Source-linked AI summary
ChatGPT Needs SPADE (Sustainability, PrivAcy, Digital divide, and Ethics) Evaluation: A Review
Sunder Ali Khowaja, Parus Khuwaja, Kapal Dev, Weizheng Wang, Lewis Nkenyereye
TL;DR
ChatGPT’s popularity has brought overlooked concerns about sustainability, privacy, digital divide, and ethics. This review develops a SPADE evaluation perspective, supports it with preliminary analysis, and proposes mitigations and policy recommendations, including discussion of the EU AI Act. It concludes that these concerns require attention while fairness and policy recommendations remain context-dependent and subject to improvement.
Problem
Existing attention to ChatGPT has focused on performance, integration, and applications, while a collective evaluation of sustainability, privacy, digital divide, and ethics is limited.
Method
The paper reviews ChatGPT-related concerns using issue-specific searches, preliminary data collection, visualizations, hypotheses, and source verification, then discusses mitigations and EU AI Act policies.
Results
The review reports concerns spanning energy consumption and carbon emissions, privacy and copyright, digital access, and bias, power, transparency, and fairness in large language models.
Takeaways & Limitations
The paper recommends SPADE evaluation for ChatGPT and subsequent conversational bots, alongside mitigations and inclusive, accountable AI policies.
Takeaways & Limitations
The authors state that perfect fairness remains difficult and ongoing, while policy recommendations are broad and context-dependent rather than exhaustive.
Abstract
from arXiv · showhide
ChatGPT is another large language model (LLM) vastly available for the consumers on their devices but due to its performance and ability to converse effectively, it has gained a huge popularity amongst research as well as industrial community. Recently, many studies have been published to show the effectiveness, efficiency, integration, and sentiments of chatGPT and other LLMs. In contrast, this study focuses on the important aspects that are mostly overlooked, i.e. sustainability, privacy, digital divide, and ethics and suggests that not only chatGPT but every subsequent entry in the category of conversational bots should undergo Sustainability, PrivAcy, Digital divide, and Ethics (SPADE) evaluation. This paper discusses in detail the issues and concerns raised over chatGPT in line with aforementioned characteristics. We also discuss the recent EU AI Act briefly in accordance with the SPADE evaluation. We support our hypothesis by some preliminary data collection and visualizations along with hypothesized facts. We also suggest mitigations and recommendations for each of the concerns. Furthermore, we also suggest some policies and recommendations for EU AI policy act concerning ethics, digital divide, and sustainability
TITLE PAGE
The paper is authored by Sunder Ali Khowaja and affiliated with institutions in Ireland, Hong Kong SAR, and South Korea.
- Sunder Ali Khowaja is identified as the corresponding author.
- The listed affiliations include Munster Technological University in Cork, Ireland, City University of Hong Kong, and Sejong University in Seoul, South Korea.
1. Introduction
The introduction situates ChatGPT and other large language models within rapid advances in generative AI and argues that they should be evaluated for sustainability, privacy, digital divide, and ethics.
- Large language models have improved through large-scale data, deep learning, and Transformer-based architectures that model contextual and long-range linguistic information.
- Generative AI systems increasingly support unimodal, cross-modal, and multimodal content generation across research and industry applications.
- The paper proposes evaluating ChatGPT and other generative AI systems through SPADE: sustainability, privacy, digital divide, and ethics.
- The authors identify unresolved policy questions involving energy use, access between developed and developing countries, user privacy, and copyrighted material under the EU AI Act.
2. Selection Criteria
The review uses issue-specific searches and filters sources for authenticity, then verifies analytical claims through additional resources.
- The authors state that no single existing study collectively addresses these four concerns for large language models and ChatGPT.
- The review organizes searches around sustainability, privacy, digital divide, and ethics, with issue-specific keywords for each concern.
- Sources were selected from authentic documentation, published papers, and interviews, while analytical figures, facts, and hypotheses were checked against additional resources.
3. Sustainability
The sustainability review examines training, inference, lifecycle energy, and financial costs of large language models, while proposing efficiency and renewable-energy mitigations.
- 3.1 Sustainability for Training LLMs like ChatGPT: 500 metric tons of CO2 were reported for GPT-3 training versus 75 metric tons for Meta’s OPT, while the paper’s estimate was lower.
- 3.2 Sustainability for LLM Lifecycle and inferences like ChatGPT: 29.3 terawatt-hours per year is the projected energy use for generative AI and LLMs, with inference expected to consume more energy than training.
- 3.3 Estimated Training Cost of LLM: GPT-3 and BLOOM are estimated to incur training costs of 1.80 and 2.29 million dollars, respectively.
- 3.3 Estimated Training Cost of LLM: Training-cost estimates use GPU-hour calculations and FLOP-based cloud pricing assumptions, including TPU performance and utilization estimates.
- 3.4 Mitigation and Recommendation: Recommended mitigations include efficient hardware, renewable energy, fine-tuned training data, optimized architectures, dynamic resource allocation, and reduced redundant computation.
4. LLM (ChatGPT) Privacy Concerns
The section describes privacy risks from ChatGPT’s data practices, copyrighted training material, exposed conversations, and open model access, then outlines measures to reduce those risks.
- ChatGPT’s training data may include personal information gathered without consent, creating privacy violations through loss of contextual integrity.The concern applies even when information was publicly available if it is reused outside the context in which it was originally shared.
- ChatGPT may reproduce or paraphrase copyrighted or proprietary material without citing or crediting the original source.
- Privacy risks also include exposed conversation titles and chat histories, leaving users’ private information potentially vulnerable.
- Open access to LLMs can enable misinformation, harassment, fraud, privacy violations, and cybercrime when ethical safeguards are limited.
- 4.1 Mitigation and Recommendation: Recommended mitigations include privacy protection, consent and user control, differential privacy, auditing, limited retention, federated learning, security, ethical policies, and user education.
5. Digital Divide
The paper presents the digital divide as unequal internet access, skills, and ChatGPT use across socioeconomic groups and countries. It reports large connectivity disparities and recommends affordability, localization, training, partnerships, and infrastructure improvements to broaden access.
- 5. Digital Divide: The paper frames ChatGPT’s distributional impact as unresolved, particularly regarding who benefits most and how low-income countries and Global South workers are affected.It also notes that LLMs such as ChatGPT are predicted to risk telemigrant jobs in developing countries.
- 5. Digital Divide: Low-income countries’ maximum average internet speed is 11 Mbps, below the median for upper-middle-income countries and nearly 6x below their maximum.The paper also reports that high-income countries’ maximum average speed is 19x and 8x higher than corresponding lower-income comparisons, although the supplied passage truncates the comparison.
- 5. Digital Divide: Students without access to ChatGPT or who do not use such tools may be disadvantaged, with access linked to basic knowledge and internet speed.The paper connects this pattern to digital inequality between students from higher-income and mid- to lower-income countries.
- 5. Digital Divide: 61.05% of ChatGPT website traffic came from high- and upper-middle-income countries, compared with 25.93% from low- and lower-middle-income countries.India alone accounted for 10.67% of the low- and lower-middle-income category’s traffic share.
- 5.1 Mitigation and Recommendation: The proposed mitigation agenda combines affordable access, reduced data usage, localization, multilingual support, capacity building, local partnerships, connectivity improvements, and user feedback.These measures are presented as ways for LLMs to promote more inclusive and relevant access across socioeconomic settings.
6. Ethics
The ethics discussion identifies privacy, bias, misinformation, academic-integrity, and other risks in ChatGPT and LLMs, then outlines mitigation and policy recommendations.
- ChatGPT and other LLMs raise ethical concerns involving privacy, bias, power, transparency, intellectual property, misinformation, and employment.
- Repeated prompting has elicited harmful responses involving torture, nationality-based profiling, surveillance, and racial bias.
- Proposed mitigations include diverse training data, counterfactual augmentation, debiasing, monitoring, education, originality policies, identity checks, and audits.
- Perfect fairness remains difficult and requires continuing improvement rather than a completed solution.
- User feedback helps identify biased, offensive, inappropriate, misleading, incomplete, or nonsensical outputs and supports ongoing model improvement.
- Academic-integrity concerns have led several states and universities to ban ChatGPT to reduce cheating and plagiarism.
8. Lessons Learned
The lessons learned call for AI policies centered on transparency, fairness, human control, safety, security, and periodic impact review, while acknowledging that recommendations depend on context.
- Recommended policy measures include transparency, fairness and bias mitigation, human oversight, safety, security, testing, validation, and monitoring.
- The recommendations are broad rather than exhaustive and may vary with the statutory body's context and requirements.
9. Conclusion
The conclusion frames SPADE as a needed evaluation for LLMs, reporting preliminary concerns about energy use, privacy, digital divides, ethics, and fair usage, alongside mitigations and EU AI Act recommendations.
- The study examines ChatGPT's sustainability, privacy, digital-divide, and ethical concerns through preliminary analysis and hypothesized findings.
- It reports energy consumption during training and inference, privacy risks in data collection and use, a digital divide across income groups, and ethical concerns over fair usage.
- The paper discusses the EU AI Act and offers mitigations, recommendations, and proposed AI-policy suggestions for the identified concerns.
Authors Contribution Statement
The contribution statement assigns authorship across sustainability, digital divide, privacy, ethics, manuscript preparation, infographics, and proofreading.
- Sunder Ali Khowaja handled sustainability and digital-divide sections, EU AI Act infographics, material compilation, and manuscript preparation.
- Parus Khuwaja and Kapal Dev contributed to the privacy section and manuscript proofreading.
- Weizheng Wang and Lewis Nkenyereye contributed to the ethics section and manuscript proofreading.
Funding
The study reports that it received no funding.
- No funding was obtained for this study.
- The manuscript identifies no funding source for the study.
- The study is reported as unfunded.
Data Availability Statement
The manuscript does not report data generation or analysis.
- The manuscript reports no data generation.
- The manuscript reports no data analysis.
- Data generation and analysis are not reported in the manuscript.