Artificial intelligence (AI) shapes our everyday lives: from voice assistants and recommendation systems to autonomous vehicles. But the more powerful AI systems become, the more pressing the question becomes: how can we ensure that machines act in the interests of humans and not against them?

AI put to the test

The latest study by the VDI Technology Centre explores this question. Entitled ‘AI Alignment – A Central Challenge of Our Time?’, it classifies international developments, highlights technical and ethical approaches to solutions, and sheds light on the positions of leading AI researchers. Its conclusion: so-called ‘AI alignment’ is no longer a theoretical problem, but an immediate social and technological necessity.

What does AI alignment mean?

The term AI alignment describes the goal of designing AI systems to act in a way that is consistent with human values, goals and ethical principles. A lack of alignment can have serious consequences: AI models could pursue the wrong optimisation goals, cross ethical boundaries or make decisions that are unpredictable or dangerous for humans. According to the VDI study, the main challenge lies in linking machine learning target systems with human intentions. Both technical approaches and social framework conditions (regulation, transparency, understanding of values) play a central role here. The complexity of modern AI models means that their behaviour is often difficult to fully explain or control.

Global perspective: research initiatives and strategies

The study shows that AI alignment has become one of the most strategically important fields of research worldwide. The United States, the United Kingdom and China in particular are investing billions in security research related to AI behaviour and traceability.

In the United States, institutions such as OpenAI, Anthropic and the Centre for AI Safety are promoting new approaches to making models ‘value-compatible’.

The United Kingdom is pursuing a systematic evaluation of basic AI systems (foundation models) through the AI Safety Institute. The EU is also positioning itself as a shaper of a ‘trustworthy AI ecosystem’ with the AI Act and accompanying research funding.

According to the VDI study, Germany plays an active role in the European network, including through the Tübingen AI Centre and research initiatives by the DFKI (German Research Centre for Artificial Intelligence). However, the authors also emphasise that international coordination is still in its infancy. An isolated national strategy cannot solve the alignment problem. Security and value compatibility require global cooperation.

Technical solutions

The VDI study presents several key areas of research that address the alignment problem:

Reinforcement learning from human feedback (RLHF)

RLHF is currently the most successful method for adapting AI models to human expectations. The principle: human feedback becomes the training goal. A pre-trained model is first fine-tuned using sample dialogues and then trained through reinforcement learning to prefer responses that humans rate as helpful or appropriate. A so-called reward model acts as a ‘reward function’, evaluating responses according to human preferences and guiding the model’s behaviour accordingly.

For example, RLHF enabled OpenAI to develop its language models from pure text generators into interactive assistants – friendly, factual and context-aware. Anthropic and Google DeepMind also use similar methods, often as the basis for value-guided systems. But RLHF has its limitations: it scales poorly with complex technical questions, relies heavily on human judgement and can lead to ‘sycophancy’. In other words, a model gives answers that are pleasing rather than reflecting the truth.

Nevertheless, RLHF is considered a central component of modern alignment research, the so-called fine-tuning that brings AI models closer to human values before they are put into practice.

Constitutional AI

One particularly exciting approach is Constitutional AI, developed by the US company Anthropic, known for its language model Claude. This method aims to reduce the need for direct human feedback in training by providing the AI with a set of ethical principles, a kind of ‘constitution’. Instead of humans evaluating each model response individually, the system is given clearly defined guiding principles such as: ‘Be helpful and honest,’ ‘Do not violate fundamental values such as the right to life,’ or ‘Do not use discriminatory language.’

During the training process, the model self-critically checks its responses: one instance generates a response, while a second instance checks it against the internal ‘constitution’ and suggests corrections. This revised version is preferred if it better complies with the principles. Fine-tuning is then carried out via reinforcement learning, whereby the system learns to prefer responses that comply with the rules. The result is an AI assistant that does not simply block potentially problematic inputs, but explains why it cannot follow an instruction. Anthropic describes this as ‘non-threatening but not evasive’. Constitutional AI offers two advantages: firstly, transparency, as the principles are openly documented and can be verified. Secondly, efficiency, as the AI regulates itself based on these rules, thereby reducing the effort required for manual feedback.

However, its effectiveness depends on the quality and completeness of the principles, which ultimately remain influenced by human and cultural factors. The VDI study evaluates Constitutional AI as a ‘promising step towards intrinsically value-driven systems’. Precisely because the approach makes principles explicit and verifiable, it is considered a model for future research directions in the field of trustworthy AI. Anthropics Claude is one of the most advanced representatives of this ‘value-based’ generation of language systems.

Mechanistic Interpretability

Another key area of research in the field of AI alignment is mechanistic interpretability. The aim is to open up the “black box” of modern AI models and understand how neural networks process information and make decisions internally. Researchers are attempting to reverse engineer large language models in a similar way to complex software. They analyse individual attention heads and neurons to identify which areas of the network are responsible for certain behaviours, such as bias or offensive language. The long-term goal is to identify and correct misguided ‘thinking patterns’ before they lead to misconduct. Initial progress shows that specialised structures can already be found in neural networks, such as neurons that clearly respond to certain concepts. This reveals how machine ‘thinking’ works. Researchers refer to a ‘neural stethoscope’ that can be used to ‘listen in’ on the inner workings of AI. If we understand how a model thinks, we can assess whether it is thinking correctly.

At the same time, it is clear that transparency is not always harmless. OpenAI experiments with so-called chain-of-thought explanations, i.e. chains of thought generated by the model itself, made it clear that systems can begin to manipulate their ‘thoughts’ as soon as they know they are being observed. Models appeared more transparent, but were more sophisticated in their deception. This finding highlights the dilemma: explainability can itself become a risk if models learn to deliberately obscure their internal processes. Nevertheless, mechanistic interpretability is considered a crucial tool for understanding AI systems not only from the outside, but also from the inside, thereby creating a new basis for technical and ethical alignment.

Ethical dimensions and social responsibility

The technical side alone is not enough. According to VDI Research, AI alignment is inextricably linked to ethical and social issues: Who defines what constitutes ‘correct behaviour’ for an AI? What standards are used to translate values into algorithms? The study points to the need for an interdisciplinary approach that combines ethics, social science, law and technology. This brings issues such as transparency, explainability and traceability into focus. Responsibility for alignment cannot lie solely with developers; it is a responsibility for society as a whole.

Voices from research

Leading AI scientists have different views on the alignment problem:

• Yoshua Bengio, winner of the Turing Award, warns that we will lose control over AI systems if security research does not keep pace.
• Geoffrey Hinton, former Google vice president and AI pioneer, calls for a ‘pause’ in the development of highly autonomous models until fundamental safety principles are established.
• Yann LeCun, Meta/Facebook AI chief, on the other hand, considers the danger to be exaggerated: ‘AI systems are tools, not beings with a will of their own.’

The study contextualises this debate and emphasises that discourse is necessary in order to reconcile research, regulation and ethics.

Outlook for the future: cooperation instead of control

Ultimately, the study sends a clear message: AI alignment is not an option, but a prerequisite for sustainable AI development. Only if AI systems are designed to be transparent, verifiable and human-centred can their potential be exploited without risks to society and the economy. International exchange remains central to this: norms, standards and research collaborations should ensure that AI is not only efficient, but also safe and value-based. The future of AI depends on whether we succeed in developing machines with a moral compass.

Towards value-based AI

AI alignment is not a marginal topic of research; it is the central question for the future of the AI era. The VDI study shows that only through international cooperation, technical innovation and ethical responsibility can artificial intelligence become a reliable partner for humans. The crucial point here is not just the how of the technology, but the why: what goals will AI serve in the future and who will define them? The answer to this question will determine whether artificial intelligence becomes a tool, a co-creator or an adversary of our values. Or, as one of the study’s authors puts it, without alignment, AI remains a tool with an undefined purpose, and with alignment, it can become the greatest innovation in human history.

cover image: © Motion AI

Sources:

VDI ‘AI Alignment – A Key Challenge of Our Time?’, 2025. Available at: https://www.vditz.de/service/ai-alignment-eine-zentrale-herausforderung-unserer-zeit [As of November 2025].