The truth and nothing but the truth

This is the chapter I chose to open my series with: a truth about what I think, a truth about AI, and a truth about AGI.

I began writing for three main reasons:

1. I miss the days when I wrote popular science articles on physics between 2017 and 2019.
2. Writing has always helped me bring order and clarity to my thoughts.
3. Sharing knowledge expands our networks, making it easier to connect with like-minded individuals and to receive timely feedback.

In the following lines, I aim to keep things concise while sharing what has been on my mind recently. This will serve as the starting point for this series of “articles” (short publications), which will cover topics I consider important at least to me. In his book A Brief History of Time, Stephen Hawking remarked that including a single equation could halve his readership. I’ll draw a parallel to that: instead of delving into mathematics or technical minutiae, I will, as much as possible, use plain language and draw on analogies when necessary.

AI is undeniably reshaping our world. What I find particularly ironic is the narrow perspective held by most users. (By “users,” I refer to anyone without expertise in how an AI model is designed, trained, or evaluated. The term is not meant pejoratively; it simply distinguishes those involved in the technical development of AI from those focused on its practical or commercial use.)

Within this group of users, a clear divide emerges between the pro-AI and the anti-AI. On one side are those eager to showcase AI’s capabilities and achievements; on the other are those waiting for the perfect opportunity to highlight its shortcomings and limitations what I like to call the “I told you so” users.

My stance lies in neither camp. Coming from a scientific background, I see little value in blindly supporting or vehemently opposing a concept. A scientist must maintain an objective perspective and a flexible mindset.

In the following lines, I will analyze what “truth” means in the context of AI. I will examine the extent to which AI can deliver truth, emphasize the importance of recognizing which truth is being delivered, and highlight the risks of relying on a one-sided truth. In subsequent sections, I will discuss the ease with which human societies can be manipulated, and I will conclude this chapter by proposing measures for the effective and responsible use of AI.

AI acts as a democratizing tool: everyone now has access to a teacher one that adapts to the learner’s pace. With just a few simple prompting strategies, such as:
“Explain X concept to a child, then to a [current position].”
one can quickly grasp new ideas and acquire knowledge, especially when paired with exercises that build progressively. Generative AI is becoming increasingly effective as a teacher. However, this strength applies primarily to factual knowledge disciplines such as mathematics, physics, and most of the natural sciences domains that can be classified as general truths.

Personally, during my work in physics, I developed a different view of scientific truth, which I will touch upon at the end of this article (not to confuse you at the outset, but rather to open your eyes to other perspectives)*.

When it comes to other truths for instance, historical truth AI depends heavily on preexisting information and processed data. If someone asks, “What happened in Tiananmen Square?” the answer varies depending on the language model. Models developed in China provide censored responses, whereas those built in Western countries typically offer more direct accounts. But this does not mean that Western models are inherently clearer or more honest. Almost every major tech company employs some form of content moderation, whether it be industry-wide refusal mechanisms or filtering for harmful, illegal, or privacy-sensitive material. Consequently, all models are subject to political bias. Moreover, English-language models often exhibit a left-leaning orientation, even when trained on datasets that aim to capture objective truths [ref][ref].

Thus, the notion of a single truth collapses. There is no absolute truth accessible through AI. It falls upon us to connect the dots, critically evaluate information, and discern whether it represents a fragment of the broader truth.

This might sound promising at first: instead of being confined to a single narrative, we now have multiple models presenting different “truths.” Yet for the masses (and I stress again, I do not use this term pejoratively), this is both critical and extremely dangerous.

AI, unlike anything before it, is the true master of pattern recognition. And by “pattern recognition,” I don’t just mean predicting outcomes based on given circumstances. Psychology has long been effective at such predictions e.g., given condition X, influenced by factor Y, the subject will likely exhibit behavior Z. AI, however, introduced a subtle but radical shift: it doesn’t merely use fixed conditions, it learns from its environment, continuously monitoring behavior to anticipate outcomes. In practice, this means: given environment X (deeply studied by AI), I (the AI) expect behavior Z if I introduce influence Y.

At today’s level of advancement, AI can already anticipate when someone is getting bored, seeking attention, or about to engage in a certain activity. This makes it a near-perfect tool for social engineering, psychological influence, and persuasion [1][2][3].

And we’re not even speaking yet of deepfakes or identity theft where a famous figure could be fabricated into saying or doing something controversial, sparking mass reactions, destabilizing governments, or crashing corporations (far more catastrophic than a celebrity scandal at a Coldplay concert 😅). The real danger lies in mass manipulation. With a single prompt, thousands of individuals could be compelled not to act, but to react to misinformation or even to misleading but technically valid arguments. (By “delusional conclusions,” I mean situations where the premises are correct but the reasoning is distorted, creating the illusion of truth.) This dynamic will only deepen the divide between those who hold power and those being subtly used and manipulated.[5][6]

Consider the case of Heaven’s Gate, the religious movement led by Marshall Applewhite and Bonnie Nettles in the 1970s, infamous for driving its members to mass suicide [4]. Now imagine such a movement at scale tens of thousands of people unknowingly bound together, subtly manipulated daily until AI systems are capable of inciting them to commit crimes, whether against themselves or others. With today’s technology, this isn’t just possible it’s frighteningly easy.

[My personal prediction:] Our society is heading toward fragmentation not complete collapse, but extreme compartmentalization. Violence may not reach dystopian extremes, but ideological divides will sharpen, tolerance will wither, and radical movements will proliferate.

I’ll close this chapter with a warning:

So, what follows this announcement? At least from my side, I try to remain optimistic and view the future as a stepping stone toward technological and societal progress. Pessimism rarely leads anywhere productive; it tends instead to produce noise, shouting, and chaos rather than practical solutions.

Let us, then, consider the brighter side: if we all recognize the repercussions of this technology, we can at least become more mindful of how we use it and of how it uses us. As I often say: “If you understand the magician’s trick… well… there’s no magic left.” Understanding is key. Being aware that everything fed to us through the tools and channels we use daily deserves deeper scrutiny is the first step.

One practical approach is to integrate references into search tools and LLMs (e.g., “According to X, Y is believed to be Z”). This would provide not an absolute truth, but at least a transparent and grounded answer. Otherwise, we risk relying solely on benchmarking tools (which themselves may carry bias) or on collective ranking systems (similar to Reddit, where answers are scored). Ranking itself is a fascinating subject, tied to algorithms and information spread if you’d like my perspective on it, I’d be glad to share.

Looking ahead, the move toward domain-specific models will convey more reliability. For example, a model specialized in German law would almost certainly be more accurate than a general-purpose model trained broadly on legal material. (This hypothesis assumes equal training data size but differing data quality general versus specialized.) Such models would also allow domain experts to meaningfully assess and validate accuracy.

On a deeper level, what we truly need is algorithmic reasoning [7][8]. Some already experiment with “chains of thought” in prompting, but we should go further investing not only in prompt design but in new tools, frameworks, and policies to advance reasoning itself. In mathematics and science, this could be integrated naturally: from deductive logic to hypothesis testing and exception handling. For philosophy, however, far more complex structures must be developed if we are ever to approach higher intelligence what we today call AGI.

In the next chapters, I will focus on understanding reasoning, free will, and critical topics such as consciousness and creativity.

If you’ve made it to the end: thank you for your time. Please share your thoughts or questions with me it is only through interaction that I can judge whether my writings bring value to others.

Written with love and care.

Mouad

*PS: As I mentioned earlier, even the scientific truth carries traces of bias or, more accurately, the scientist does. During my bachelor’s years, I observed two recurring forms of bias. The first is confirmation bias: a theory is constructed merely to reinforce a previously assumed fact, and any contradictory data is dismissed as noise something unworthy of attention. The second is projection bias: a phenomenon is forced onto a theory, presuming its validity before it has been properly tested. Think, for example, of blackbody radiation and the presumed continuity of light in the late 19th century.

There’s much more to say on this, but the key point is this: there is no absolute truth only relative truth. And here’s the part that may bend your mind a little: relative truth itself is conditioned by the observable frame. The moment we alter the conditions of observation… guess what? The truth shifts with it.*

[1] “Political Bias in Large Language Models" (Choudhary, 2024) Article link

[2] "Measuring Political Bias in Large Language Models" (Bang et al., 2024) Article link

[3] "LLMs are Left-Leaning Liberals: The Hidden Political Bias of Large Language Models" (2025) Article link

[4] https://en.wikipedia.org/wiki/Heaven's_Gate_(religious_group)
[5] “On manipulation by emotional AI: UK adults’ views and governance implications” Article link
[6] “Computational propaganda and misinformation: AI technologies as tools of media
manipulation” Article link
[7] “How Smart Are Reasoning Models in 2025?” (@LouisBouchard, 2025) (youtube video)
[8] “Algorithmic reasoning in large language models and neuro-symbolic architectures” (Flavio Petruzzellis, 2024) Article link