ChatGPT is a Remarkable Tool -- For Experts

• ChatGPT is described as useful across research, education, creative writing, programming, and medicine, including support for text tasks and personalized learning.The introduction also lists common uses such as summarizing, drafting, code assistance, and solving problems.Source: §1 · pp. 1–2
• The paper focuses on professional ChatGPT use, where accuracy, privacy, copyright, transparency, and domain-specific ethical concerns require attention.The introduction highlights risks in high-impact patient interactions, creative work, programming, and education.Source: §1 · p. 2
• The paper frames responsible AI use across sectors as requiring attention to privacy, algorithmic bias, and transparency.Source: §1 · p. 2
• The paper distinguishes its practical tools for individual users from prior work focused on ethical challenges at regulatory and institutional levels.It positions heuristics and flowcharts as guidance for maximizing benefits while handling practical limitations.Source: §1 · pp. 2–3

2.1 Research and Academic Usage

• Researchers can use ChatGPT from literature review through data collection, analysis, drafting, and editing, while studies also explore ranking [38], stance classification [82], and machine translation.ChatGPT rankings showed some consistency with human rankings; it performed well on common stance-classification datasets and could explain predictions. Translation performance was competitive for high-resource European languages, and GPT-4 improved performance for low-resource or distant languages.Source: §2.1 · pp. 4–5
• In research, ChatGPT can make significant, plausible-sounding errors that mislead non-experts, including fabricated quotations and sources, and it may mishandle context-specific proofreading.Version 3.5 was observed to provide inaccurate quotes and attribute them to nonexistent sources.Source: §2.1 · p. 4

2.2 Opportunities and Challenges in Education

• ChatGPT can provide instructors with lesson-planning support and students with personalized information and assistance, but unreliable responses, academic-integrity risks, and concerns about educators’ changing roles qualify its use.Its responses can be incorrect or illogical, and small changes in phrasing may affect answers; risks also include cheating, bias, fake information, and harm to creative writing and other higher-order skills.Source: §2.2 · pp. 5, 7, 8
• Manual algebra hints produced significantly higher learning gains than ChatGPT hints, and 30 % of ChatGPT hints were rejected for wrong answers or solution steps, underscoring the need for human supervision.The comparison covered Elementary and Intermediate Algebra; learning gains with ChatGPT hints were not statistically significant.Source: §2.2 · p. 7

2.3 Programming Assistance

• Programming assistance spans learning, information gathering, code explanation, and code generation, but effective use relies on appropriate prompting, software-development expertise, and rigorous testing.A recommended “Knowledge Generation” approach uses detailed prompts describing architecture, design guidelines, and step-by-step plans, followed by code inspection and testing.Source: §2.3 · pp. 8–9
• ChatGPT can give incorrect code explanations, and programmers should retain control rather than outsource development of critical software systems.Explanations may focus on misleading variable names or fail to capture the code’s intended meaning.Source: §2.3 · p. 9
• ChatGPT can help debug code, explain errors, suggest corrections, and improve readability, but developers should verify its outputs and remain responsible for project-specific requirements.Providing programming language and environment details can improve bug hunting; ChatGPT may miss conventions and performance, security, privacy, or accessibility needs.Source: §2.3 · pp. 10–11
• 77.5% success rate: ChatGPT outperformed traditional program-repair methods and performed competitively with deep-learning approaches in a bug-fixing evaluation.Human input helped improve the automated system, but verifying ChatGPT’s answers can require significant effort.Source: §2.3 · p. 10

2.4 Mathematics Tasks

• ChatGPT can solve basic mathematical problems and explain concepts, but it often gives incorrect answers confidently and struggles with complex operations, algebraic representations, and deeper logical reasoning.It may produce random digits for difficult calculations, and a study found its mathematical abilities below those of an average mathematics graduate student [27].Source: §2.4 · p. 12

2.5 Healthcare

• ChatGPT should not replace professional medical advice because its decision-making, potential biases, inaccurate or outdated information, and limited reliability raise concerns in this high-risk domain.The passages also identify ethical and legal concerns, including authorship, authenticity, copyright, and the lack of personal and emotional insights.Source: §2.5 · pp. 13–14
• ChatGPT has exhibited considerable healthcare capabilities, assisting professionals by deciphering intricate medical information, and has been accurate in addressing medical queries across numerous specialties.Source: §2.5 · p. 14

• ChatGPT may give incorrect responses, inaccurate code, outdated information, and biased or inappropriate outputs, with limited reasoning and self-correction and possible overconfidence.Source: §3 · p. 14
• Its in-depth explanations, examples, and references can make occasional inaccuracies and fundamental errors difficult to navigate.Source: §3 · p. 14
• The paper recommends using ChatGPT cautiously, independently verifying its outputs with reliable tools, and relying on users with domain expertise.Source: §3 · p. 14

3.1 Incorrect Responses - Confidently

• ChatGPT can present inaccurate information with unwarranted certainty, so users should approach its responses cautiously and verify them independently.Confident errors may confuse students and non-experts, especially in medical or scientific contexts.Source: §3.1 · p. 14

3.1.2 Inaccurate Code

• On ChatGPT3.5’s 30 prompt variants, 93.3% of functions and 96.6% of function usages were correct, versus 43.3% of running examples and 60% of explanations.The paper concludes that generated code should be examined and tested before use.Source: §3.1.2 · p. 15

3.1.3 Information Is Not Up-to-Date

• ChatGPT cannot assimilate information after its training cutoff; real-time data supplied as query context does not become retrievable in later queries.Source: §3.1.3 · p. 15

3.1.4 Limited Logical Reasoning

• Logical failures span formal entailment and finance, law, and medicine; a prompt encouraging reasoning also led ChatGPT4 to turn a correct answer into an incorrect one.In the medical example, ChatGPT4 gave the correct answer but an incorrect explanation.Source: §3.1.4 · pp. 15, 18

3.1.5 Incapacity of Self-Correction

• ChatGPT may fail to acknowledge contradictions within a response and instead try to hide them; it can acknowledge mistakes in a later response.ChatGPT4 also appears to notice a reasoning problem but cannot correct it while responding.Source: §3.1.4 · p. 18, §3.1.5 · p. 21

3.1.6 Over Self-Confidence

• Examples in Figures 2–9 show ChatGPT and ChatGPT4 often sounding highly confident even when incorrect, making critical evaluation important.Source: §3.1.6 · p. 21

3.2 Privacy Issues

• Data privacy concerns the secure storage, access, retention, and protection of sensitive data, supporting confidentiality, availability, and integrity.The section also discusses regulatory frameworks and prior recommendations including user control, data minimization, security, and auditing.Source: §3.2 · pp. 21–23
• A March 2023 breach exposed some users’ conversation headings to other users, illustrating privacy risks from collecting ChatGPT dialogues.The bug was quickly resolved.Source: §3.2 · p. 22
• On April 28, 2023, ChatGPT returned to Italy after OpenAI added data-use information, an EU training-data objection option, and age verification.Italy had temporarily suspended the service amid concerns that included data collection, age verification, generated inaccuracies, and a recent breach.Source: §3.2 · p. 22
• For businesses, generative AI use may expose confidential customer data or trade secrets and create risks of data leaks and legal repercussions.A report also warns about potential future use of confidential data by AI companies.Source: §3.2 · p. 22
• Privacy protection in sensitive domains requires safeguards for personal, patient, and company information, with users, organizations, governments, and technology companies all playing roles.The section emphasizes that professionals handling sensitive data must preserve confidentiality and make informed decisions about how information is handled.Source: §3.2 · p. 23

3.3 Copyrights Issues

• Copyright concerns arise both from potentially copyrighted training data and from generated works whose ownership can depend on dataset origin and similarity to source works [65].The section also notes that ChatGPT’s lack of explicit source attribution raises concerns about possible infringement.Source: §3.3 · p. 23
• ChatGPT responses can closely resemble copyrighted sources without attribution, including definitions and code snippets in niche languages.The section gives an example of a generated white-paper definition that Gordon Graham said closely resembled his own.Source: §3.3 · p. 24
• ChatGPT-generated code for the 8 queens problem closely resembled published Prolog code without credit, and omitted a command present in the original.The original code included the “use module” command needed for the generated code to execute.Source: §3.3 · p. 24, Figure 10 · p. 24
• The legal system has not fully adapted to AI-generated content, and companies and individuals are pursuing disputes to assert and protect their rights.The section also reports that certain generative AI models may incorporate others’ content, including code, and that identical outputs may raise infringement concerns.Source: §3.3 · pp. 23–24
• The section describes competing views on whether AI developers, the AI system, or users who prompt it should be considered authors and rights holders.It also notes that US guidance recognizes only human authors and that AI-generated outputs themselves do not qualify for copyright protection.Source: §3.3 · p. 25
• In 2022, the US Copyright Office Review Board denied protection to an artwork created entirely by AI because it lacked human creative intervention.The Board reasoned that a purely mechanical process without a human author’s creative intervention does not meet the human-creation requirement.Source: §3.3 · p. 25
• A graphic novel’s human-selected and arranged composition and words received copyright protection, while its Midjourney-generated images did not.The example illustrates the distinction between human contributions and AI-generated material in a single work.Source: §3.3 · p. 25
• Copyright protection for AI-generated works depends on meaningful human creative contribution; ownership questions remain complex and legal guidance is still evolving.The section notes that rights vary across jurisdictions and that policies and guidelines are still needed.Source: §3.3 · p. 25

3.4 Algorithmical Bias

• Bias may reflect training data, model specifications, algorithmic constraints, product design, and policy decisions, rather than training data alone.The section also describes bias as potentially introduced during data collection, algorithm design, and implementation.Source: §3.4 · p. 26
• ChatGPT can produce biased, discriminatory, or harmful responses, and users have found ways to bypass filters and guardrails.Figure 11 shows a dialog ending in a jailbreak that returns biased Python code.Source: §3.4 · p. 26, Figure 11 · p. 27
• A cited study found racial differences in toxicity and reported that assigning ChatGPT a persona significantly increased toxicity, with outputs engaging in harmful stereotypes [20].The section notes that some racial queries are filtered most of the time, although racial and stereotypical text may still appear.Source: §3.4 · p. 26
• Hartmann et al. [34] prompted ChatGPT with 630 political statements from two leading voting advice applications and the nation-agnostic political compass test, finding converging evidence of a pro-environmental, left-libertarian political orientation.Source: §3.4 · pp. 26, 28
• The section reports that ChatGPT responses may not reflect objective truth or a complete and unbiased perspective, so users are advised to examine outputs critically.It suggests that personal judgment can help identify potential bias or inaccuracy.Source: §3.4 · p. 28

4 Flow Charts for Efficient ChatGPT Usage

• For inaccurate or unhelpful responses, the cited prompting guidance recommends correcting the response and using clear goals, concise messages, and specific questions.The guidance also advises natural-language prompts and avoiding jargon, technical, vague, or ambiguous wording.Source: §4 · p. 28
• The safety checklist recommends cross-checking information with multiple sources, requesting sources, checking for updates, and citing sources to avoid copyright violations.These steps address accuracy, time-sensitive information, and attribution.Source: §4 · p. 28
• Figure 13 outlines iterative interaction with ChatGPT to obtain informative responses or programming code until a satisfactory solution is reached.Figure 12 depicts ways ChatGPT can assist at stages of the research process.Source: §4 · p. 28, Figure 12 · p. 29, Figure 13 · p. 30
• The paper says step-by-step prompting may improve accuracy, while asking ChatGPT for a definite answer before an explanation may yield an inaccurate response.The section states that ChatGPT may commit to its output and fail to acknowledge an obvious mistake.Source: §4 · pp. 28, 30
• In healthcare, military operations, and engineering, users must exercise extra caution and consider tool limitations, privacy, trustworthiness, and responsibility.The section recommends a checklist for responsible use in critical areas such as healthcare and engineering.Source: §4 · pp. 28, 30

5 Conclusions

• ChatGPT can augment productivity, problem-solving, and writing across scientific research, mathematics, education, programming, and healthcare, but over-reliance carries ethical and reliability risks.The risks include incorrect responses, erroneous code, limited logical reasoning, and potential overconfidence.Source: §5 · pp. 30, 32
• ChatGPT should augment rather than replace human expertise, and human–AI collaboration can foster innovation and discoveries.Source: §5 · p. 32
• The authors recommend iterative interaction and independent verification, with professionals leveraging the tool's strengths thoughtfully.Their methods and flowcharts are based on comprehensive experimental studies.Source: §5 · p. 32
• Future research should develop collaborative environments and workflows for complex tasks, investigate AI bias, and evaluate educational outcomes and healthcare mental-health support.The proposed healthcare research includes empathetic dialogue, symptom detection, and monitoring users’ mental states, with ethical implications, limitations, and benefits studied.Source: §5 · p. 32
• The authors call for further study of AI explainability, user feedback, and ethical questions about AI–human creative work, including copyright and the balance of contributions.Source: §5 · p. 32
• Appropriate regulations, clear procedures, and effective working rules could improve chatbot output quality and support safer, more efficient use across domains.Source: §5 · p. 32