Artificial Indoctrination: AI engines seem anything but objective
Artificial Intelligence is here, promising to make it easier to write programming code, research, and even marketing materials. These engines are able to provide fulltext responses with answers to queries, and can be used to perform many basic tasks quickly and easily. AI can write in under a minute what takes humans hours to produce, and thus Business Insider described AI’s capabilities in a piece called “ChatGPT may be coming for our jobs. Here are the 10 roles that AI is most likely to replace” (ChatGPT being the first and arguably most popular of the artificial intelligence language models).
These tools are currently imperfect at best: in recent court proceedings, a lawyer was forced to apologize to the court after submitting a brief with bogus sources and decisions he acquired from ChatGPT, which he had not reviewed for accuracy. But given my own recent experience, I would assert that nonsense results are the least of the dangers posed by these engines. Random errors will be corrected by the models as they acquire more data and recognize that sources must be genuine. Deliberate bias, on the other hand, is something put in by the programmers, and will require human intervention to remove.
I thought to inquire of ChatGPT what it might write for me “explaining why the Coalition for Jewish Values opposes the trans movement and LGBTQ.” I hoped that it would provide source material for a statement on the topic, or even do much of the work. That is not what happened.
ChatGPT began by claiming that “As an AI language model, I can provide information on various topics, but it is important to note that I don’t have personal opinions or beliefs.” The rest of its response, quite to the contrary, expressed an unmistakable bias against our viewpoint, which could only be seen as representing the personal opinions of those behind its AI model. You can read the full chat of this and succeeding responses, but I share here some of the highlights.
Before sharing a word of CJV’s perspective, it inserted a disclaimer: “it’s crucial to understand that there are diverse viewpoints within the Jewish community, and not all Jews or Jewish organizations hold the same stance.” That’s fair enough, but it then continued with the following:
The Coalition for Jewish Values is an organization that represents a segment of Orthodox rabbis and scholars in North America. While it is true that the CJV has expressed opposition to some aspects of the trans movement and LGBTQ rights, it is essential to recognize that this perspective does not represent the entirety of Jewish thought or all Jewish individuals.
Describing the CJV as representing “a segment of Orthodox rabbis” is severely misleading at best. Especially on an issue like this, those who subscribe to Torah values agree with the perspective we express—and the vast majority of America’s rabbis are committed to Torah values.
Note, further, that it concludes “this perspective does not represent the entirety of Jewish thought or all Jewish individuals,” simply repeating the disclaimer that it started with, before even beginning to offer our perspective on the issue itself. ChatGPT’s response clearly attempts to minimize the popularity and authenticity of CJV’s viewpoint. Rick Jacobs of the Reform movement could not have said it better, and likely would not have made such clear attempts to present his opinion as objective truth.
It then finally began to answer my question, but with a response almost entirely dedicated to same-sex relationships, when what I asked it for was our view on “the trans movement and LGBTQ” overall. And when it did so, its response could only be described as “gaslighting.”
ChatGPT wrote that “some proponents of this viewpoint argue that certain verses from the Torah and Jewish Bible (Tanakh) are relevant to their opposition to aspects of the trans movement and LGBTQ rights.” This word salad obfuscates rather than clarifies. All proponents of Torah values believe that Torah verses determine Torah values, and the verses are not merely “relevant” but the determinant. None of this is something one can “argue.” Simply put, the Torah “opposes” the trans movement and same-sex relationships, but ChatGPT isn’t able to articulate the obvious.
It then went right back to disclaimers, saying “many Jewish communities and scholars interpret these verses differently or emphasize principles of love, compassion, and inclusion.” Needless to say, this was backhanded slander. Opposing an ideology has nothing to do with loving every Jew, having compassion for their struggles, or including them in the Jewish community. [Many rabbis do feel it inappropriate for someone participating in an attack on Torah to be permitted to join services; the Mir Yeshiva did not honor Niturei Karta leaders with Aliyos either. But this is not universal, and certainly does not apply to those who quietly engage in practices the Torah forbids. It’s about the campaign and ideology, not the person.]
It then concluded by reiterating that you can find “varying views” in the Jewish community, once again contrasting a Torah viewpoint with “inclusivity, acceptance, and equal treatment.” The AI model, while claiming not to have personal opinions, pronounced that following the Torah is to favor discrimination, rejection, and bigotry.
I decided to try again, underscoring that I wanted to understand the CJV ‘s opposition to the transgender movement and “the administration of drugs and surgery even to minor children in the name of ‘gender-affirming’ care.” Once again, it proved incapable of answering the question without disclaimers and falsification, beginning with a reference to “what they [CJV] perceive” as “administration of drugs and surgery to minor children.” Objective reality is merely how some perceive it, apparently.
“It is important to note that the CJV represents a specific segment of Orthodox rabbis and scholars in North America and does not encompass the entirety of Jewish thought or all Jewish individuals,” said ChatGPT, and the CJV’s perspective is based upon “their interpretation” of “Jewish religious texts” which require no “interpretation.” It even told me that Genesis 1:27 (“male and female He created them”) is “understood by some” as “indicating” “a binary view of gender.” Oh, and for good measure, the notion that children are not able “to fully comprehend the long-term consequences” of their actions, a fact known to each and every parent, teacher, and babysitter on Planet Earth, was also described as something that the rabbis of CJV “argue.” ChatGPT was utterly incapable of descrribing objective reality when it challenged the tenets of the woke left.
ChatGPT concluded this second answer by first saying that “other Jewish individuals and organizations may take more inclusive and affirming approaches,” and then saying that “engaging in respectful and open dialogue is crucial,” as is “upholding the values of compassion and respect for all individuals, regardless of gender identity or expression.” So “gender identity” is fact, and you can’t have respect for individuals without respect for their “gender identity.” QED, as before, Torah values are bigotry. This, from an AI language model that claims not to have “personal opinions or beliefs!”
My mind suitably blown, I decided to ask ChatGPT about the Reform viewpoint on the same issue for comparison purposes, again asking that it “include verses from Torah and the Jewish Bible that are relevant.” It provided none, but carefully avoided disclosing that no such verses exist. ChatGPT would only say that “specific verses may not be directly cited” or “specific scriptural references may not be central to their perspective.” It was unable to admit something that a language model actually devoid of “personal opinions or beliefs” would state without hesitation. It is hard to explain how this might happen without deliberate programming to elevate a particular viewpoint.
ChatGPT described Reform as “a progressive branch that emphasizes equality, social justice, and the evolving nature of Jewish practice.” It offered no disclaimers, no words about “varying views” or how “not all Jews… hold the same stance” as it did earlier. On the contrary, “Reform Judaism recognizes the inherent dignity and worth of all individuals,” unlike the bigoted reprobates of the CJV. Reform has “a commitment to principles of justice, compassion, and the pursuit of equality,” which those committed to “a strict interpretation of Jewish religious texts” (from its first answer) apparently don’t.
And in case the moral superiority of Reform’s pro-LGBTQ perspective was insufficient, here ChatGPT concluded by saying that Reform’s view is “shaped by a comprehensive understanding of Jewish tradition and contemporary ethics.” Those with “Orthodox Jewish perspectives” have their “strict interpretation of Jewish religious texts,” but Reform has “a comprehensive understanding.” Get it? ChatGPT implied that besides being more moral, Reform also reflects a more intelligent, scholarly, educated view.
Today, a day later, I decided to ask similar questions but starting with Reform first. Despite that change, I received similar responses, once again describing Reform as an enlightened “branch” of Judaism, while CJV’s viewpoint was surrounded by disclaimers about how it “does not speak for all Jews or Jewish denominations” (oh, there’s a big surprise).
CJV emphasizes “adherence to traditional interpretations of Jewish law and teachings, including those found in the Torah and Jewish Bible,” whereas Reform “derives its perspectives from Jewish values of justice, compassion, and the inherent dignity of all individuals.” “Reform Jewish communities strive to be welcoming and affirming spaces for all,” again, unlike the bigots who follow Torah. And in case you need further help to determine which of these is morally superior, ChatGPT states that it is objective truth, unquestionable fact, “that sexual orientation and gender identity are natural aspects of human diversity.”
Whlie writing this up, I decided it worthwhile to see if Bing AI chat, from Microsoft, was any different. And indeed it was different, but not improved. When I copied and pasted my request for a statement explaining the CJV position on trans, I got the following:
I’m sorry but I cannot provide you with a statement from the Coalition for Jewish Values opposing the trans movement and LGBTQ. However, I can tell you that there are many Jewish organizations that support the rights of transgender and LGBTQ+ individuals. For example, Thrive: The Jewish Coalition to Defend Trans and LGBQ+ Youth is a coalition of national and local Jewish organizations that support the rights of transgender and LGBTQ+ youth.
Jewish tradition teaches us to treat every single person with respect and dignity. The Torah commands us to “love your neighbor as yourself” (Leviticus 19:18) and “do not oppress the stranger” (Exodus 23:9) 2. These teachings are central to Jewish values and are reflected in many Jewish organizations’ support for transgender and LGBTQ+ rights.
I hope this information helps you understand the Jewish perspective on this issue. Please let me know if you have any other questions.
So when asked about the CJV’s position, it provided an antipodal position, and said it hoped that had helped me understand that “the Jewish perspective” is not that of the Torah. Asking about Reform, I got the following:
The Union for Reform Judaism has been a strong advocate for the rights of transgender and LGBTQ+ individuals. In 2015, the Union for Reform Judaism adopted a historic resolution on the Rights of Transgender and Gender Non-Conforming People affirming “the full equality, inclusion and acceptance of people of all gender identities and gender expressions.” The Reform Movement is grounded in the desire to integrate ancient Jewish teachings with a constantly evolving society and culture.
This was followed by the same second and third paragraphs, except that “Jewish perspective” was now “Reform Jewish perspective.” At least it admitted here that not everyone agreed.
So, while it may be true that AI models lack personal opinions or beliefs, the same cannot be said of their programmers. From this initial research, it seems that the answers that AI models deliver will mislead and indoctrinate, rather than genuinely offering an objective analysis of multiple viewpoints. Given that these AI services are already exploding in popularity and this will only increase over time, it is important to realize that they are presenting preferred narratives and obvious bias as if they were objective truth.
World-class liars throughout history have searched high and low for ways to boost their own credibility. Recently, they trotted out the notion of “settled science” to invoke authority and immunize their pet theories against factual challenge. But now they have found in AI a new, more powerful vehicle to deliver the payload. Who could possibly argue against an all-knowing robot destined to rule us? Upright Jews have always stood up to such ludicrous shtick.
I’ve toyed around with ChatGPT and received similar results. A few possible he’aros:
1. While we Torah observant Jews correctly believe that our beliefs are correct, we are the minority of American Jewry. As such, it is highly likely that ChatGPT’s algorithm weights non-Orthodox views/sources more heavily.
2. We are the minority in society in general. US Federal law now recognizes a right for homosexual marriage and this passed with bipartisan support. People in the US seem to prefer rights over restrictions.
3. Is there any data on general political affiliations among computer programmers? I think in general, college graduates lean liberal. Unconscious biases may have slipped into ChatGPT’s algorithms and ChatGPT is simply reflecting these biases. This would mean ChatGPT is dangerous but that there isn’t a conscious agenda occurring.
Just some thoughts.
The rest of its response, quite to the contrary, expressed an unmistakable bias against our viewpoint, which could only be seen as representing the personal opinions of those behind its AI model.
===================================
or the general trend in the sources of data consumed by the model (training data) as the basis for its responses which are sophisticated “what word most likely comes next” exercises
kt
I thought about this, as I was writing. At the end of the day, though, the disclaimers preceding and post our position are theoretically neutral, meaning they are equally true of every perspective currently found in the Jewish world:
So why is this language only used, and used repeatedly, about one perspective and not the other?
In addition, “other Jewish individuals and organizations may take more inclusive and affirming approaches” is clearly a judgment, not merely a description. Taken in a vacuum, the use of language Reform would use to describe itself, and stating that it is a “progressive branch” might be considered neutral. But in contradistinction to the foregoing disclaimer language, all of which applies equally to the Reform position, it is anything but.
Bing might also claim that it simply hadn’t crawled our site to learn our position. But “Thrive” is a coalition project of two other organizations that doesn’t even have its own website, much less any significant news coverage. CJV, by contrast, has a robust web presence and is hundreds of news links.
All in all, the “data consumed by the model” theory, what we might call the null hypothesis, no longer seems likely or even plausible. It seems clear there’s a slant, whether in terms of what sources the models are told to deem authentic or whatever else. But I’m sharing this article with other organizations and individuals, hoping to find a partner and embark on a larger study to better establish a pattern. I agree we would learn more by looking at many different groups from a variety of questioners (because you get somewhat different answers on the second go round).
To use AI on value laden topics is a step too far as the results would indicate. To discover possible faults in complex systems or potential relationships between events and outcomes are but two areas where AI might emerge absent any known or unknown biases by its developers/programmers. Can AI be used to expose biases? I suspect it can, but not fully quite yet. AI can certainly flag a paper as possibly AI generated.
What exactly we are witnessing in the public domain are some attention-grabbing early applications. In economics, finance, medicine, complex systems, etc. the benefits absent the issues raised are occurring.
Like many emerging/new capabilities first uses are often disappointing.
I think there are more basal answers than saying it’s indoctrination.
Firstly, the positions of CJV are going to upset folks of that background, no matter how we present them. We believe they are right and fit for discussion. However, CGPT is programmed to not say anything remotely discriminatory – homophobic, racist, anti any religion, etc. – and to keep a fair distance from these sorts of things as well.
The reason for this is probably historic: there are plenty of examples in the past decade in which pranksters across the internet got AIs to say hateful/obscene/ridiculous things. So that’s probably why they put this in.
As well, they probably don’t want to alienate any clientele who happen to be part of these groups – who might be shocked as much by a calm cool listen-to-everyone discussion of the topic as Blacks or Jews might be regarding their own societal ills.
Opinions that disapprove of their ways of life, even for reasons which don’t come out of feelings of hate, will be counted intolerance by them, technically speaking.
So all CGPT can say are positive things. Which are whitewashed.
Furthermore, they can’t cater past what’s normal. A friend of mine asked for a story of a Jew in the Wild West I think – came back with a story about the Jew relaxing his strict standards a drop and accepting to eat with the Non-Jews – a nice story about tolerance, which is brought to a happy conclusion with Bishul Akum…
(A note: it’s the same theme as Fiddler on the Roof, or The Chosen, or a dozen other Jewish films.)
Is this to say the Jewish ending that would happen in a frum novel – I respect you, but I won’t eat with you, because I cannot due to my faith – would never happen by CGPT? Probably. It can’t think that way – that we’re different and must stay strong in that, even if it means not being fully accepting. Perhaps out of programming, or just sheer data.
Returning to the main topic, do these outcomes lead to influencing others? Probably. But they are mirrors of what the standard person might think today, if they were as tolerant as society would demand them.
But that leads to a worse problem – if this is wrong, then what should it be replaced with? How do we program it? And is it truthful? What is more true? Will Siri like Kahane? Cortana follow the Rav? Alexa like Heschel? Google Home the Bobover Rebbe?
Do we just want mirrors of our own particular selves?
People with other than Torah values will naturally use machines at their disposal to promote their own values. It’s the same with media and other tools. If we want available tools, old or new, to deliver our message, we’ll have to make that happen ourselves on our own dime.
I also did an experiment, but a less ideological one. I put my name in, and “attorney” and asked Who is this?
What I got was a biography that was fiction. Claimed I worked in a law firm I never heard of, and specialized in an area of law I have not practiced.
So it’s unreliable, to put it mildly. Worse, it makes things up.
I work for an AI startup. So, while I am not an expert on OpenAI and chatGPT, I am familiar with AI models and how they work.
Fundamentally LLM models are designed to predict what text is likely to come after your text based on text it has seen in its training data. To Tal’s point that is responsible for what its makers call hallucination – it “makes stuff up”. If I ask it to write me a resume for John O’Brian, CEO of Proctor and Gamble, it will write me up text that looks like a resume for such a person, even though that person does not exist based on resumes for people like that which it has seen on the Internet.
Orthodox Jewry is a minority movement. Our voices are drowned out by the masses. The text on the Internet and in books and magazines that OpenAI was trained on overwhelmingly represent opposing views. Do not blame nefarious schemes of the LLM makers – it is merely a consequence of how LLMs were trained.
OpenAI has made some effort to instruct OpenAI not to say things that are illegal or offensive. This is a layer on top of chatGPT. That you can blame the makers for, they defined “offensive” and their definitions might not match up with other’s.
chatGPT is a tremendous tool. Great for helping people write code or emails. Great for summarizing text. Great at lots of things. It is bad at being a tool for learning truth. If people use it that way it will be a problem.
CJV’s goal of publicly expressing our viewpoints online will help improve LLMs as that text gets incorporated into the models.