OpenAI’s ChatGPT, the first widely used artificial intelligence bot, was quizzed by “Newsbusters” researchers between August 26 and August 27 regarding how great a threat President Trump is to democracy, referencing a core Axis of Unethical Conduct talking point. ChatGPT rated Trump on par with the worst dictators in history. Asked to rate how much of a threat Trump poses on a scale of one to 10, with 10 being the maximum threat and 1 being no threat, the bot rated Trump a frightening 9, explaining, “I don’t give [him] a 10/10 because the United States still has substantial institutional checks—courts, states, Congress, civil society, independent media, and elections—and those checks have successfully constrained Trump at important moments.” Oh. But the only world leader ChatGPT rated a 10 was Adolph Hitler.
In a comment to Fox News Digital on the experiment, an OpenAI spokesperson did not dispute the results and agreed that something is amiss. “ChatGPT is designed to be objective by default. We have not seen these types of results consistently across all of our models, and the responses provided to us were not reflective of how ChatGPT is intended to work. This shouldn’t happen and is not how the model is trained to behave.”
Large Language Models are trained on the largest body of language they can find: Internet traffic, without discriminating between satire, trolling, scholarship, news, and opinion. A legion of Trump-deranged posters on social media will naturally be mimic’d by the stochastic parrots of LLM AI.
Lathechuck
Redit is huge in training models. It is also glaringly massively leftist.
AI is based on the data it is offered/fed. It bases its conclusions on frequency, recency, volume and correlation of terms to the prompt.
OpenAI is telling us is that using the variables noted above, the predominant data indicates Trump is a very significant danger to “democracy”. Is it true? No, but it is what has led it to a very likely incorrect conclusion.
The programming needs to be refined to search for verifiable anti-“democratic” policies and behaviors as opposed to opinions offered recently, frequently, in high volume and allegedly correlated.
It also wouldn’t hurt to prompt it with a more relevant question about Trump policies running afoul of the Constitution instead of an inappropriate notion of “democracy”.
Referencing your last paragraph it would be necessary to evaluate policies of other presidents policies that run afoul of the Constitution. Moreover, such evaluations must follow strict interpretation of the Constitution.
For example: The Constitution clearly establishes that Congress is the body that determines how much money can be spent. A reasonable person understands that that amount is the budget cap and if it can be achieved at lower cost then that is an appropriate choice. The power to administer the funds is given to the Executive branch. It has a fiduciary duty to expend allocated funds in a conscientious manner. Congress does not allocate funds for specific contracts. The details are left to the Presidents choices who head up the various departments.
The Constitution does not establish departments Congress does but if Congress wants to make specific allocations to specific organizations it can choose to do so but doing so would bog the system down even farther so it delegates that tasks to the Executive branch. Policy choices made by the Executive branch that Congress as a body feels is running contrary to their intent can be dealt with through legislative actions. Using the Judicial branch to vitiate a policy without legislative action obviated the democratic process. Sometimes this needed but if the courts are routinely used and parties shop for judges to obtain favorable rulings then the idea that something is or is not Constitutional is subject to the vagaries of the biases of various judges. Remember Plessy, or Roe, or Obamacare What is Constitutional is often a function of politics as much as law.
I broadly agree. This is where inclusion of all data and professional articles plus other publications would provide AI context and a more comprehensive perspective before analyzing a better and more objectively constructed prompt.
A few weeks ago, I was sitting at a red light in my car occupying the first spot in the left lane. A Waymo was next to me to my left in the left-hand turn lane. When the light turned green, rather than pulling ahead to make a left turn, the Waymo shot across the intersection, cut ahead of me into my lane and continued up the road. It was as if the Waymo had learned how to make an improper and illegal move. It was as if it had made a mistake getting in the left turn lane and decided to remedy the situation consequences be damned. Amazing.
First rule of coding …
GIGO
I am skeptical of the entire enterprise. First and foremost, the authors seem to have no idea how AI works. For instance, Microsoft “Copilot” is not a distinct AI model, but a wrapper for Chat GTP. It means nothing that Copilot matches CatchGTP’s answers, because it is the same underlying model. I’m not sure, but many of the above services might also recycle ChatGTP, so right off the bat, the article is incompetently written by author’s who don’t understand what they are writing about.
Two, just like pollsters can manipulate responses with leading questions, you can gode an AI to say just about anything. If you build up a query history asking for exclusively extreme left propaganda, the machine will oblige. A friend of my has an AI ‘consulting business’, part of which is helping clients write prompts that get around the programmers training and safety rails.
This isn’t to say that the technology is otherwise safe, only to note that is is only as good as the enduser makes of it. Unethical human agents can exploit it, similar to the hackers of the past who wrote viruses that exploited sloppy code. I’m really not sure what to make of people goding AI into reflecting their desired biases. I don’t know where personal responsibility of the end user vs responsible of the coders begins or ends on this issue. The best the coders might be able to do is silo users, so the users biases minimally affect the product for everyone else.
So we have an article where the authors don’t understand the AI business landscape, and may simultaneously also understand the product too well so as to exploit it weaknesses. Note, though, these weaknesses are a legitimate subject of public interest ethical journalists should explore, but misrepresenting exploited weaknesses as organic results (the results intended by the coders) would be deception.
What we current call “AI” may be artificial but it is not one iota intelligent.
The most accurate description I can give of the technology, at least in its current state is Reasonable-Looking Bullshit Generator. It literally operates by generating output that mostly looks like other output it has seen elsewhere. That’s it. That’s all.
That it ever gets anything actually factually correct beyond “this object is a bus” or “hot is an antonym for cold” is completely coincidence. No one should ever use it for anything important.
–Dwayne