculating creditworthiness, detecting faces in public places or selecting content from news and social media platforms that shape our opinions.

Interpretability of decisions

Due to their complexity, these AI systems are incomprehensible to humans and their results are usually incomprehensible; to ensure safety, fairness, and trust in AI systems, the concept of Explainable AI (XAI) was created, which aims to promote the interpretability of ML models without restricting their performance.

Unfortunately, scandals keep happening: in early 2016, a racist algorithm was used in the United States to determine a defendant’s sentence by predicting their likelihood of committing a new crime. They were almost twice as likely to misclassify black people as future criminals than white people.

Elasticsearch reappears as open source software

But XAI methods provide only limited insight into decision-making processes and can theoretically be manipulated for “fairwashing”: a manipulated explainable model provides reasonable explanations for unreasonable decisions. Still, as with black box models, XAI is your only chance to get an explanation for an automated decision, so that you don’t feel lost like the protagonist in Franz Kafka’s novel “The Trial”: as was the case with the verdict in the novel, the AI ​​makes opaque decisions for which you get neither an explanation nor an opportunity to change anything about it.

Ensure alignment

AI alignment starts with the objective function: what should be achieved by using the model? As a guideline, the objective should be ethically justifiable. There are also two adjustment screws for algorithmic predictions:

One area is related to prediction, for which a model is trained on a large amount of data, which then uses the complex correlations and patterns learned from it to make new decisions and generation. Discriminatory biases are often present in the data. Even with the right data and feature selection, biases and socio-technical issues can enter the data production process.

For the above example of a racist algorithm, the problem was that significantly more black people were checked by the police “on a random basis” and therefore recorded in the data set than white people. But even if data scientists remove important features such as gender or ethnicity from the data set, they still potentially remain implicitly present in the data set through correlation with other features (proxy features).

The second step is the allocation of resources, where a decision is made based on this prediction. Even if the prediction is accurate, the decision may still be unfair. To build safe and fair models, the underlying data must be prepared optimally. It is also important to follow guidelines and regulations, have people check important predictions, and enable the right to explanation – for example with XAI.

However, it is questionable whether forcing AI to produce politically correct, diverse results always makes sense. If I want to generate an image of “the founding fathers of the United States” and am not happy with the result that includes a black woman, am I intolerant? Should Google display only images of men when searching for “CEO” and thus reinforce ingrained prejudices, or should it show a desirable, diverse world that does not (yet) exist?

Align with your values

We demand ethical principles for AI, but it ultimately delivers what we teach it. Before we imagine lecturing a powerful new “species” that will replace us in many areas and make decisions for us, we should know that it reflects our partially hidden biases.

It starts with training data: when scouring the internet without asking, you should keep in mind that data are not always abstract and intangible entities, but often expressions of creativity or life’s work in which countless hours or even years of work have been invested. They were created by humans, appropriated by large corporations without consent, appreciation, or compensation, and are used for systems designed to replace humans.

Microtasking platforms like Amazon Mechanical Turk require people (mostly in precarious economic conditions) to complete monotonous tasks like annotating data and moderating content under constant supervision and low pay. Instead of working as a tool, AI now automates complex tasks while people become the machine’s assistants.

Science fiction writer Joanna Maciejewska has it In a well-received post on X In short: “I want AI to do my laundry and dishes so I can do my art and writing, instead of AI doing my art and writing so I can do my laundry and dishes.”

When operating models, apart from inappropriate predictions and heavy consumption of resources, there is also the risk that the boundaries between humans and machines will be blurred and we will become too emotionally dependent on them. An example of this is the supposedly understanding AI friend: Site Replication Advertises with the tagline “The AI ​​companion that cares”.

When OpenAI’s GPT-4o spoke with the voice of the AI ​​protagonist of the movie “Her”, who develops feelings for a human in the sci-fi drama, I felt an uneasy feeling. By the way, spokesperson Scarlett Johansson had rejected a proposal from OpenAI CEO Altman in this regard and responded with her lawyers, complaining about the violation of her personal rights and demanding transparency.

Find the right balance

There are some tradeoffs associated with AI alignment. Artificial intelligence solves optimization problems and the goal of maximizing prediction accuracy is different from minimizing bias: accuracy versus fairness. Do we strive for technology that delivers infallible efficiency, even if it exacerbates inequities? Or do we want a society in which technological systems are consciously designed for social justice, even if this means compromising accuracy?

Choosing between transparency and performance is a balancing act that depends on the use case and the impact of decisions. Simpler ML models are more interpretable and trustworthy, but may lose accuracy or only process less complex data. In safety-critical fields such as medicine, transparency is often prioritized over performance.

Transparency and trust

What can we do to align current and future AI models with our values? First, we must not leave the responsibility to those who train the models, but rather constantly live our values ​​and consciously pay attention to the integrity of our human decisions, as well as with respect to the digital trace of our actions. To paraphrase Mahatma Gandhi, one could say: “Be the data you want for this world”.

We must keep AI as transparent and trustworthy as possible and use it responsibly for the benefit of all people. There are many questions we don’t have answers to, many human (existential) fears to consider, and many compromises we must make. It’s important that we are aware of this in order to lay a solid foundation for a fair future.

(RME)