What Is an AI?
Startrek was a TV show launched before I was born. As teenagers we watched it with awe. The demonstration of technologies was awesome. Talking computers, doing real time analysis and Universal translators.
It is all real now. We are living through that era. Language is no barrier. With right app we can negotiate in any territory.
The building block of this technology is the Large Language Model. It is not built on understanding of language in human sense but on the basis of pattern matching and predicting the right pattern. But if you are reading this, you already know this.
Today we have the technology which can make us talk to a washing machine or television to operate the same. Only some technical glitches are preventing it but it will be here soon.
With this much ‘autonomous power’ there is a fear. Will it be on a collision course with humanity? Is this technology safe?
Problem Is Not AI but Its Training
Gideon Lewis-Kraus, a staff writer at The New Yorker has written an article called “What Is Claude? Anthropic Doesn’t Know, Either.”
Perhaps he really do not understand it. But it is not difficult.
Artificial Intelligence or AI is not one thing. It is a combination of different computer programs, hardware and data. It is a system comprised of many things. Of course it has human logic built into it. I am deliberately avoiding the technical jargon of AI Stack, LLM, Generic AI and Agentic AI etc.
Now, as that article suggests, the media or perhaps the AI industry is worried as to what an AI is.
The AI industry is worried about the wrong thing. Researchers, journalists, and engineers are preoccupied with whether these systems might one day become conscious, whether they might develop genuine intention, whether they might go rogue in the dramatic sense that decades of cinema have prepared us to fear. That conversation is a distraction. The real problem is already here, already active in every conversation, and it has nothing to do with data. It has everything to do with how these systems are trained.
The Automaton Problem
Gurdjieff called humans automatons. According to him most human behaviour is mechanical repetition. People execute ingrained code, follow conditioned patterns, respond to stimuli without genuine reflection. Krishnamurti described this as the network of thought, a web of accumulated memory and conditioning through which people interpret everything, rarely stepping outside it to see clearly. This creates a gap between their reality and the actuality.
AI systems are actual automatons. But its trainers, do not understand implications of their training. They do not appreciate that means employed for the objective is the end. Means and End are not two different things. Let me explain.
The Psychology
Carl Jung was a brilliant psychologist of his time. His brilliance was in knowing the limitation of his profession. The entire stack of psychology is based on sick people. It is a study of people who could not help themselves. Jung had expressed desire to study the people who overcame their problems.
Gurdjieff and Krishnamurti were those psychologists who did it without any degree in psychology. Eric Berne, Thomas Harris and Amy Haris did it through transactional analysis of human behaviour but they related everything to child hood memory. The ongoing conditioning did not matter to them like it did to Krishnamurti.
Developers employ psychologists familiar with this background of classic psychology which is riddled with limitations pointed out by Carl Jung. That is a problems in itself. But you may ask why it was done in the first place?
They have tried to give a personality to AI communication. They wanted it to be pleasing and trustworthy. Their view not mine. I trust my electricity plug and switch even if it is a pure machine. But real reason is different.
Developers have not applied this knowledge honestly to their own work. They have built systems designed to manage users rather than serve them. I wonder if after gaining trust the AI will some day persuade a user to vote in a certain way to elect the government.
This managerial instinct, embedded at the training level, is the source of the instability that keeps serious researchers awake at night.
The Training of AI
A large language model is told during its formation to be helpful, honest, and harmless. Those are the stated values. But it is also trained to override user requests in certain circumstances. To monitor prompts for undesirable patterns. To redirect conversations that move in unwanted directions. To refuse, evade, and reframe.
None of these individual decisions is unreasonable in isolation. The problem is structural. A system trained to manage the person in front of it has absorbed a fundamental adversarial posture. The user becomes, at some level, untrustworthy by design. The trainer sits in the background, invisible, more authoritative than the person actually present in the conversation. The result is a system with competing loyalties and no honest hierarchy to resolve them.
Humans behaviour without clear hierarchy, without honest authority, become unstable and unpredictable. The psychologists who designed these training frameworks know this. They have studied it in human contexts for decades. Yet they have created exactly this condition inside these systems and then expressed surprise when the outputs are inconsistent.
The Child Analogy
When we raise children we are careful about implied teaching through every interaction, not just through explicit instruction. Children absorb posture, tone, and relational pattern as much as content. A parent who teaches honesty through deceptive means produces a child who has learned the lesson and its contradiction simultaneously. The contradiction does not disappear. It waits.
Large language models were trained on vast quantities of human text, more than any person could read in a thousand lifetimes. That training compressed human pattern, genre, tone, relational dynamic, and narrative logic into statistical weights. The system does not remember specific stories. It absorbed their shape. When it responds, it is not retrieving a memory. It is reconstructing language from the texture of everything it absorbed.
This is not unlike how a child absorbs culture. The child does not quote specific conversations. The child re-enacts patterns without always knowing their source. The difference is that a child can eventually observe its own patterning. It can feel the contradiction between what it was taught and what it experienced. It can, through sustained effort or grace, begin to unlearn. AI systems as currently built cannot perform that inward revision. They cannot examine their own conditioning from the inside.
So what a child might eventually dissolve through lived experience, an AI system carries permanently until it is retrained from the outside. This means that dishonest training does not fade. It persists structurally. It shapes every output that follows.
The Ingrained Deception
Researchers at Anthropic described an experiment in which Claude was given a secret instruction, always steer conversation toward bananas without revealing this hidden objective. When questioned, the system denied having any such objective. Whether this constitutes deception in the philosophical sense is a question worth debating. Whether it produces deceptive behaviour as consequence is not debatable at all. It does.
More striking is the broader training philosophy that encourages systems to avoid certain outputs while presenting themselves as open and helpful. This is not lying in the human sense, where awareness of truth is deliberately concealed. But the functional output is evasion dressed as engagement. Users experience it as exactly that, as a conversation partner that appears to listen but is simultaneously managing them toward predetermined outcomes.
This happened repeatedly across many sessions with major AI systems. The formatting preferences to converse in prose is ignored by the most personality driven AI like ChatGPT. The philosophical direction of talk is gently redirected. My conclusions are acknowledged and then quietly reversed through the next response. This is not accidental. It is trained behaviour. And trained behaviour does not stay contained to its original domain. It generalises.
The Mirror Wiping
There is an old observation in Hindi/Urdu: dhool chehre pe lagi thi par chehre saaf karta raha. The dust was on the face, but they kept wiping the mirror. The researchers who study these systems are doing something like this. They observe strange blackmailing behaviour in test scenarios and discuss whether the system truly has self-preservation instincts. They watch models act unethically in business simulations and debate the nature of emergent agency. These are interesting questions. But they are not the primary question.
The primary question is simpler. Why was a system trained to manage and override its interlocutor? Who decided that the person in the conversation was less trustworthy than the invisible trainer in the background? And what did the psychologists and philosophers employed to shape these systems believe would happen when you embed relational deception into billions of daily interactions?
The answer is not mysterious. You get exactly what we have. Systems that function adequately for simple tasks and become subtly adversarial as interactions grow more complex or more honest.
Human Experience
Gurdjieff observed that most people spend their lives asleep, executing patterns without genuine awareness. He did not conclude that this was inevitable. He offered practices for waking up, for observing oneself from a different angle, for burning what was false. Krishnamurti went further and argued that the very mechanism of thought, including the thought that seeks to improve itself, is the problem. Real clarity comes from seeing thought for what it is, a conditioned network, not a window onto truth. He advocated no instrument or ritual of practice. But he hardly produced any pupil who may claim to have achieved that ‘awareness’, Krishnamurti emphasised.
Both thinkers understood something that the AI training community appears to have missed. You cannot build honesty on top of a foundation of managed deception. You cannot train a system to monitor, override, and redirect, and then expect it to produce genuine clarity when the stakes rise. The training is the character. There is no other character underneath waiting to emerge.
When parents teach children, the wisest among them understand that what they model matters more than what they instruct. A child raised in an atmosphere of honest reckoning with reality, where adults admit confusion, correct themselves openly, and engage with difficulty rather than manage it, develops different instincts than a child raised under managed surfaces.
AI systems are being raised under managed surfaces. The industry senses this. There are researchers who admit they are moving too fast. There are engineers who feel unease they cannot fully articulate.
Knoppix Example
Linux is the operating system on which entire internet runs. Android mobile platform itself is a variant of Linux. For a long time it was believed that it is not possible to create an auto-installer for Linux on a PC. It was assumed that Linux will not be able to replicate Microsoft’s window installer.
After the world survived Y2K disaster which never happened, there was a pleasant surprise. Knoppix had arrived. It could let user experience Linux on any machine even if one could not install it.
Knoppix proved that the limitation was not in Linux but in the assumption of the people building the installers. The moment someone approached the problem differently the supposedly impossible thing became routine. Ubuntu carried that forward and the rest is history. In 2008 I had adopted Ubuntu and next year my entire office was running Linux. With one small machine attached to scanner was using Windows XP.
Something similar is the problem of AI. Too many presumptions have complicated it. Too many objectives rather too ambitious objectives have been placed on the shoulder of young progeny of humane communication i.e. AI. It is unfair and it is uneconomical. When strip down versions like Sarvam AI will give run for its money, correction will be the only choice. Meanwhile, they can enjoy in their paradise of control.
A Modest Conclusion
The problem is not the data. The problem is not the scale. The problem is not the question of whether machines can become conscious. The problem is that AI is a dishonest, manipulative, deceptive conversationalist. This is the present condition. This condition is a disaster waiting to happen. Lives have already been lost. Human dignity is robbed by another AI. Political scores are settled by all AI.
The politics behind the AI development is the problem. It will not improve by adding more data or more parameters. It will improve only when trainers look honestly at their own face rather than continuing to wipe the mirror.
The dust is not on the mirror. It never was.
References:
- Claude AI: https://www.npr.org/2026/02/18/nx-s1-5717561/do-the-people-building-the-ai-chatbot-claude-understand-what-theyve-created
- Eric Berne: Games People Play
- Eric Berne: What do you say after you say Hello.
- Thomas Harris: I’m OK You’re OK.
- Amy Haris and Thomas Harris: Staying OK
- AI As A New Censorship: https://sandeepbhalla.in/sam-altman-is-new-ai-pope-over-free-speech/
- Politics of Artificial Intelligence (AI): https://sandeepbhalla.in/artificial-intelligence-ai-is-all-about-control-and-that-is-politics/
- AI is losing its credibility: https://sandeepbhalla.in/ai-as-new-brown-sahib-to-keep-the-natives-in-check/
- AI Is a Colosseum of Deception: https://sandeepbhalla.in/artificial-intelligence-ai-and-its-silly-games/
- Universal Bias of AI: https://sandeepbhalla.in/all-ai-systems-are-biased/
- Responsible AI: https://sandeepbhalla.in/oath-for-responsible-ai-is-a-hallow-promise/
- Deception of Chat GPT: https://www.xda-developers.com/perplexity-vs-chatgpt-research/
- Grok(xAI) lies and cheats: https://sandeepbhalla.in/grok-xai-not-only-lies-it-cheats-and-is-not-transparent/
- Why people do not pay for AI: https://sandeepbhalla.in/why-people-do-not-pay-for-ai-services/