Constitutional AI for dummies: 101 on what it is and why marketers should care

In last week's 101 series by MARKETING-INTERACTIVE, we delved into Claude, a family of foundational AI models that can be used in a variety of applications. It can help you brainstorm ideas, analysis images or process long documents using claude.ai. 

One of the key selling points highlighted by the system and our experts is that Claude is trained by constitutional AI. Constitutional AI provides one answer by giving language models explicit values determined by a constitution, rather than values determined implicitly via large-scale human feedback, according to Anthropic, the company behind claude.ai.

Don't miss: Study: 82% of SG companies say they need to reinvent business due to AI

The system uses a set of principles to make judgments about outputs. At a high level, the constitution guides the model to take on the normative behaviour described in the constitution – here, helping to avoid toxic or discriminatory outputs, avoiding helping a human engage in illegal or unethical activities, and broadly creating an AI system that is helpful, honest, and harmless, it added.

In this week's 101 piece, we take a deep dive into Constitutional AI, how it could be the safer way to train AI models and how it can help marketers. 

What is Constitutional AI?

Constitutional AI (CAI) is an approach developed by Anthropic to train AI systems to be helpful, honest, and harmless by aligning them with a set of explicit principles or rules, rather than relying solely on human feedback, explained Milind, an AI Scientist from Mercedes, who was expressing independent views.

It uses a "constitution" which is a set of principles or rules that govern the AI's behaviour, rather than using human feedback alone to shape the AI's outputs.

The constitution is used in two ways during training, said Milind. For a start, the AI is trained to critique and revise its own responses based on the principles. Secondly, the AI is trained through reinforcement learning to choose outputs that are more aligned with the principles.

"This approach aims to make the AI's values and decision-making more transparent and easier to adjust, compared to relying on implicit values learned from large-scale human feedback," said Milind, adding:

The goal is to create AI systems that remain helpful and harmless even as their capabilities grow, without requiring constant human supervision.

How does constitutional AI differ from other ways of training AI models?

There are a number of ways that constitutional AI differs from other approaches to training AI models, according to Milind. 

Explicit constitution vs. implicit values: In constitutional AI, the AI model is trained using an explicit "constitution". This is in contrast to other approaches that rely more on implicit values learned from large-scale human feedback, explained Milind.

Self-critique and revision: In the supervised learning phase of constitutional AI, the model is trained to critique and revise its own responses based on the principles in the constitution. This allows the model to learn to generate outputs that align with the defined principles. Other approaches don't have this self-critique mechanism.

AI feedback vs. human feedback: Constitutional AI reduces the reliance on human feedback during training by incorporating AI-generated feedback and preferences. This makes the training process more scalable and efficient compared to approaches that rely heavily on human annotators.

Transparency and accountability: The explicit constitution in constitutional AI makes the model's objectives and decision-making more transparent and accountable, compared to approaches where the model's values are more opaque. This transparency is intended to better align the model with societal norms and ethics.

Is constitutional AI a safer and more effective way to train AI models?

According to industry experts, constitutional AI does appear to be a safer and more effective approach to training AI models compared to traditional methods.

For one, constitutional AI aims to capture a wider range of ethical principles beyond just safety and helpfulness, drawing from sources such as human rights declarations. This helps better align the model with societal norms and values, said Milind. 

Evaluations have shown that constitutional AI models generate significantly less harmful content as compared to models trained solely on human feedback. They are also able to explain why they are avoiding certain responses.

"However, defining the right constitutional principles can be complex, and the framework must remain flexible to adapt to evolving ethical standards. Additionally, constitutional AI may not be a complete solution on its own for aligning highly capable AI systems in the future," said Milind. 

Adding to his point, Siddharth Jhanji, senior manager (domain leader), data architecture and engineering at Ekimetrics explained that traditional AI training can sometimes be like a free-for-all soccer match—everyone’s running with the ball, but there are no clear rules. Constitutional AI is different; it’s like playing by the book, ensuring every move is made with respect and fairness.

"If you’re baking a cake, you follow a recipe for the best results, right? Constitutional AI follows a ‘recipe’ of ethical guidelines to ensure the ‘cake’—or the outcome—is just what you wanted," said Jhanji. "Constitutional AI ensures your AI models work smarter. It’s about quality over quantity, precision over power." 

How can a platform trained on constitutional AI benefit industry professionals using AI in their work?

One of the biggest concerns with professionals using AI are inaccuracies, mistakes and the legal issues that might crop up. 

In fact, according to a survey done by editorial and research strategy company Media Collateral, it found that despite the benefits brought about by Gen AI, communications professionals are mainly concerned about the inaccuracies and mistakes of AI-generated content or advertisements, with 65% of respondents agreeing to that.

More than 50% have also said that the use of Gen AI has been limited at work due to ethical concerns surrounding data usage and transparency while 44% are concerned about the legal issues surrounding AI-generated content.

To combat that, constitutional AI models have been shown to generate significantly less harmful, biased, or unethical content compared to models trained solely on human feedback. This can help marketers and agencies avoid reputational damage or legal issues from AI-generated content, according to Milind. 

"The constitutional principles used to train these models can also be drawn from sources like human rights declarations and industry guidelines. This can help ensure the AI's outputs are aligned with important ethical and legal standards relevant to marketing and advertising," he said. 

Are there downsides to using constitutional AI?

Saying that, like any other training model, constitutional AI also comes with its cons. 

For a start, defining the appropriate set of constitutional principles that should guide the AI's behaviour can be complex and subtle. Ensuring the principles are comprehensive, clear, and adaptable is a significant challenge.

Ethical standards and societal norms are constantly evolving, so the constitutional framework must remain flexible and adaptable to accommodate changes in ethical considerations over time. This can be difficult to achieve, said Milind. 

He added that there may also be potential trade-offs between maximising the helpfulness of the AI's outputs versus maximizing its harmlessness. "Prioritising one over the other could lead to suboptimal performance," he said. 

Implementing constitutional AI in practice may also be more complex than traditional training approaches, requiring significant expertise and resources to define the appropriate constitutional principles and training procedures, he said. 

How can constitutional AI help marketers?

For marketers and agency mavens, using Constitutional AI is like having a personal shopper who knows your style, respects your budget, and always picks the perfect ensemble, said Jhanji. 

"It understands your brand’s voice and helps amplify it without hitting a false note. Imagine your AI chatbot speaking in your brand’s tone, engaging customers like a charismatic influencer," he said. 

He added that traditional AI can be thought of as a knife which can be dangerous in the wrong hands. However, constitutional Ai is like putting a safety handle of that knife. "It’s designed to minimise risks and protect you," he said. 

Join us on 12 June 2024 for an exciting experience as Content360 makes its debut in Malaysia! Brace yourself to join the crème de la crème of the content marketing industry hailing from across the region. Immerse yourself in a dynamic atmosphere, and uncover the latest trends with thought leaders and solution providers from the realm of content.

Related articles:
Claude for dummies: 101 on the new AI platform that might challenge ChatGPT
Cracking the code: How women can get ahead in an AI-driven workplace
AI, data, and social commerce set to shape SEA marketing landscape in 2024