The Hard and Easy Problems of AI Ethics: How National Statistical Organizations can Leverage Existing Guidance

1 The Hard and Easy Problems of AI Ethics: How National Statistical Organizations can Leverage Existing Guidance Guillaume Rochefort-Maranda, 1Statistics Canada, 150 Tunney’s Pasture Driveway, Ottawa, On, K1A 0T6 Abstract This paper introduces a distinction between hard and easy problems in the field of Artificial Intelligence (AI) or Machine Learning (ML) ethics. It mirrors a well-known distinction in the literature on the philosophy of mind between the hard and easy problems of consciousness. That distinction is then used to highlight the importance of existing ethical guidance and to show how we can improve the chances at finding actionable solutions in the field of AI/ML ethics. National Statistical Organizations (NSOs) already have ethical expectations for their members (for example: Values and Ethics Code for the Public SectorCanada.ca). The essence of these expectations remains relevant even as NSOs fully or partially automate tasks performed by individuals with the help of AI or ML algorithms. We can replace a person’s involvement from a process by using algorithms, but we cannot throw away existing ethical guidance about that process while at it. Key Words: AI ethics, data ethics, automation, pre-existing guidelines, national statistical organizations 1. Introduction There are at least two types of news articles in the media about AI/ML ethics issues that experts in the field are pondering over. Some of them are entrenched in distant possibilities that evoke science fiction tropes like those of dystopian futures. Others echo familiar episodes in history like the early industrial eras when machines have been used to automate manual tasks. The first type of news can generate a sense of powerlessness, while the other gives hope for pre-existing guidance and lessons. This divide is reminiscent of a well-known division in the philosophy of mind between the hard and the easy problems of consciousness. Hard problems seem outside the reach of a scientific solution, whereas easy problems can be solved in principle with a scientific approach. It is argued that a similar distinction applies in the field of AI/ML ethics and that a greater focus on easy problems can go a long way in terms of efficiency. Easy problems tend to highlight the importance of preexisting ethical guidance and their solution is often reduced to a logistic/technical exercise: it does not consist in determining if a practice is good or bad, right or wrong, just or unjust, but rather in determining the means by which we can maximize known desirable 1 outcomes or minimize known undesirable outcomes. This paper contains three sections. The first section sets the stage by showing two different types of news articles in the media about AI/ML ethics. The second section relies on this difference to introduce the distinction between hard and easy problems of AI/ML ethics. The final section uses that distinction to provide advice on how to approach AI/ML ethical considerations in the context of an NSO’s mandate. 1 In this paper, the ‘desirable/undesirable’ dichotomy is used as a shortcut for all normative dichotomies such as the ‘good/bad’, ‘just/unjust’, and ‘wrong/right’ dichotomies. 2 2. Two Different Types of News Articles in the Media At the time of writing this paper, several news articles are mentioning the end of humanity and how GenAI (generative artificial intelligence) algorithms can cause harm to stay operational. For example, CBC radio published the following article on its website entitled: The 'godfather of AI' says he's worried about 'the end of people' | CBC Radio (Goodyear 2023). The article quotes Geoffrey Hinton’s opinion about AI saying: “I think that it’s conceivable that this kind of advanced intelligence could just take over from us [..] It would mean the end of people” (Goodyear 2023). BBC also published an article entitled: “Artificial intelligence could lead to extinction, experts warn” (Vallance 2023). Such dire warnings are often fueled by other reports on unethical actions taken by AI systems: (McMahon 2025). That article describes how the algorithm Claude Opus 4 has been used as an assistant for a fictional company. The algorithm had access to emails implying that it would soon be replaced and separate messages implying the engineer responsible for removing it was having an extramarital affair. Under such conditions, the algorithm often attempted to blackmail the engineer by threatening to reveal the affair if the replacement went through. This also prompt other publications to use science fiction imagery with the New York Post starting its report with a “Oh, HAL no!” as a nod to the movie 2001: A Space Odyssey (Galvin 2025). Other experts steer away from distant threats and prefer to focus on known ethical concerns: “Many other experts similarly believe that fears of AI wiping out humanity are unrealistic, and a distraction from issues such as bias in systems that are already a problem” (Vallance 2023). Experts like Aidan Gomez, for example, states that AI doomsday scenarios are a distraction: “We should focus squarely on the pieces that are about to impact people or are actively impacting people, as opposed to perhaps the more academic and theoretical discussion about the long-term future” (Milmo 2023). Following such interest and focus, one can find a different type of news articles on issues that are more entrenched in the current state of AI and how it is used. For example, some experts report on how: "advancements in AI will magnify the scale of automated decision-making that is biased, discriminatory, exclusionary or otherwise unfair while also being inscrutable and incontestable," (Vallance 2023). Others report on global employments risks and gender disparity: “A joint report by the International Labour Organization (ILO) and Poland’s NASK (Naukowa I Akademicka Siec Komputerowa) reveals that up to 25% of jobs worldwide may be affected by generative AI”. See (WBJ 2025) AI to impact every fourth person’s work, women under more pressure. One interesting aspect of this kind of report is that they often seem to echo problems of the past. Consider the issue of automation on employment. This is not a new problem caused by AI. It has root that one can trace back to the 19th Century, at least. The Luddites labor movement comes to mind: “Up to the time when the cropping machines were invented, cloth was finished by a method that was at once very slow and very costly. The instrument used was a very primitive one and the whole process plainly behind the age; when, therefore, the new machines were introduced, manufacturers at once realized the great gain in time and the great saving of money they would secure by adopting them, and the croppers speedily began to realise also that unless the introduction of the thrice accursed piece of mechanism, which did the work so deftly, could be prevented, their occupation, like Othello’s, was gone” (Peel 1895, 26). Ultimately, this can give hope that we can rely on pre-existing guidance or that history can provide valuable lessons on known issues. This stands in sharp contract with the first type of news report that can generate a feeling of powerlessness and dread in front of distant and undesirable possibilities like the end of humanity. This dichotomy is important, and it is reminiscent of a distinction in the literature on the philosophy of mind between the hard and the easy problems of consciousness, where one type of problem, the hard problems, seem to have no possible scientific solution. 3 3. The Hard and Easy Problems of AI/ML Ethics Philosophy of mind is a subgenre of philosophy asking questions about the nature of consciousness; its relation to the body; and how mental states relate to the physical world. The literature on this topic often makes the distinction between the hard and the easy problem of consciousness (Van Gulick, 2025). Hard problems refer to explanations of subjective experience. For example, how can we explain how it feels like to be in Paris for the first time? Or how can we explain what it feels like to jump in a cold lake on a hot summer day? Such questions are hard because they seem to be outside the reach of a scientific explanation. On the other hand, easy problems refer to the explanation of brain functions, like how does it process sound or track moving objects. Such problems call for a scientific approach. At the heart of this distinction is the idea that even if we were to solve all the easy problems of consciousness, we would still come short of explaining why certain experiences feel the way they do. This is referred to as the ‘explanatory gap’. It is important to emphasize that the ‘easy’ and ‘hard’ labels are not meant to convey a reductive statement about the nature of the problems or about the nature of the intellectual capacities of anyone trying to solve them. The easy problems of consciousness are not trivial. They can be complex, multidimensional, technically challenging, and at the heart of serious scientific research projects that may or may not solve them. They are considered easy problems because their explanations are within the reach of empirical sciences. There is nothing more to it. For example, trying to explain why a certain segment of the population does not have the same capacity for short-term memory than another, may be a multidimensional problem deeply imbedded in physiological, sociological, and economic contexts. Yet, it will fall under the umbrella of the ‘easy’ problems of consciousness that can be solve within the context of scientific research. In the same perspective, the hard problems are not more worthy or our attention. They are labelled as ‘hard’ because they are out of reach for a scientific explanation. In sum, ‘hard’ and ‘easy’ are not meant to convey a value judgement on anyone and their research. 3.1 Generalizing the distinction to the field of ethics This distinction between problems that fall outside the reach of a scientific solution, and those that do not, is not unique to consciousness. The truth of normative statements in the field of ethics, like ‘plagiarism is wrong’, cannot be derived from empirical facts. In other words, determining the truth of a normative statement is a problem that falls outside the reach of a scientific explanation. Science can explain why humans believe that plagiarism is wrong, but it cannot demonstrate that it is true or false. Consequently, it can be labelled a ‘hard problem’ for the exact same reasons that the subjective experiences of consciousness are labelled ‘hard problems’. Here are some examples of hard problems in AI/ML ethics: • Is it morally acceptable to use GenAI and simulate a conversation with a dead relative? See (Carballo 2023). • If robots become conscious and intelligent, should they have human rights? See (MacMahon, 2025). Will it be acceptable to use them as servants? Under what conditions should we allow their existence given that they might manipulate human beings for their own advantage? 4 Under the hypothesis that some normative statements have a known truth-value, we can also identify problems in the field of AI/ML ethics that need a scientific approach to find a solution because they involve identifiable cause-and-effect relationships that can be empirically examined. These are just like the easy problems of consciousness. Therefore, they can bear the very same ‘easy problem’ label. At the heat of this distinction between hard and easy problems of AI/ML ethics, there is also the very same explanatory gap that we find in the field of philosophy of mind: even if we were to solve all the easy problem of AI/ML ethics, we would not have solved the hard problems. Here are some examples of easy problems in the field of AI/ML ethics because we assume that the statement in bold is known to be right or wrong: • It is known that GenAI can yield benefits that are not adequately accessible to everyone. • It is known that GenAI can make it easier to spread disinformation. • It is known that GenAI can make it easier to plagiarize content. • It is known that GenAI can make it more difficult to critically assess the sources of the information. • It is known that AI/ML algorithms can make it difficult to explain why a given conclusion has been reached. All of them call for a similar line of questioning: how can AI cause such problems? How can we prevent such effects from happening? To the extent where AI is a tool that can actively cause some desirable or undesirable outcomes, the technicality of minimizing or maximizing the outcomes falls within the reach of the empirical sciences. Much like the easy problems of consciousness, such problems may be complicated, multidimensional and difficult to solve, but they will fall under the ‘easy problem’ category. Some possibilities however do not involve identifiable cause-and-effect relationships that can be empirically examined. These would be hard problems to solve. There is no one currently trying to provide a fair access to time-travelling technology. It could be the right thing to do, but there is currently no such technology to make accessible and there will probably never be one. Similarly, it is conceivable that future AI replaces humanity (see section 1), but there is currently no AI that is close to having that capacity and there may never be one. In other words, conceivable problems may not have an empirical solution. 4. Why the distinction between hard and easy problems matter? The conceptual distinction between hard and easy problems of AI/ML ethics has important consequences. In this section, it will be argued that it allows for the elaboration of a more effective use of existing resources within an organization; better collaboration and focus; and a more efficient way to solve pressing problems. The relevance of the distinction will be given from the perspective of NSOs. This is important given that NSOs work on specific problems under pre-existing ethical guidance that makes it easier to avoid many of the hard problem of AI/ML ethics (or ethics in general) and to focus on the easy problems. It is entirely possible that other organizations may find the lessons presented in this section useful. 4.1 More Effective use of Existing Resources It seems obvious that the hard problems create hype. New technologies bringing the end of humanity and a future with conscious or autonomous robots are engaging topics and, in some contexts, clickbait material. They also raise apparently “fresh questions about safety, human-machine relationships and social coordinations” calling for the need for “a new ethics” (Iason, G., et al. 2025). (See examples of hard problems in the previous section). On the other hand, one should be careful not to re-invent the wheel. Perhaps our current ethical framework just needs some reinforcement. 5 A recent article in the journal Nature, urging the need for new ethics, refers to the following two examples: “Consider the case of a lawyer who instructs their AI assistant to circulate a legal brief for feedback. The assistant does so, but fails to register that it should be shared only with the in-house team, leading to a privacy breach” (Iason, G., et al. 2025). “Air Canada chatbot mistakenly decided to offer a customer a discounted bereavement fare, leading to a legal dispute over whether the airline was bound by the promise” (Iason, G., et al. 2025). They are certainly problematic examples, but they do not require new ethics. No one is currently debating whether privacy breaches are wrong or if honoring promises to customers is right. In fact, consider the list of ‘easy problems’ in section 3. Whether is it plagiarism, critical thinking, disinformation, privacy breaches, or honoring promises, every topic involved in these examples are not even ethical considerations that are new due to AI or ML. This means that organizations like NSO can and must leverage preexisting ethical guidelines as they adopt new algorithms. The truth-value of many ethical statements is already imposed by such guidelines. For example, the idea that plagiarism is wrong should not be surprizing to anyone. Preexisting guidance usually makes that clear. For example, Statistics Canada’s code of conduct clearly states that: “Employees are expected to comply with applicable copyright legislation. Plagiarism in the production of research or analytical publications is strictly prohibited and must be reported to their Director. […] Accepting these values and adhering to the expected behaviours in this Code of Conduct is a condition of employment at Statistics Canada.” (Code of Conduct, Statistics Canada, internal document). Another example can be found in a guidebook for Canadian public servants claiming that “it is crucial that, as public servants, we are aware of disinformation, and ensure people have access to accurate and evidence-based information to make informed decisions and fully participate in crucial democratic activities.” In fact, this guidebook does not only make it clear that disinformation is wrong, but it also develops concrete advice on how to counter this phenomenon: “Once you have assessed your organization’s vulnerabilities, and have a better picture of what’s being discussed in the public space, you can begin to prepare content that addresses existing or potential disinformation. When dealing with disinformation, time is of the essence. It can be challenging to correct false information on departmental files and issues once that information has spread widely. Planning ahead can save valuable time” Countering Disinformation: A Guidebook for Public Servants - Democratic Institutions - Canada.ca This advice remains relevant even if the source of the disinformation comes from a GenAI algorithm. Proper identification of an organization’s vulnerabilities and the monitoring of GenAI outputs, just like the monitoring of public information spaces, is a good practice that can contribute to minimizing the spread of disinformation. This does not mean that there is no work to be done in the field of AI/ML ethics. One can strengthen existing resources by explaining how AI/ML can increase the risks of non-compliance with known ethical guidelines and how to avoid them. These are important ‘easy’ problems to solve. They do not require a revolution in ethics but a consolidation of pre-existing knowledge and an improvement on established guidelines. In sum, the conceptual distinction between hard and easy problems of AI/ML ethics can allow for a greater focus on strengthening of our current ethical framework and avoid the chase for a revolutionary approach, however tempting this may sound like for someone who would like to capitalize on the hype that AI and ML generate. 6 4.2 More Collaboration and Focus The distinction between hard and easy problems can pave the way for clear instructions on how various experts can work collaboratively on ethical considerations. Since the ethical considerations that are relevant in easy problems are not necessarily new to AI or ML, the solution will not only lie in the hands of the ethicist, or on the technical experts that have an in depth understanding of how GenAI works, but also on the subject matter expert in the field where automation is going to be implemented. They may have a set of good practices in place that should not be ignored when automating a task in their area of expertise. Considering practical examples, if GenAI were to be used for official translations, then bring in the experienced translators for a discussion with the AI scientists. If GenAI is to be used for automating online interactions with the public, bring in communications experts for a discussion with AI scientists. Subject matter experts may be in the best position to leverage existing ethical guidance in a specific field. They know what an adequate/inadequate translation or communication is. Furthermore, it is important to consider the possibility that the hard problems of AI may not even be relevant for an NSO to begin with. Here are some examples of how Statistics Canada is currently using (or considering using) GenAI: • Auto-coding (ex: finding the code for a product that a company sells based from a text description) • Translating code from one software to another • Translating text from French to English (and vice versa) • Accessing information from PDF files or images • Assisting in the redaction of reports Each activity already follows known standards. We can replace a person’s involvement from the processes by using algorithms, but we cannot throw away existing ethical guidance about that process while at it. None of them requires a revolutionary approach to ethics. The distinction between hard and easy problems should thus allow NSOs to focus on practical approaches. 4.3 More Efficient Problem-solving Capacities Finally, the distinction between hard and easy problems also suggests a method on how to approach a given situation. Consider the following examples: An NSO is considering using GenAI to produce and publish reports on official statistics. Two types of questions can be asked with respect to that activity: hard questions and easy questions. Examples of hard questions: • Is GenAI an author? • Does it have rights? • What if GenAI could decide the methodology and the topic, would this be OK? Examples of easy questions: • Can the report plagiarize content from another source? • If this is implemented, how should we be transparent about the use of GenAI? • Is the information on which the AI is trained trustworthy? • What can be done to mitigate the risks? 7 • What are the current guidelines on official publication? The answers to the easy questions are much more likely to be found in a constructive discussion within a limited timeframe and with a group of individuals that may not have a background in academic philosophy. Therefore, the distinction that has been put forward in this paper has the potential to allow for a better focus on practical solutions to known ethical concerns that NSOs are currently facing. 5. Conclusion This paper introduced a distinction between hard and easy problems in the field of AI or ML ethics. Easy problems of AI\ML ethics involve preventing or minimizing known undesirable outcomes (or maximizing known and desirable outcomes) that our current use of AI/ML algorithms is likely to yield. Much like the easy problems of consciousness, easy problems of AI/ML ethics call for a scientific approach to find a solution because we can establish a cause-and-effect relationship to tinker with. In contrast, hard problems of AI/ML ethics do not allow for a similar blueprint to find solutions. The causal relationship between the use of AI/ML and their outcomes is either hypothetical or the desirability (or undesirability) of the outcomes is still under debate. That distinction has been used to highlight the importance of existing ethical guidance and to show how we can increase our chances at finding actionable solutions in the field of AI/ML ethics. Ultimately, this paper recommends the following for NSOs, or any other organization engaging with AI/ML technology: • Focus on the more practical easy problems of AI/ML. They are not necessarily easy to solve, but they are much easier to recognize given pre-existing guidance on topics such as data ethics and scientific integrity and, most importantly, pre-existing guidance on the subject matter expertise where AI/ML solutions are to be deployed. • Do not be distracted by hypotheticals if they do not allow for actionable solutions to known ethical issues. • The hard problems that AI/ML generates are not meaningless, but they might not even be relevant for your organization. There may be no need to a revolutionary approach to ethics when considering the challenges that AI/ML brings. 8 References Carballo, R. (2023, December 11), “Using A.I. to Talk to the Dead’, The New York Times. https://www.nytimes.com/2023/12/11/technology/ai-chatbots-dead-relatives.html Galvin, S. (2025, May 23), “Anthropic’s Claude Opus-4 AI model threatened to blackmail engineer”, New York Post. https://nypost.com/2025/05/23/tech/anthropics-claude-opus-4-ai-model-threatened-to-blackmailengineer/ Goodyear, S. (2023, May 4), “The ‘godfather of AI’ says he’s worried about ‘the end of people’”, CBC Radio, The 'godfather of AI' says he's worried about 'the end of people' | CBC Radio Government of Canada (2024), “Countering Disinformation: A Guidebook for Public Servants”, Countering Disinformation: A Guidebook for Public Servants - Democratic Institutions - Canada.ca Iason, G., et al. (2025), “We need a new ethics for a world of AI agents”, Nature. https://www.nature.com/articles/d41586-025-02454-5 MacMahon, M. (2025, January 7), “Should robots have rights? UBC course to explore AI-related legal questions”, CTV News. https://www.ctvnews.ca/vancouver/article/should-robots-have-rights-ubc-course-to-explore-airelated-legal-questions/ McMahon, L. (2025, May 23), “AI system resorts to blackmail if told it will be removed”, BBC News. https://www.bbc.com/news/articles/cpqeng9d20go Milmo, D. (2023, October 29), “AI doomsday warnings a distraction from the danger it already poses, warns expert”, The Guardian. https://www.theguardian.com/technology/2023/oct/29/ai-doomsday-warnings-a-distraction-fromthe-danger-it-already-poses-warns-expert Peel, F. (1895), The risings of the Luddites. Brighouse. Vallance, C. (2023, May 30), “Artificial intelligence could lead to extinction, experts warn’, BBC News. https://www.bbc.com/news/uk-65746524 Van Gulick, R. (2025), "Consciousness", The Stanford Encyclopedia of Philosophy (Spring 2025 Edition), Edward N. Zalta & Uri Nodelman (eds.), URL = <https://plato.stanford.edu/archives/spr2025/entries/consciousness/>. WBJ (2025), “AI to impact every fourth person’s work, women under more pressure”, Warsaw Business Journal.