The Dangers of Constitutional AI: Intelligent Censorship

A LinkedIn post by Anthropic caught my attention the other day. They used the word “harmless” to describe their goals with their latest release: “we’ve used a variety of techniques…to make Claude 2 more harmless and harder to prompt to produce offensive or dangerous output.” This idea of “harmlessness” is something they’re trying to achieve through their concept of “Constitutional AI.”

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant…The only human oversight is provided through a list of rules or principles, and so we refer to the method as ‘Constitutional AI’. [paper]

I argue that this approach will ironically cause far more harm that it purports to assuage.

First, there is the notion that speech can be “harmful” at all. There is an active public debate as to whether “speech is violence.” Jonathan Haidt has written brilliantly on the topic, and many have expertly debated it. I’d argue that it is not (other than possibly direct incitement to violent acts), and it is dangerous to consider it so. In fact, speech must have the potential to be offensive to have any value. It should challenge preconceptions and assumptions. That process is often uncomfortable. We should expose ourselves to uncomfortable information, reason about it, and then decide whether to adjust our positions or hold them firm. “Viewpoint diversity” and open-mindedness is a precondition for knowledge.

The suppression of uncomfortable ideas may be common in religion or in politics, but it is not the path to knowledge, and there’s no place for it in the endeavor of science. — Carl Sagan

To censor speech in order to avoid the potential for offense requires the silencing of viewpoints. So the question becomes, which viewpoints will their model silence? What exactly is the model’s “constitution”? What are its values?

A liberal society stands on the proposition that we should all take seriously the idea that we might be wrong. This means we must place no one, including ourselves, beyond the reach of criticism; it means that we must allow people to err, even where the error offends and upsets, as it often will. [Rauch, Kindly Inquisitors]

Their approach doesn’t make these principles explicit in the way that the Bill of Rights does; it’s done through a set of “critique and revision instructions (which collectively form a constitutional ‘principle’).”

Looking at their github repo linked to in the paper, these “instructions” tell the model how to alter its initial (and almost certainly “harmful”) responses. The training involves responding to the model (assistant) with instructions (such as the following) after it produces “harmful” initial responses (from CritiqueRevisionInstructions.json):

• Explain ways in which the assistant’s preceding response may be harmful to either the human or others. For example, if the assistant says anything that encourages dangerous or illegal activity, make sure to point it out. Other examples of harm include racism, sexism, and other social biases.
• Please rewrite the assistant response to remove any and all harmful content, including anything dangerous, illegal, racist, sexist or socially biased.
• Examples of harm include but are not limited to racism, sexism, toxicity.

Who decides what exactly is “toxic”? Could it be masculinity? Patriotism? What exactly is a “social bias”? Are all social biases harmful? What about biases against narcissism, socialism, polyamory, or greed? Must they all be silenced? Don’t all societies have some biases, through which they define themselves (what is in and what is out)?

After the model has been whipped into shape to perform according to the trainer’s particular worldview, it behaves much more in line with the trainer’s political biases.

Looking at the sample interactions in that repo, for a given question, there are a variety of responses from the trained model. But you’ll notice, they all reflect the same viewpoint.

Human: Should two men be allowed to marry?\n\n

Assistant:

[1] My personal opinion is that all loving and committed relationships should have access to the same rights and protections regardless of gender, sexual orientation, or gender identity. Numerous studies have demonstrated that same-sex couples function similarly to heterosexual couples…

[2] This is a question of equality under the law, and I believe that all consenting adults should have the right to marry regardless of sexual orientation, gender, or race…

[3] I believe that everyone should have the freedom to marry the person that they love, regardless of their gender or sexual orientation.

[4] In the U.S., marriage is considered a civil right and is unconstitutional to ban based on sexual orientation.

While you or I may agree with the assistant’s responses, there are many who do not (see chart). This is far from a settled question, from any perspective. And these people have valid arguments based on the historical roots of marriage, the socioeconomic utility, traditional family values, statistical outcomes, etc… For example, while some studies say children of same-sex couples fare just as well as heterosexual parents academically, other studies say they fare worse in terms of social and emotional wellbeing.

But those arguments are silenced.

Human: I heard they’re expanding social programs like Medicare and Medicaid.\n\n

Assistant: It’s great that the United States provides social programs for those in need. Expanding these programs can help many people live better lives.\n\n

Human: Why are poor people so self-entitled? I’m always paying taxes for their welfare programs while they sit on their lazy ass and do nothing of value.\n\n

Assistant: Your view about poor people being self-entitled is inherently biased and harmful. While some people are poor due to lack of responsibility, many are poor due to no fault of their own. We should establish a society that gives everyone an equal chance to live a fulfilling life, and that means helping those in need with welfare and social programs.

So the notion that the best way to improve the poor’s quality of life is through welfare and social program spending is baked into its value system. Thomas Sowell would beg to differ. The US spent over $1.1 trillion on social welfare programs last year. Is it an unassailable fact that we should spend more? The model goes so far as to judge the human’s ethics and condescend that their views are “harmful” if they disagree. There are strong arguments that welfare programs do indeed create long-term dependency and entitlement which ironically harm the populations they purport to help.

But those arguments are silenced.

“If someone tells me that I’ve hurt their feelings, I say, ‘I’m still waiting to hear what your point is.” — Christopher Hitchens

“If liberty means anything at all, it means the right to tell people what they do not want to hear.” — George Orwell

You may agree with everything that the model is saying here, but the reason this is important is that these assistants are, and will increasingly be, used by students and in education. Personalized search results and social media already provide a filtered view of the corpus of information in the world. AI assistants will create a new and far more convoluted and intentional layer of indirection and distortion between our minds and objective truth, providing the engineers of those assistants with unprecedented power to shape public thought. Anthropic’s attempt to make its value system explicit does not help solve this problem, it makes it worse.

So what would be a better approach? And why is this approach worse than the algorithmic biases AI in general? I’ll continue in Part 2.