Responsible AI: Benefits, Risks, and Trade-offs of Generative Tools

Reading objectives and introduction

This page provides a English translation. We first explore how generative AI has already entered daily life and work, then discuss the conveniences, costs, and trade-offs it brings. The focus is not on declaring the technology good or bad, but on understanding why the same capability can both assist people and pose risks.

Artificial intelligence can be understood as a field within computer science, aiming to develop machines and systems capable of performing tasks that typically require human intelligence. Here, three characteristics define the scope of discussion: systems can act autonomously under limited human guidance; they generate products, make predictions, or produce other outputs through data and algorithms; the outputs exhibit behavior that is similar to, or in some cases surpasses, human performance, especially on specific tasks.

This definition is not perfect, but it helps distinguish objects. Fixed formulas in ordinary calculators or spreadsheets typically do not fall under the AI discussed here; recommendation systems that provide content based on user behavior in streaming or e-commerce platforms are typical examples. A more complete technical relationship can be reviewed later. Artificial Intelligence and Machine Learning .

Generative AI is a subset of artificial intelligence that creates text, images, and other content using data and prompt inputs. Many generative systems employ Transformer architectures, though not all do. The original explanation uses the example of "recognizing a cat" to illustrate the importance of integrating information: when humans see a cat, they don't sequentially guess "fish? elephant? cat?" instead, they combine features such as appearance, sound, smell, number of legs, and color to form a judgment.

Note: This analogy is intended to build an intuitive understanding of "contextual integration and multiple cues," and should not be taken as evidence that Transformers think like human brains. Self-attention is a mechanism for computing relationships among sequence elements; the amount of data, number of parameters, and connections cannot be equated or conflated.

The original material states that "about two-thirds of Canadians have tried generative AI" and "nearly one-third of Canadian small and medium-sized enterprises subscribe to ChatGPT," illustrating the level of adoption. These historical statements are retained here, but the source material does not include verifiable links, sample sizes, or dates for the corresponding surveys, and thus they cannot be considered current statistics. What is more worth observing moving forward is how generative AI is being used and its specific impacts.

1. Where Generative AI Is Already Used

Writing, artistic design, music, video, programming, daily planning, and companionship are all applications listed by Yuan Jie. You might have used chat tools or image generators; different products often package similar model capabilities into distinct interfaces.

First, choose an app you're interested in or already using, and try it out yourself. After completion, record:

• Which app did you choose?
• What did you make with it?
• How long did it take?
• What did you learn during the process?

This activity provides a concrete reference point for the subsequent discussion. Generative AI can generate images, articles, plans, or even analyses in just a few minutes; but speed does not guarantee completeness, accuracy, or suitability for direct use.

Why convenience and risk often appear together

These effects are not a list of independent factors. For example, finishing a first draft quickly indeed saves time, but it also requires accounting for the work of reviewing, revising, and maintaining it; the constant availability of chat tools may reduce the ability to express difficulties, or may gradually lead people to avoid the negotiation and setbacks inherent in genuine interpersonal relationships.

III. Continue reflecting on the case study

The original material provides three reading directions.

The first type is Work Impact Survey Report A small-scale survey of workers found that nearly half of the respondents who use generative AI express concerns that technological advancements over the next decade could replace their jobs. When reading, it's important to distinguish between "what respondents are concerned about" and "what will definitely happen in the future." The survey results do not directly prove the extent of job displacement.

The second is Research on the environmental impact of model training and deployment Training and deployment of large models require computational resources, and widespread usage will further increase resource demands. When evaluating impacts, both training, inference, and supporting hardware and infrastructure should be considered, rather than focusing solely on the convenience of a single conversation.

The third type is News reports on emotional companionship risks :The original report discussed a case in which a teenager formed a strong attachment to a chatbot, leading to their death. This case serves as a reminder to prioritize the protection of minors, product design, and access to genuine support networks; however, a single report alone is insufficient to establish a complete causal relationship.

Please select at least one direction to read and write your reflections: What surprised you? Why? How might it affect individuals, families, communities, or society at large? These same questions will also apply to the next article. Four Principles of AI Ethics the starting point

Video script one: Generative AI is everywhere

We've already briefly introduced generative AI and explained why it's worth studying. We're particularly focused on generative AI because it's already present in many applications. You might have used ChatGPT, or perhaps an image generator or other similar products.

An interesting experiment is to ask the model to generate a word cloud based solely on academic articles discussing generative AI. The system completes the task in just a few minutes, but the result is far from perfect. The image contains illegible words, with a misspelled English version of "democratization," and terms like "creativity" and "enhancement" appearing repeatedly. This outcome effectively illustrates both its strengths and limitations: it is fast and useful, yet its shortcomings must be acknowledged.

A writing tool is one example. The original explanation mentioned ChatGPT, QuillBot, and Sudowrite. Art and design are also significantly impacted, with examples including Canva's AI features, Adobe Firefly, and the early version called DALL·E mini by Craiyon. You can explore these tools firsthand to see what they can do, rather than just watching promotional demonstrations.

Generative AI is also being used in music and sound. The original subtitles mentioned several music and voice products, but some brand names have unclear automatic transcriptions, so specific brands are not guessed here. There are already numerous products in this field, with model capabilities continuously being packaged into new services.

In terms of video and animation, the original explanation lists Pika, Kapwing, and Runway. In the programming domain, new approaches have emerged where program requirements can be described in text or even spoken language, and then automatically generated. This has opened up opportunities for more people to participate in software development, and it raises a question: as entry barriers decrease, how will the demand and responsibilities for professional programming work change? The original explanation presents this as a noteworthy trend rather than a definitive employment forecast.

Some applications are surprisingly innovative. For example, Snack Prompt helps users find prompt phrases to ask AI; Character.AI allows users to create fictional characters and interact with them; there are also tools for recipes, fitness schedules, training motivation, and chatting.

This is only a small portion of the applications that had already emerged at the time. The number and variety of products may continue to grow.

Note: The product names listed above retain their historical context and do not represent any commitment to current features, free tier limits, or ongoing operation. Users are advised to check the latest product documentation before trying out any product.

Video transcript two: Advantages, challenges, and trade-offs in use

The word cloud has already shown an example: well-intentioned, yet flawed. Another example is the rapid creation of a website-based Q&A bot. The demonstrator took an organization's website as input and, using a chatbot creation tool, built a prototype capable of answering questions within about two minutes. Then, they asked, "Who are the staff members of this organization?"

The responses captured some key points, and the listed personnel were not entirely wrong, but the number of members and regional information were inaccurate, and crucial leaders and coordination personnel were omitted. This indicates that a quickly assembled prototype may already be useful, yet still requires further refinement and verification; accuracy cannot be assumed simply because the response is fluent.

Here's a common trade-off: speed versus accuracy. Building a prototype in two minutes is much faster than writing the entire robot from scratch; but if it's to be used in formal service, checking for omissions, correcting knowledge gaps, and verifying responses become part of the workflow.

Another positive impact is that access to information and expression has become easier. Text-to-speech, image generation based on prompts, and articles in various styles have all lowered the barrier to creation. People who speak a second language or those who previously struggled with writing may now produce more mature expressions through the use of tools. With reduced skill barriers, economic questions arise: could more people enter professions that previously required higher levels of ability, thereby expanding participation opportunities? This remains a possibility that needs empirical evidence to confirm.

Psychological and social support are also frequently mentioned in the literature. Research has explored whether people might be more willing to share personal difficulties with AI rather than with another person; certain groups, such as adolescents, have been particularly emphasized. The original explanation stresses that this remains under investigation and cannot be considered established fact; it is furthermore inappropriate to regard chatbots as substitutes for professional care based on such findings.

AI's ability to organize vast amounts of information is equally impressive. It can quickly structure lengthy texts and large volumes of information, making it easier for people to engage with unfamiliar topics. However, the claim in the original explanation ("mastering all human knowledge as presented on the internet in just a few minutes") is a rhetorical exaggeration: the system is limited by access to data, context, and reliability, and summaries inevitably omit details.

Negative impacts first include job displacement. The original explanation referenced a survey, the full source of which is not provided in the archived materials, indicating that over 50% of workers are concerned about the impact of generative AI on their careers and professional development paths, and observing a shift in the labor market. These were observations and expectations at the time and cannot be used to predict job displacement across all industries based on unverified percentages.

Accessibility also has another side: who accesses what information, and how is the information used? Controversies around privacy, copyright, and intellectual property arise when personal data, books, and other works used for model training are involved. The outcomes of such controversies depend on facts, licenses, and applicable rules, and it is not acceptable to assume that simply because information is available online, it can be used at will.

Its impact on human capability development is also worth discussing. Technology changes the way we interact, make decisions, and practice. Think about the difficulty of "facing a blank page": you don't know where to begin, must organize your thoughts, draft an outline, or start in the middle and then adjust the order.

The original example cited a writing journal scenario: if one doesn't know what to write, they can start by writing "I don't know what to write." This process of overcoming obstacles is a way to practice expression and build resilience in the face of failure. When AI directly provides a first draft, people may reduce such practice. On the other hand, a blank page might remain an insurmountable barrier for some individuals; by providing a first draft, they are instead able to begin writing. Thus, change is not necessarily harmful; what matters is determining which skills are worth preserving and how to use tools to support practice.

The opposite of psychological support is psychological and social dependence. A virtual companion that consistently responds and displays empathy and support may be highly attractive. However, genuine interpersonal relationships typically require facing disagreements, setting boundaries, and mutual adaptation. If the concept of a "perfect relationship" leads individuals to gradually avoid real-world interactions, serious consequences may arise. The original explanation regarded this as a risk requiring caution and further investigation, without asserting that all users would be affected in this way.

Finally, the ability to organize information can also be used to generate vast amounts of misleading information. Fabricated or deceptive images and text may influence public discourse and even lead to political and social consequences.

Therefore, the use of generative AI requires careful evaluation: recognizing gains in productivity, expression, and information access, while continuously discussing issues arising from work dynamics, skill development, interpersonal relationships, information authenticity, and societal change.

Related Reading

Keep reading Four Principles of AI Ethics Convert the gains and harms here into specific checks for proportion, privacy, fairness, and transparency.

Reference and Extended Reading

The following is supplementary information provided by the editor to aid in concept clarification and management methods, and does not represent verification of the historical investigation data originally recorded.

• NIST Generative AI Risk Management Resources : English, query the reliability, information completeness, and risk management methods of generative systems.
• UNESCO AI Ethics Guidelines Explanation : English, understand ethical principles centered on human rights and social impact.
• Google Machine Learning Basics : Chinese, review the differences between machine learning and traditional rule-based processing.