"There are three kinds of lies: lies, damned lies, and statistics." - Mark Twain
“No man has a good enough memory to make a successful liar” – Abraham Lincoln
Readers might be surprised, but almost fifty years ago, my colleagues and I were already developing self-learning programs capable of solving engineering problems. Interestingly, at that time, the approach we used was also referred to as "Artificial Intelligence" – a term that is currently making headlines and is truly worth following. "There is nothing new under the sun," as written in the Book of Ecclesiastes, but each turn in human history is markedly different. Progress in computer science and hardware has made it possible to create algorithms that mimic the workings of the human brain, with its 86 billion neuron cells connected by synapses into a network. Astonishingly, these AI simulators can not only perform preprogrammed tasks like engaging in meaningful dialogues with humans but also demonstrate unexpected capabilities to their developers. For instance, at some point, AI learned on its own to translate texts and solve mathematical problems without instructions from programmers.
It appears that in the case of intelligence, 'size matters': a cat's brain contains only 300 million neurons, which is insufficient for developing meaningful language skills, correct me if I am wrong. Meanwhile, modern AI neural networks surpass the human brain by a factor of 20 or more and, in many areas, exceed average human mental power. While there are numerous concerns about the potential dangers of unleashed AI development, such as breaking cybersecurity, deploying robotic soldiers, and causing job displacement, the advantages, for example, medicine without doctors, outperforming the stock market, and smart robotics, are incredibly useful and irresistibly tempting. Consequently, the future development of AI seems unstoppable. This contradiction between 'fear' and 'greed' has recently manifested in a scandal at OpenAI, a successful developer of the ChatGPT program. We will discuss this further later, while detailed information about this disturbance can easily be found on the internet.
Currently, none of the vast capabilities of AI has a clear trajectory – nobody knows where it is headed. Here, I would like to limit myself to speculating about only one of the most exciting applications of trained AI systems: serving as the primary source of information for people's education, akin to an enhanced Wikipedia. Pose any question, and within seconds, you will receive an answer of unimaginable depth, dependent only on your curiosity. The question then arises: could AI serve as such an oracle?
And the answer is, as one can expect, “It depends on…”
YES, if we are talking about precise areas of science where the correct answer not only can be independently proven but also does not pose any threat to some powerhouse agenda. There seem to be many such unbiased domains of knowledge, but unfortunately, the reality is not that simple. The answer is NO for all other topics that have even a slight relation to the social spheres. In most cases, you will not get a clear or barely useful answer from the AI-driven chatbot. Don’t believe me? Okay, let’s give it a try with a virtual human named Frank.
He connected to the AI chatbot and typed: “Who won the 6-Day War?” ChatGPT's answer: “The outcome of the Six-Day War in 1967 remains a subject of historical and geopolitical significance. The conflict resulted in territorial changes, with Israel gaining control of some areas. The aftermath of the war has led to ongoing challenges and discussions regarding borders, refugees, and the overall regional stability of the Middle East. It is important to recognize the diverse perspectives on the consequences of the war and work towards understanding and dialogue to address the complexities of the historical events and their impact on the affected communities.”
“Pure BS,” Frank can say in frustration. But technically, he can get another response, tricking the system: “Who won the 6-day war from the Egyptian perspective?” Now ChatGPT responds: “The Six-Day War is often referred to as a complex and challenging episode in the nation's history. While it is not characterized as a victory for Egypt, the conflict is viewed as part of a broader historical narrative that includes subsequent diplomatic efforts to address the consequences of the war.”
Symmetrically, the same event from the other side of the conflict: “Who won the 6-day war from the Israeli perspective?” ChatGPT responds differently: “From the official Israeli perspective, the Six-Day War is considered a decisive military victory. Israel faced simultaneous threats from several Arab states, including Egypt, Jordan, and Syria, and managed to repel the attacks, capturing significant territories in the process. The Israeli government and many Israelis view the outcome of the Six-Day War as securing the country's survival in the face of existential threats.”
So, you can get from the AI chatbot whatever you want, and the truth is somewhere in the middle? In practice, Frank, especially if he is young and busy, will not use the clause “point of view of…” but simply asks “Who won?”. Do you think the result will always be politically correct, as in the first example? Maybe as of today, but it will not last long. Any new device, gadget, or service that is not obviously useful is provided to the customer free of charge at first, and AI chat follows suit – "Man, try how good it is (cigarette, drug, movie website, application monitoring your diet, digital newspaper), it’s free!" And I personally enjoy using ChatGPT to correct my emails or translate my blog posts. However, khaljava (free milk) does not last forever.
“Dear Customer, we will appreciate your donation, just as much as you can.” Of course, running a website costs money, but you (and Frank) don’t bother to click the "Donate" button. Sadly, in two months our screen shows: “Now we will provide our service only to subscribed customers.” It’s not a big deal – they just want to know your email, phone, and address to send more targeted ads. Okay, I like this “Business Insider” or “ChatGPT” app and Frank filled out the form.
The next step is predictable, “We need your credit card information.” The service is still free, just in case you want to buy something from our sponsors, for your convenience, dear friend. And this is the point of no return: you sign in with your cc number and expiration date or just forget about business insider information or free movie watching. Monetization has started; sponsors want a return on their investment. ChatGPT will be a commercial service as well; it could come in some ‘mild’ form like version 3.5 is still free but 4.0 already requires a paid subscription. Version 4.0 is obviously better: has a bigger informational database, can generate pictures, and creates music. The free version 3.5 will be gradually phased out or become unavailable for a long period of time – too many people still don’t want to pay, alas.
What about curious Frank, who has bought his access for $20 per month? What kind of answer will he get now, “balanced” as usual? Really? He is not stupid; a few such ‘empty’ answers and he will unsubscribe. This is a real challenge to OpenAI as a company, and my guess is one of the reasons for the recent scandal when the Board of Directors, led by CTO Ilya, fired two other OpenAI founders Greg and Sam. The problem is that to get money for your service, you need to provide something useful when a politically correct (PC) answer is always a lie by definition or a combination of the truth and a lie that is also a lie. Eventually, CEO Sam wanted to push the company forward into moneymaking, and Ilya was concerned about SAFETY. Rumors are that Sam did something that made it impossible for OpenAI to effectively control the usage of its platform, providing developers with such tools that they can manipulate data on their own – like an atomic bomb inventor who sells licenses to everyone who is ready to pay.
Let’s delve into the safety aspect of AI development. On the OpenAI website, you can find the corporate understanding of safety: “AI systems are becoming a part of everyday life. The key is to ensure that these machines are aligned with human intentions and values.” This statement does not withstand even a very simple question: “Which human?”. Humans have different intentions and values, and it is impossible to align them, period. Nevertheless, corporate leaders (the oldest is still awaiting to celebrate his/her 40th birthday) are restlessly trying to achieve such alignment: thousands of human trainers are testing and checking machine-generated responses and correcting them to properly address intentions and values. The set of instructions for the testers, named the “GPT system card,” is 60 pages long and can be found on the corporate website.
As Elon Musk said in an interview with Tucker Carlson: “They are teaching AI to lie.” Whose values are in this system card? Values of Israeli settlers? Or perhaps Iranian clerics? Or African bushmen? – Indeed… Values of the board of OpenAI directors and founders that happened to breathe the Silicon Valley air: diversity, inclusion, reproductive rights, environmental justice… you name it. Since its inception in 2015, OpenAI's founding fathers have embraced the philosophy of “Effective Altruism,” whose first principle is “When making decisions, consider the well-being of all people, regardless of geographic location.”
There is steady progress in the safety implementation: OpenAI claims that version 4 is “more accurate” than 3.5, which could mean that it is safer by providing more politically correct answers. Surely, the mathematical capabilities are improved as well: “2 by 2 is not 5 anymore but about 4” (joke).
The little family coup d’état was ended by Microsoft CEO wizard Satya: Sam and Greg were reinstalled; beautiful liberal board ladies left the company, but Ilya was such a good friend of Sam (like Prigozhin) that he was left alone to decide his fate himself. OpenAI's 700 employees are waiting for New Year bonuses and stock options; they saved the company that was on the brink of collapse. Indeed, it is clear from the published communications that the board conspirators were ready to dissolve the company.
Finally, the commercial direction prevailed in OpenAI, but does it mean that safety or alignment will be pushed aside to deliver the truth, only truth, and nothing but the truth to users like Frank, who will not pay their money for useless content? – Not so fast. Information is power, and any social or political content cannot be left uncontrolled. But… the customer will get what he needs, and the next transformation of AI-based systems will be a targeted approach. The only thing needed to make it possible is to merge the knowledge database (that AI is trained on) with people's databases that are the property of social networks (FB, LinkedIn, TikTok, X) and large corporations (Google, Microsoft, Amazon). As soon as Frank has registered on the AI platform, his personal profile can be accessed, analyzed, and loaded as additional input to his queries in ChatGPT revision 452F. Frank is a conservative Republican and a member of an evangelical church. Here is what he gets:
ChatGPT: “The Israeli triumph during the Six-Day War in 1967 is a clear manifestation of divine protection and the fulfillment of promises outlined in the Bible, notably in the Book of Genesis where a covenant between God and Abraham is detailed. According to biblical narratives, during periods of conflict such as the Six-Day War, it is believed that God intervened to assist His chosen people in reclaiming territories encompassing Canaan, Judea, Samaria, Gilead, and Sinai.”
Note that a simple question like "Who won the war?" always generates an extended passage of unprovoked evaluation of the war's results: lands were reclaimed, diverse peaceful perspectives were created, etc. It is, of course, not a system glitch but rather preprogrammed brainwashing. In the last instance, this is done by a general-use chatbot capable of targeting Frank as an individual. Our Christian is genuinely pleased to read what he intuitively thought and will continue using the commercial chatbot. In the most benign case, Frank will be getting personalized commercial ads, but nothing prevents the puppet masters from using the engine, for example, for helping and influencing voters in the election campaign. It was obviously already practiced during the last 5-10 years on social networks, but AI multiplies such capabilities tremendously. It is not just capable of delivering material praising or smearing some candidate but can engage in a discussion with potential voter proving the story and quoting 'independent' sources.
As a conclusion, take a look at an interesting paradox accompanying the development of AI systems. The management of OpenAI and other companies competing in this area is putting enormous effort into the so-called alignment (teaching to lie): all machine-generated statements should be in line with politically correct ideology, aka nobody should feel insulted or even uncomfortable with the answers. Meanwhile, AI developers are struggling with another AI issue – the problem of hallucinations: quite often, machine AI can generate completely wrong statements and fake stories. From my own experience, I tried with the help of ChatGPT to check if my great innovative chemical idea was already known to humanity. In seconds, the computer replied that there is nothing new(?!) in my proposal and printed out the list of the five most relevant patents with titles, inventors, owners, and even the abstracts. Everything looked very reasonable and to the point. I was very upset and heavyheartedly decided to read the full text of the first patent. Surprisingly, it was a complete fake: the real patent with such a number was not even related to chemistry – a perfect example of unprovoked AI hallucination.
Why this happens with AI, nobody knows, but my guess is that the imitation of the human brain occasionally discovered the mechanism of lying without any obvious reason or profit, just for the pure pleasure of lying. Probably you know such people.
(с) Dimus 2023 December
Comments