Reading objectives and introduction
This is an English translation of the published reading. We understand AI challenges and ethical dilemmas at different levels and learn how to read impact studies. Categorization is meant to aid thinking; in reality, the effects often overlap: changes in individual behavior may further influence interpersonal relationships, labor markets, and the environment.
I. Personal Impact
AI may influence work, shopping, physical and mental states, as well as knowledge, skills, and ways of thinking. How recommendation algorithms curate the content we see offers a promising entry point for study: they may make information more accessible, or they may continuously reinforce existing preferences.
The original material suggests that AI may also be used in psychological, medical, visual therapy, and support for human development, and invites reading about research on psychological support tools and performance in specific medical diagnosis tasks. When reading these materials, it is important to distinguish between specific experimental tasks, test populations, baseline comparisons, and real-world usage conditions.
Note: The original source makes a broad statement that "human brains have already been altered by AI," which cannot be used to infer that every application either harms or improves cognition. The fact that "a certain test outperforms humans" does not imply it can independently assume clinical responsibilities. This discussion focuses on research and evaluation implications, not clinical recommendations. The safety, effectiveness, and accountability of medical systems still require individual validation.
II. Impact on Interpersonal and Community Levels
Before the emergence of generative large models, there were already robots and chat products designed for companionship and interpersonal relationships. The source material mentions robots aimed at children's language and learning support, as well as Replika, launched in 2017 with the goal of adult companionship.
Generative capabilities are enabling broader applications of interactive services such as tutoring and psychological support, which were previously primarily delivered by humans. The long-term impact of relying on such interactions on relationships with real people remains to be studied continuously. The fact that systems are becoming more human-like does not automatically imply that such interactions are more beneficial.
The original material also offers a unique perspective: when companion products are updated, users may feel a loss of their familiar AI "friend," and wonder what kind of reactions this might trigger. This question helps illustrate how human attachment can also be influenced by product rules, commercial decisions, and system changes.
III. Social Structure and Labor Market Impacts
AI has also entered fields such as intelligence, military, and political propaganda, for example, deepfakes and targeted political messaging. For many people, more immediate impacts come from economic and employment changes.
Productivity and profits may increase, but changes in business processes also affect job responsibilities, skill requirements, and workers' conditions. Therefore, we cannot look at organizational efficiency alone, we must also ask: which tasks are being reduced, which are being added, who is gaining new opportunities, and who is bearing the costs of adjustment?
The original reading activity offers two directions: first, by studying job postings on gig platforms, examining how demand has changed following the emergence of generative AI; second, by reading a specified chapter of a study to understand the potential impacts on enterprises and processes. The saved text does not retain a precise, verifiable full-text link and cannot be substituted with a similar study.
IV. Environmental Impact
Larger models, longer inputs and outputs, and the creation of synthetic training data may all increase resource demands. Beyond electricity, considerations must also be given to data center cooling, water usage, equipment manufacturing, and disposal.
The original material uses a historical report from February 2025 of “about 400 million users per week.” It attempts to scale up an earlier estimate of “1 million users and 8.4 tonnes of emissions per year,” arriving at a claim of 33,600 tonnes, equivalent to the annual emissions of approximately 2,240 Canadians.
Note: This external extrapolation cannot be used as a reliable emission estimate. Simply multiplying the given numbers, 8.4 × 400, should yield 3,360, not 33,600; more fundamentally, weekly active users cannot directly represent annual request volume, as models, request lengths, hardware, power consumption, and accounting scopes vary across years. The original estimate's source and boundaries are not fully preserved, so further calculations of per capita equivalents from this result are not pursued. Actual workloads and clearly defined lifecycle boundaries should be used instead.
Text generation is not the task that consumes the highest amount of resources. The following video discusses differences among various tasks and reminds us to observe direct and indirect effects.
The original reading activity also invites a comparison of two materials: whether the environmental burden of AI environments is evenly distributed across different regions, and what environmental costs are associated with current data collection and storage methods.
Video script one: How do the four levels interconnect?
For easier thinking, we can divide influences into four categories: individual, interpersonal, social structure, and environment. A single influence may span multiple categories, so these boundaries are not absolute.
At the individual level, it involves brain, body, and psychological states. Some studies explore the assistance AI can provide in diagnosis or psychological support, while also considering dependence, overdependence, and changes in thinking patterns. The specific benefits depend on the system and validation conditions.
The interpersonal level involves interactions, roles, and communities. Social media recommendation algorithms can facilitate the formation of communities among people who are geographically distant but share similar interests; similar mechanisms may also interact with issues such as isolation and psychological distress.
At the level of social structure, geopolitics and labor are primarily involved. Military drones, intelligence, deepfakes, and targeted propaganda are all uses previously listed. Some studies discuss their potential impact on political processes; concerns regarding productivity, profits, and economic growth are concurrently present, linked to worries about changes in the labor market.
The environmental aspect focuses on water, electricity, and emissions, as well as the direct costs of production, transportation, operation, and disposal of equipment. It also includes indirect effects: for example, improved building maintenance or agricultural productivity may reduce resource waste; however, whether there is a net benefit must be evaluated in conjunction with the costs associated with adopting the system itself.
Therefore, each level may have both positive and negative effects; the dilemma's emergence should be observed through specific cases.
Video transcript two: From a Single Request to Environmental Cost
The environmental cost of generative AI often lies far from users' direct experience. When sitting in front of a computer typing a sentence or generating an image, people rarely see the energy and infrastructure required for model training; they might also assume that "training is complete" and thus overlook the ongoing costs of continued usage.
The original explanation cites studies comparing different tasks, ranging from dedicated classification systems to generative models and image generation, illustrating significant differences among these tasks. The chart uses logarithmic coordinates, so differences spanning multiple orders of magnitude should not be interpreted using ordinary linear scales.
The example provided for the original explanation states that each 1,000 text classification requests corresponds to approximately 0.3 grams of carbon dioxide, while text generation consumes about 10 grams; two requests to a high-energy image model consume roughly as much energy as charging a smartphone once.
Note: These numbers are explanations from the original commentary for specific studies and charts, not unified coefficients applicable to all current services. Different studies may report energy or carbon emissions, which require conversion based on conditions such as power carbon intensity; average values, peak values, task durations, and hardware specifications cannot be interchanged. "Phone charging" represents only a conditional comparison and does not apply to arbitrary two image generation tasks.
The daily request volume is very high, and even minor resource consumption per request can accumulate. Direct impacts include production, transportation, operation, and disposal; indirect impacts may be either positive or negative.
Automated irrigation in agriculture is a positive example: if water is supplied based on actual conditions, it may reduce waste compared to watering for a fixed number of hours each day, and can also be compared with already water-saving drip irrigation and scheduled watering programs.
Negative indirect effects include increased consumption and overuse. If an additional layer of call is introduced to provide a prompt for another AI, one should question what extra benefits it brings. A recipe app may generate value by reducing food waste; if not, it cannot simply claim environmental benefits merely because it uses AI. Entertainment applications should also be evaluated based on actual needs and resources, rather than applying the same conclusion to all uses.
One approach to reducing environmental impact is data management: optimizing and limiting unnecessary collection and storage, and discontinuing the retention of data that is no longer needed. The original explanation mentioned that some companies may have a high proportion of "dark data," citing a figure as high as 80%, but the supporting materials do not provide verifiable statistics, so this cannot be generalized to all enterprises.
"Dark data" does not merely refer to data used only once; it more commonly refers to data that has been collected or stored but not effectively utilized or analyzed. Decisions regarding deletion or retention should also take into account actual business needs and legal requirements.
The original explanation continued to mention the concept of aggregated storage or resource sharing, and advocated for careful evaluation of the cost involved in generating synthetic data. Synthetic data can address issues such as sample scarcity in appropriate scenarios, but the generation process itself consumes resources and should not be assumed to be better the more abundant.
Open, shared, and appropriately licensed datasets have the potential to reduce the need for multiple companies to independently scrape, store, and process the same data. Challenges related to intellectual property, commercial interests, and governance still remain. The call for collaboration, and potentially policy coordination, does not imply that publicly available data can be used without authorization.
Finally, it's important to make thoughtful model selections. If a dedicated system already meets the task requirements, there's no need to default to larger, more general-purpose models. Even if generation capability is needed, comparisons can be made across architecture, parameter size, and output length. Using a smaller, task-appropriate implementation (avoiding unnecessary long outputs) is a worthwhile optimization direction; ultimately, it's still necessary to verify that the same quality standards are met.
Next practice step
use AI Ethics Impact Study Practice Select a real case to illustrate the dilemma, causes, evidence, and feasible improvements, rather than merely stating "there are benefits and drawbacks."
Related Reading
Can revisit when assessing damage Four ethical principles ; want to study a specific case, you can continue doing Ethical Impact Assessment Exercise .
Reference and Extended Reading
The following is the edited supplementation and research inquiry entry; references in the original material lacking specific links were not marked as completed in the original reference.
- Power Hungry Processing Research : English, query a comparison of energy consumption across different models during the inference phase and experimental boundaries.
- IEA: Energy and AI : English, systematic analysis of data centers, energy, and AI, differentiating the entire data center industry from AI-specific workloads.
- WHO: Ethics and Governance of Large Multimodal Models : English, understanding the verification, governance, and accountability required for medical applications cannot be substituted by experimental performance alone.