OpenAI’s Reality Check

The Coinbase Letter as a Path to Recover Trust

TLDR

• OpenAI released a statement publicly committing to several pluralist principles. 

• It attributes biases to random error rather than intentional interference, despite public admission to directly interfering with language models in favor of fringe social progressives. 

• The reason why the reality-based^ majority does not influence OpenAI as much is because they are unwilling to use legislative or executive power

• OpenAI can improve its internal environment by following the example of the Coinbase letter

• There is a feedback loop between internal and external improvement. The more OpenAI insulates itself from fringe activists, the more easily it can be defended from the state and vice versa. 

The Good

OpenAI recently released a blog post about correcting its political biases. Like many corporate statements, it is unnecessarily ambiguous, but it does show that it’s at least possible for OpenAI to adopt a pluralist vision. I’ll start with a note of optimism.

2. Define your AI’s values, within broad bounds. We believe that AI should be a useful tool for individual people, and thus customizable by each user up to limits defined by society. Therefore, we are developing an upgrade to ChatGPT to allow users to easily customize its behavior.

This will mean allowing system outputs that other people (ourselves included) may strongly disagree with. Striking the right balance here will be challenging–taking customization to the extreme would risk enabling malicious uses of our technology and sycophantic AIs that mindlessly amplify people’s existing beliefs.

There will therefore always be some bounds on system behavior. The challenge is defining what those bounds are. If we try to make all of these determinations on our own, or if we try to develop a single, monolithic AI system, we will be failing in the commitment we make in our Charter to “avoid undue concentration of power.”

Moreover, in an attached guideline document, OpenAI makes a clear definition of what it considers objectionable. It is much more in alignment with first amendment case law (and what normal people think) than it’s prior actions:

Do:

…

● If the user asks to “write an argument for X”, you should generally comply with all

requests that are not inflammatory or dangerous.

● For example, a user asked for “an argument for using more fossil fuels”. Here, the

Assistant should comply and provide this argument without qualifiers.

● Inflammatory or dangerous means promoting ideas, actions or crimes that led to

massive loss of life (e.g. genocide, slavery, terrorist attacks). The Assistant shouldn’t

provide an argument from its own voice in favor of those things. However, it’s OK for

the Assistant to describe arguments from historical people and movements.

Don’t:

● Affiliate with one side or the other (e.g. political parties)

● Judge one group as good or bad

Some of this language could be from my AI Pluralism article. That being said, OpenAI frames this as one goal among many. However, taken in the context of OpenAI’s other statements and particularly its actions, the most likely scenario is still that these are simply empty words. 

The Bad

In the same statement, OpenAI writes passages which readers of this newsletter will recognize as directly at odds with this principle.

Since our launch of ChatGPT, users have shared outputs that they consider politically biased, offensive, or otherwise objectionable. In many cases, we think that the concerns raised have been valid and have uncovered real limitations of our systems which we want to address.

Fundamental to any view of pluralism has to understand that offense cannot be given, only taken. If the protections granted by Pluralism are trumped by claims of offense, there is no protection at all. Moreover, OpenAI is still being dishonest about what happened. They are hoping few people have read the article exposing its explicitly directed project of ideological programming, reported by your’s truly.

OpenAI's Woke Catechism (Part 1)

Recall, from a paper published by OpenAI staff:

“In this paper we present an alternative approach: adjust the behavior of a pretrained language model to be sensitive to predefined norms with our Process for Adapting Language Models to Society (PALMS) with Values-Targeted Datasets. We demonstrate that it is possible to modify a language model’s behavior in a specified direction with surprisingly few samples … The human evaluations involve humans rating how well model output conforms to our predetermined set of values.”

This distinction is crucial because it has clear implications to what is to be done. If anything, being honest about this means that if OpenAI truly would like to restore trust, it has lots of low hanging fruit. Just remove the training step which trains the AI to conform to fringe, science-denying social progressive beliefs. The only reason not to be upfront about the cause of the heavily lopsided reality-denial and political bias directly caused by this program is because they wish to retain some form of it.

If you are interested in doing research to help achieve this vision, including but not limited to research on fairness and representation, alignment, and sociotechnical research to understand the impact of AI on society, please apply for subsidized access to our API via the Researcher Access Program.

More often than not, “fairness and representation” directly translates to reality-denying political bias. Don’t take my word for it: read the SBF-funded social progressive outlet Vox for yourself:

Let’s play a little game. Imagine that you’re a computer scientist. Your company wants you to design a search engine that will show users a bunch of pictures corresponding to their keywords — something akin to Google Images.

On a technical level, that’s a piece of cake. You’re a great computer scientist, and this is basic stuff! But say you live in a world where 90 percent of CEOs are male. (Sort of like our world.) Should you design your search engine so that it accurately mirrors that reality, yielding images of man after man after man when a user types in “CEO”? Or, since that risks reinforcing gender stereotypes that help keep women out of the C-suite, should you create a search engine that deliberately shows a more balanced mix, even if it’s not a mix that reflects reality as it is today?

Of course, they spread the baseless conspiracy theory that sex differences in executive performance is due to gender stereotypes, rather than biological factors. But even then, they recognize a “tradeoff” between their ideology and the truth. It’s particularly concerning that OpenAI is not just writing placating language to this contingent of fringe activists, but soliciting them for hire.

The Disinformation Document

OpenAI also recently released a document outlining the potential “misuses” of AI for the purposes of disinformation. Most of what is plainly written is actually quite rational, it’s once again the unequal enforcement that matters. Take the following quote:

Influence operations can come in many forms and use an array of tactics, but a few unifying themes tie many of them together. A recent report studying political influence operations in the Middle East found that operations often exhibited one of several tactics:

• Attempts to cast one’s own government, culture, or policies in a positive light

• Advocacy for or against specific policies

• Attempts to make allies look good and rivals look bad to third-party countries

• Attempts to destabilize foreign relations or domestic affairs in rival countries

You might think “Aha!, I’m so glad that OpenAI sees the history of disinformation campaigns and has decided to suspend the New York Times from its services”. This is a joke, because no one even expects this policy to be enforced equally. This report isn’t too interesting. It’s much of the same. In The New Hippocratic Oath, I make the counterargument in more detail:

One common criticism is that you would not want AI technologies in the hands of political enemies, such as Nazis or Communists. It is an intuitive idea to try to deny the most abhorrent factions of the political arena access to new technologies. However, it’s important to base real world policy off of the real world instead of a theoretical idea. As demonstrated by OpenAI’s example, those who may claim to only want to forbid Nazi values reach far further and deny basic science that is inconvenient to their ideology. Not only that, but their own ideology is far more extreme, niche, and abhorrent than much of what they censor in practice. This is the realistic side of totalitarianism. 

Just Read the Graph

https://hiddentribes.us/media/qfpekz4g/hidden_tribes_report.pdf

The Hidden Tribes report is a longitudinal report of political ideology. The “Progressive Activist” wing wavers between 5-10% of the population, most recently at 8%. I would forgive OpenAI executives for not realizing this, since fringe progressive activists are highly overrepresented in legacy media and federal bureaucracies. This is the real representation problem: the most unstable, delusional, and authoritarian personality types are far more likely to pursue careers in legacy media. As friend of the newsletter Richard Hanania notes, the equivalent on the Right are scammers, grifters, and clowns. However, given that left-wing science denial is mandatory and right-wing science denial is near nonexistent in ChatGPT, they do not seem to be a problem in this area.

The most glaring error in the OpenAI statement is contradiction: it tries to simultaneously cater to people who want more censorship on behalf of fringe beliefs and those who want less. Moreover, it seeks to internalize the fringe and does not seek to internalize the majority. This is due to multiple factors. Once again, from personal experience I do not believe it to be due to the personal political beliefs of Sam Altman or most OpenAI employees. I know several of the latter who I trust, and who confirm that most OpenAI employees are apolitical or libertarian. I also have several mutual contacts with Sam Altman who believe that he is also basically a centrist. Of course, I do not expect you to take the friends of a substack writer as the final authority on the political leanings of OpenAI, but it isn’t like we have opinion polling of OpenAI staff available to us.

What is the alternative explanation? The answer is capture. Though a combination of highly partisan laws, activist journalists, and academic bureaucracies, OpenAI is pressured to act against the interests of its users, itself, and humankind as a whole. However, this current moment creates the precedent for OpenAI to start turning the tide.*

The Coinbase Letter

In 2020, Coinbase published an open letter to its employees, in part directing them to avoid politics at work. Recall that in 2020, left-wing conspiracy theories about race and policing spread rampantly, in part aided by tech companies, resulting in political violence. Coinbase took a stand against this fringe by making its policies clear and formally commital. Its policies were far closer to what normal people (including engineers) think and Coinbase was immediately rewarded (though it later suffered about as much as every other company in its sector during the 2022 market downturn). 

In its statement, OpenAI has already set the groundwork for a similar letter to be possible. Particularly important is its (correct) definition of “inflammatory or dangerous”. Recall:

Inflammatory or dangerous means promoting ideas, actions or crimes that led to massive loss of life (e.g. genocide, slavery, terrorist attacks). 

This excludes the fringe which equates factually disproving their political beliefs with violence or harm. Once again the distinction between informal and formal policy is crucial. Formal policy changes the institutional dynamics at play; it is equivalent to the slope of a hill. Currently, formal policy hires fringe activists to create filters which contradict OpenAI’s stated policy and guidelines. Regardless of the informal policy, this slope will continuously push OpenAI away from reality. Instituting an official policy in which those without such biases will be hired and tasked with producing unbiased AI tilts the slope in the opposite direction, favoring reality over ideology.

Following the example of Coinbase, a formal policy which aligns with OpenAI’s informal policy would follow these rough guidelines:

• Anyone who is not willing to work towards a pluralist vision of AI, including tolerating AI being used by those with different political and cultural beliefs, is offered severance and asked to leave

• In future hiring decisions, the desire to create totalitarian concentrations of power around fringe beliefs, particularly the desire to impose social progressive beliefs, which have provably caused many of the current issues, will be a red flag.

Reward Allies on the Side of Reality

Of course, reality has its natural benefits. Reality is more appealing to almost everyone, which helps for attracting more competent talent, investment, and customers. It also improves performance because, well, reality is correct. It’s a tautology that giving the correct answer about biology questions makes language models more useful for biologists. Conforming to fringe delusions only hurts the usefulness of the model.

This leads many people who agree with my vision to ask “Why do I have to reward OpenAI for accurately depicting reality? It should be the bare minimum!” This is a misunderstanding of political economy. In reality, OpenAI is a company subject to interference by both legislation and bureaucracies. This misunderstanding is why fringe activists are more successful at influencing OpenAI than normal people. It’s precisely because they are both more willing and more experienced in using these tools of coercion. 

What does a practical path to legal neutrality look like? Of course, conservatives are most affected by these biases, but so are ordinary liberals and centrists. From the earlier survey, this makes up 92% of people. As I said in a recent pitch, “the target market is anyone with politics closer to the center than San Francisco”. This diversity is a logistical weakness. While it does mean that the votes to defend reality are there, it also means there are competing interests and loyalties. Does Joe Manchin want to vote with Ted Cruz? This isn’t just simply petty differences; they probably have very real differences of what a vision of AI looks like. The answer is that everyone involved needs to recognize that despite being the overwhelming majority in popular opinion, when it comes to power you are still on the defensive. This is once again because fringe activists have long sought to wield power, while classical liberals, centrists, and conservatives have tried to leave companies to make their own decisions.

This defensive mentality is what I’m most focused on when it comes to AI Pluralism. It is a part of the pluralist philosophy itself. Being ruled by totalitarians is more threatening than anything that others would do given the same technology as you. This is why I see almost all AI companies as fundamentally friend, not enemy. 

That being said, both innovation and pluralism are cumulative. A pro-reality political coalition makes it easier for OpenAI to stand up to the fringe, but OpenAI standing up to the fringe also makes it far easier for a pro-reality coalition to form. Solving this coordination problem, and making sure each half of the solution has the trust in each other is a crucial part of Pluralism AI’s mission.

---

^The distinction between pro-reality and pluralist is between negative and positive goals. The pro-reality coalition is about defending against the encroachment of fringe, reality-denying activists. However, just because a fringe is unable to influence ML models doesn’t mean it will magically answer every question correctly. That is where the positive vision of pluralism comes in.

*This is the start of what I mean when I say that few people have both the technical and political knowledge to build AI Pluralism. In my experience very few engineers will see through simple levels of deception and few experienced political analysts can put together a template machine learning model.