This entry records one chain of questions from two angles side by side. Each question splits into two blocks: how the question got reached and what was observed, and the factual content of what AI answered at the time. All five questions came out of a datasheet assignment. My accuracy was 91%, the highest in the class was 95%, and under the constraint of avoiding overfitting nobody got everything right.
# One: Is the Risk of Trusting a Person the Same as Trusting AI?
Neither 91% nor 95% is a perfect score. People get things wrong, and so does AI. With the error rates sitting that close together, whether trusting a person carries the same risk as trusting AI became the first question.
- High-credibility settings: data analysis, language translation, code generation, first-pass medical image screening. Errors in these are relatively easy to verify.
- Settings that need care: resume screening, credit scoring, recidivism risk prediction. AI reproduces the bias sitting in its training data.
- Closed, single-task model output is verifiable and credible. Generative AI is a probability predictor and will hallucinate.
- Facts arriving without a source attached (numbers, dates, legal or medical assertions) should not be taken at face value.
- Trust should be built on repeated calibration, and it depends on the user's ability to verify.
- Whatever gets typed in may end up as training data, so private information should stay out of it.
# Two: How Much Similarity Is There Between AI and People?
Once trust holds conditionally, the next question is how similar AI and people actually are. With enough similarity, the trust standard applied to people every day could in theory transfer to AI.
- AI has no consciousness, no desire, no moral sense. Its position is that of a tool.
- Humans shape AI through training data, deployment context and legal constraints.
- AI reshapes humans in return: more reliance on external memory, repetitive skills losing value, judgement and creativity gaining value, and the form of human interaction shifting.
- Responsibility is asymmetric. AI carries none for its own behaviour, and the weight falls on developers, deployers and users.
- Humans should hold the supervisor role.
# Three: Where Do People Tend to Place Their Trust?
One observable fact: with error rates close together, people hand responsibility to a person who cannot be controlled, and extend lower trust to AI that can be controlled and traced. Every AI step leaves a record and can be walked back to a specific stage. Most human decision processes cannot be reconstructed in full. That gap forms the third question.
- Human trust splits into three kinds: predictable consistency (probability built from repeated verification), aligned intent (their goal matches yours), and a traceable chain of responsibility (someone is held to account when it breaks, and the system gets improved).
- What humans trust most is a verified fragile order: proven reliable within a limited scope, with a mechanism for correcting errors kept in place.
- AI lacks all three at once: it updates and hallucinates, holds no internal intent, and scatters responsibility when something goes wrong.
- The sharper question is whether a system can be designed so that people trust the process of using AI.
# Four: Will This Wave of AI Turn Out to Be a Bubble?
With what people actually trust clarified, the question goes back to the ground: whether this half-trusting state can hold up the current wave. I asked AI for three reasons it would be a bubble and three reasons it would not, then sorted them myself.
Three reasons it would be a bubble
- The gap between promise and reality: AGI stays out of near-term reach, stuck on causal reasoning, energy consumption and hallucination.
- Missing profit models and missing killer applications, with valuations resting on imagined futures.
- A collapse of trust triggering a regulatory and liability storm, potentially freezing the industry.
Three reasons it would not
- The underlying technology is already embedded in manufacturing, logistics, healthcare and finance, delivering real efficiency gains.
- Labour processes have already been reshaped around AI, and that shift runs one way.
- National security and the arms race create hard demand decoupled from the commercial cycle.
- The core distinction: the high-valuation narratives chasing AGI and human replacement may burst, while the narrow AI tools embedded in industry will stay.
# Five: Does Today's AI Still Sit Inside the Hard-Coded Category?
The AI that companies actually adopt, automation like n8n for instance, handles fixed and definable processes. The premise of adoption is that it works inside a certain region. Society's trust in people rests on the assumption that people grow. Set against that, people are also grown into their current state by experience and data, which forms the last question: whether people and AI sit in the same category.
- Whether it counts as "AI" depends on the definition: measured by understanding and consciousness, it does not; measured by "completing tasks that used to require human intelligence," it does.
- Narrow AI holds patterns without understanding: recognising a cat means having learned the correlation between pixel arrangements and the label "cat," and a cartoon cat may well get the same verdict.
- More precise terms for it: statistical learning system, function approximator, pattern matcher.
- It is practically useful and will displace part of human labour. Measured against human-level intelligence, it falls short.
Laid out in order, the five questions form two parallel lines: one tracks the observations derived from the assignment, the other tracks what AI answered for each question. The two lines meet at a single point — the premise that humans are more trustworthy than AI holds unsteadily once error rate, traceability and being constituted by data are all on the table.
Note
This entry starts from a datasheet assignment scored at 91% accuracy. The five questions appear in the order they were actually asked, each pairing my own line of reasoning with what AI answered.