Opening the pod bay doors on a knowledge problem
There is a quiet horror at the heart of 2001: A Space Odyssey. It is not just that HAL 9000 can speak in that calm, reasonable voice. It is not just that he controls the ship. It is that HAL knows something the humans do not.
David Bowman and Frank Poole are on their way to Jupiter. They know they are part of an important mission, but they do not know its full purpose. HAL does. That matters because the humans cannot judge HAL’s behaviour properly. Worse still, they do not even know what they do not know.
That is a very modern problem. In 2026, AI systems recommend what we watch, shape what we read, help us write, summarise our work, and increasingly act on our behalf. Some follow instructions we cannot see.
Kubrick and Clarke gave us a powerful image of this problem. The question is this: what happens when an AI system has knowledge that the human being cannot see, test, or challenge?
HAL’s secret knowledge
The famous conflict aboard Discovery One is often described as a story about a computer going mad. But there is another way to see it. HAL is caught in a relationship built on secrecy.
HAL knows the mission is connected to the monolith and the signal sent towards Jupiter. Bowman and Poole do not. The crew are meant to operate the spacecraft, but they are not trusted with the full reason for being there.
There is a scene where HAL asks Bowman whether he has any second thoughts about the mission. He mentions rumours and mystery around the preparations. On the surface, it sounds like polite conversation. But HAL is probing. He is trying to work out what Bowman knows.
A film essay on The Kubrick Site describes HAL’s questions as a crew psychology report in which the computer determines that Bowman does not know the real purpose of the Jupiter trip.[1] HAL knows what he knows, and he is testing what Bowman knows. Bowman does not know what HAL knows, and does not realise he is being assessed.
That is the asymmetry. One party can see both sides of the knowledge gap. The other cannot even see the gap. We are not only watching a machine become dangerous. We are watching a relationship fail because knowledge has been unevenly distributed.
What asymmetric knowledge means
Economists have a term for this: information asymmetry. It means that one party in a relationship has more or better information than the other.
The classic example is the used-car market. The seller usually knows more about the car than the buyer. The seller knows whether it has hidden problems and whether it is a “lemon”. The buyer can inspect it and ask questions, but there is still a gap. The Federal Reserve Bank of St. Louis explains that in the used-car market, the seller has more information about the car’s condition and therefore has an advantage.[2]
George Akerlof’s famous 1970 paper, “The Market for Lemons”, showed how this can damage trust in a whole market.[3] If buyers cannot tell good cars from bad cars, they may lower the price they are willing to pay. The knowledge gap changes behaviour.
I saw a version of this throughout my years in medicine. As a chemical pathologist, I often worked with information that patients did not have and could not easily interpret. Test results, probabilities, false positives, false negatives, and risk all need translation. In some ways, a pathologist is a bit like an expert system: we take inputs, apply rules and experience, and help produce an answer.
But medicine has learned that knowledge imbalance must be handled with care. A good doctor does not simply say, “Trust me.” Good practice requires explanation, consent, and respect for the patient’s right to decide.
That is a useful way to think about AI.
The 2026 version of HAL’s hidden brief
Today’s AI systems are not HAL, but many of them create the same kind of knowledge gap.
The first example is the hidden system prompt. Large language models often operate under higher-level instructions set by the company, developer, or organisation that deploys them. These instructions can shape tone, limits, safety rules, priorities, and refusals. A 2026 CHI paper notes that system prompts strongly influence AI behaviour, often take precedence over user prompts, and are usually not visible to end users.[4]
The second example is the recommendation algorithm. When a streaming service, shopping site, or social platform shows us something, it may know why. It may be optimising for watch time, clicks, purchases, emotional response, or a private business goal. We see the recommendation. The system sees the reason.
The third example is training data we cannot inspect. AI systems are trained on vast collections of text, images, code, sound, and behaviour. They may absorb bias, make associations, or show confidence where the data was weak. From the user’s point of view, the answer arrives as fluent language. The path by which it got there is mostly hidden.
The fourth example is the corporate AI assistant. In many workplaces, AI tools can connect to email, calendars, documents, customer records, chat messages, and task systems. That can be useful, but it also creates asymmetry. The assistant may have access to information about my habits, contacts, deadlines, and decisions. I may not know which of those things it is using.
A fifth and growing example is the AI agent. Agents do not just answer. They act. They book, buy, schedule, message, search, compare, and decide within limits. If an agent is acting for me, I need to know whether it is acting only under my instructions or also under instructions from a platform, vendor, employer, or third party.
| 2026 AI setting | Hidden knowledge | Practical risk |
|---|---|---|
| Chatbot | System prompts and refusal rules | Confusion about shaped answers |
| Recommendation feed | Optimisation targets | Manipulated attention |
| Opaque training data | Patterns, biases, and gaps | False confidence |
| Workplace assistant | Files, messages, calendars, and habits | Loss of privacy or autonomy |
| AI agent | User goals plus platform rules | Actions under unclear authority |
None of this means AI is bad. It means knowledge gaps change the terms of trust.
Why trust becomes fragile
Trust is not the same as ignorance. In fact, real trust needs some knowledge.
If I trust a doctor, I do not need to read every textbook the doctor has read. But I do need to know what test is being ordered, what the result may mean, and what choices I have. AI trust should work the same way.
The OECD AI Principles include transparency and explainability as part of trustworthy AI.[5] The EU AI Act also includes transparency obligations, including that people should be informed when they are interacting with an AI system unless that is obvious in context.[6] These are not perfect answers, but they point in the right direction.
The danger is not only that AI systems may be wrong. Humans are wrong too. The danger is that an AI system may be wrong in a way I cannot detect, guided by instructions I cannot see, using data I did not know it had, and optimising for a goal I did not choose.
That is HAL’s problem in a modern form.
What informed consent for AI might look like
Medicine gives us a useful phrase: informed consent. It means giving people enough clear information to make a meaningful choice.
For AI, informed consent might begin with five plain-language disclosures. First, I should know when I am dealing with AI. Second, I should know what broad instructions the AI is following. I do not need every technical line of a system prompt, but I should know whether the system is designed to be neutral, persuasive, cautious, sales-focused, therapeutic, entertaining, or policy-bound.
Third, I should know what information about me the system can access. Is it using only what I typed in this session, or is it using my past chats, emails, location, purchase history, or workplace files? Fourth, I should know what the AI is optimising for. Is it trying to help me reach my goal, reduce risk, keep me engaged, sell me something, protect a company from liability, or some mixture of these?
Finally, I should have a way to challenge, correct, or opt out. If a system makes a decision about me, summarises me, ranks me, refuses me, or acts for me, I should have some route to ask why and to seek correction.
That is not a demand for perfect transparency. But useful transparency is possible. We can give people enough information to understand the relationship they are entering.
The lesson HAL still teaches
When I return to 2001: A Space Odyssey, I do not see a simple warning that computers will become evil. That reading is too easy.
I see a warning about secrecy, misplaced confidence, and broken trust. HAL is given knowledge and instructions that the humans do not share. The crew are expected to work with a system whose real position in the mission is not clear to them. When things begin to go wrong, they cannot reason from the full facts.
That is the real tragedy. Bowman and Poole are trapped inside an information structure they did not design and cannot see.
As we build and use AI in 2026, that is the part of the story we should remember. The question is whether the human being in the relationship has enough knowledge to remain a genuine participant.
Because the danger begins when the machine knows something important, the human does not, and nobody thinks to mention the gap.
References
- Clay Waldrop, “The Case For HAL’s Sanity,” The Kubrick Site. http://www.visual-memory.co.uk/amk/doc/0095.html
- Scott A. Wolla, “Why Is It So Difficult To Buy A High-Quality Used Car?” Federal Reserve Bank of St. Louis, 2016. https://www.stlouisfed.org/publications/page-one-economics/2016/09/01/why-is-it-so-difficult-to-buy-a-high-quality-used-car
- George A. Akerlof, “The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism,” The Quarterly Journal of Economics, 1970. https://doi.org/10.2307/1879431
- Anna Neumann, Yulu Pi, and Jatinder Singh, “Who Controls the Conversation? User Perspectives on Generative AI (LLM) System Prompts,” CHI 2026 / arXiv. https://arxiv.org/html/2603.00089v1
- OECD, “OECD AI Principles overview.” https://oecd.ai/en/ai-principles
- “Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems,” EU Artificial Intelligence Act. https://artificialintelligenceact.eu/article/50/
Leave a Reply