Exam practice

• Intelligence is what a system can do: solve problems, learn, use language. It can be tested. Example: a translation model can be tested on how well it translates Nepali news into English.
• Consciousness is whether a system experiences anything, whether it is like something to be it. There is no agreed test. Example: whether a chatbot that says it is sad feels anything; its words alone don’t settle it.
• Responsibility is who answers for what a system does, and whether it can be wronged. It is decided by people, in ethics and law. Example: when a bank’s model wrongly refuses a loan, the bank answers for it, not the model.

The three can come apart: a system can be very capable and, as far as anyone knows, feel nothing, and however capable it is, people stay responsible.

For: if consciousness comes from how information is organized and used (functionalism), there is nothing special about neurons: a machine organized the right way would be conscious too. A 2023 study of AI against scientific theories of consciousness found no obvious technical barrier to building systems with the features those theories name. And we already accept that animals with very different brains are conscious.

Against: consciousness may depend on the biology of living brains, as Searle argued: a simulation of a brain no more feels than a simulated storm is wet. And a machine trained on human writing will talk about feelings whether or not it has any, so we may never have good evidence.

Own view: any view, if argued from the evidence, separates what is known from what is open, and answers the other side’s strongest point. For example: “I think it is possible but not shown, because no theory rules it out and none confirms it; so we should neither treat today’s chatbots as feeling beings nor rule out the question forever.”

Many people made choices that led to the crash: the engineers whose software kept misclassifying the pedestrian; the managers who left emergency braking to a single safety driver watching for hours, with the car’s own safety systems switched off; the state that allowed tests with weak oversight; and the safety driver, who was looking at her phone. Only the driver was prosecuted, and pleaded guilty to endangerment; prosecutors found no basis for criminal charges against Uber.
By control, the company had most power over the design: it decided how the system worked and how it was tested, and the NTSB named its inadequate safety culture as a contributing cause. By knowledge, a company running such tests should know that people watching an automated system for hours lose attention (the NTSB found it had no real way to deal with this); the driver knew she should watch the road. So a fair answer holds both responsible, the company for the system and the driver for her task, and asks why the person with least control carried the punishment (the moral crumple zone).

It plans an AI Regulation Council (chaired by the communications minister) to develop standards and rules, a National AI Centre, standards for data, algorithms, transparency and accountability, ways to deal with harmful AI-made content such as deepfakes, data protection laws, AI education from school to university, training AI specialists, data centers and support for startups, and a review every two years.
A policy is the government’s plan: it binds nobody and punishes nobody, so a person harmed by AI must rely on existing laws such as the Privacy Act, 2075 and the Electronic Transactions Act, 2063.
It becomes effective when it is put into practice: laws and directives with clear duties; a council and center with a budget and independent expert staff; ways for people to complain and get decisions reviewed; and measures of progress, checked at each review.

A public register lists every algorithm a public body uses, with its purpose, the data it uses, how officials act on its results and whom to contact (as Amsterdam and Helsinki began doing in 2020). An impact assessment is a set of questions a body answers before using a system, about what it decides, about whom, how bad a mistake would be, its data and testing; the answers set its level of risk, and the level sets the safeguards (as Canada’s Algorithmic Impact Assessment does).
Together: the assessment makes the body think about harm before use and puts strong safeguards (human review, notice, appeal) only where the risk is high; the register lets citizens, journalists and researchers see where algorithms are used and question them. Both are cheap compared with a ban, and they make it possible to find and fix unfair systems.

Criticism. The problem is real (copying, wrong answers, privacy), but a ban is the heaviest tool, used here for every use, including good ones: a teacher making practice questions, a blind student using voice tools, a class learning how AI works, as the National AI Policy itself plans. It is not proportionate; it is hard to enforce, since most students have phones; and it isn’t clear what counts as “AI” (a spell-checker? a translation app?).
A better proposal. Each school writes a policy on AI for schoolwork (as in Grade 9’s project): students may use AI where the teacher allows it for a task, and must say how they used it; exams and graded work done in class stay without AI; teachers are trained to design tasks that need students’ own thinking; no student’s personal data is entered into AI tools; the policy is reviewed each year with what teachers and students report.

A professional code of ethics is a statement by a profession’s association of how its members should behave towards the public, clients, employers and each other. It differs from a law: it is written by the profession, not Parliament; breaking it is not a crime, but can cost membership (or, in some professions, a license); and it gives principles to weigh, not rules for every case. It covers choices that are legal but still right or wrong.
Shared principles (any three): put the public’s safety, health and welfare first; be honest about what a system can and can’t do; be fair and don’t discriminate; work only within your competence and keep it up to date; report dangers to the public.

The software in the engine’s computer recognized a laboratory emissions test from the steering wheel’s position, the speed, how long the engine had run and the air pressure; during a test it ran the emissions controls fully, and on the road it turned them down, so the cars emitted up to 40 times the legal limit of nitrogen oxides. Researchers at West Virginia University found the difference by testing the cars on real roads in 2013 (published in 2014), and the US Environmental Protection Agency announced the violation on 18 September 2015.
It broke 1.2 (avoid harm: pollution harming people’s health), 1.3 (be honest: it was made to deceive regulators and buyers), 2.3 (respect the rules: it was built to break clean-air law while appearing to follow it) and the public good of 1.1 and 3.1.
An engineer could refuse, in writing, giving the reasons; raise the concern with managers and those above them; and if the company went ahead, report it to the regulator (blow the whistle), keeping records. The ACM Code’s 1.2 supports each step.

One possible design. Problem: landslides in the monsoon block roads and bury houses; people often have no warning. Technology: sensors on known unstable slopes (soil moisture, tilt, rainfall), sending readings over the mobile network to a server that sends a text-message warning to people below when readings pass a threshold; simpler options (a rain gauge and a phone tree) are compared, and the sensors are chosen because they warn at night and in places nobody watches.
Stakeholders, benefits and harms:

Design choice: the threshold: low gives early warnings but many false alarms; high gives few false alarms but less time. Chosen: two levels, “be ready” and “leave now”. Small test: sensors on two slopes for one monsoon, measuring how many warnings came before movement, how many were false, and whether people acted on them.

Answers will differ. A good answer names a thread honestly (for example, Data and analysis), links it to real careers (a data analyst at a bank or a health program in Nepal; remote data analysis for a company abroad), and gives specific next steps with sources: for example, SQL beyond Grade 10 (SQLite’s documentation and a practice site), statistics for data (a free university course), and a portfolio analysis of open data from Nepal’s statistics office, with a written report. It says how it would check its progress: a project finished, a course completed, a piece of work a client paid for.

The official documentation is written by the makers, for the version you have, and says exactly what a function does, including the surprising cases. Tutorials cover common uses and can be out of date; AI answers can be confidently wrong (Grade 10). Example: round(2.5) gives 2 and round(2.675, 2) gives 2.67 in Python; a tutorial might not mention it, but the documentation explains that exact halves round to the even choice and that 2.675 can’t be stored exactly in binary. Release notes also say what changed in a new version, so you know what may break.