Dear colleagues,
Artificial intelligence (AI) has the potential to revolutionize industries, but it also risks perpetuating bias, inequality, and reinforcing societal power imbalances. Biased algorithms and discriminatory datasets can disadvantage marginalized groups, limiting opportunities and fairness in AI-driven decision-making.
To mitigate these risks, what strategies should AI developers implement to ensure transparency, accountability, and fairness in AI models? How can ethical AI principles be integrated into model training, data collection, and real-world deployment to protect underrepresented communities?
Developers can ensure ethical, fair, and responsible AI development by incorporating transparency, bias mitigation, and accountability measures in the AI lifecycle. These include using diverse and representative data sets, regular auditing of models for biases, ensuring explainability, and adhering to ethical principles such as fairness, privacy, and security. Working with interdisciplinary teams, stakeholders, and policymakers also aligns AI systems with societal values and legal expectations. In addition, continued monitoring and updating of AI models ensure they remain ethical and fair over time.
Sondos Feriel Naoum
Dear Shahnawaz Mohammed,
Thank you very much for your interest and I appreciate your effort. I hope to see you in this field again.
All the best ,
sondos.
Deep learning models, despite their powerful capabilities, often function as "black boxes", making it difficult to comprehend their decision-making processes. This lack of transparency creates several key challenges:
Enhancing Model Transparency
Several methods have been developed to improve deep learning interpretability:
- Feature Importance Analysis: Identifies key input features influencing predictions (SHAP values, permutation importance).
- Visualization Techniques: Tools like saliency maps and activation maps help visualize how models process information.
- Rule Extraction: Converts complex model behavior into human-readable decision rules.
- Attention Mechanisms: Highlights critical input sections that impact predictions.
- Surrogate Models: Uses a simplified model to approximate and explain the behavior of a more complex deep learning model.
Additional Considerations
- Accuracy vs. Interpretability Trade-off: Enhancing explainability may slightly reduce model performance, requiring a balance.
- Context Sensitivity: The need for transparency varies by application domain—higher stakes require more explanation.
- Human-Centered Explanations: Interpretations should be clear, meaningful, and tailored to users' expertise levels.
I really recommend you to go to youtube and have a look at the several things people build with AI, like for example one young man who build an automatic tracking device that always marks the forehead. If you linked it to a weapon... I guess you understand what one can do with it. And those are scripts everyone can build. You do not have to work for a big company. You do not even need a data center. Those are scripts that can run on cheap systems like rasberry pi devices. Those are scripts that can be easily run connected to drones.
Those scripts are already out. Everybody can script them. You cannot controll that anymore.
You are only thinking of huge AIs. As if it is all about the big companies and underestimating everything else.
https://www.youtube.com/watch?v=N5vzpi4yF7A
Ultimately, the successful deployment of an IAI system hinges on a balanced approach to data collection, preprocessing, and evaluation. By systematically addressing the key elements of your IAI, you can make informed decisions about the value and feasibility of investing in these systems...
HOW TO FIND THE RIGHT BALANCE OF DATA FOR YOUR INDUSTRIAL AI SYSTEM
"Understanding the inner workings of your Industrial Artificial Intelligence (IAI) system is crucial if you want it to add measurable value to your manufacturing operations. In this blog, we will dig into one important aspect of every AI — the inputs, a.k.a. your data. Including the right type and the right amount of data is key to success when it comes to AI.
To understand the full potential of an IAI system, it’s important to look at what data, assumptions, rules, and shortcuts feed into the system during its training, testing, development, and deployment stages. You won’t be able to know or anticipate everything that goes into AI from your system, but understanding these things (to the extent possible) can give you a better understanding and more realistic expectations of how your AI will behave..."
https://insights.cermacademy.com/464-how-to-find-the-right-balance-of-data-for-your-industrial-ai-system-michael-sharp/
Copyright’s Big Win in the First Decided US Artificial Intelligence Case
"Back in March of 2023, when there were only a handful of cases alleging copyright infringement for training purposes by AI companies, I predicted that we would soon have some guidance from the court in Thomson Reuters Enterprise Center GMBH and West Publishing Corp. V Ross Intelligence, Inc. Predicting the timing of court decisions is a fool’s errand, and this fool was repeatedly wrong in his predictions on timing. Nonetheless, on February 11, the Ross case did in fact become the first US decision on the merits to directly address copying to train AI. Now we have a clear decision, and it is favorable for rightsholders.
The case arose out of the surreptitious copying of the entire Westlaw database (after having been denied a license) by a company that wanted to create an arguably competing product. Unlike some of the generative AI cases, there was no claim that the AI’s output included the copyrightable content. As the Court noted, this was not about generative AI. This case was purely about training...
The Ross case might not be the same as a generative AI case, but it is more similar than anything else decided to date. This case involves training, as do the other pending AI cases.
Does this mean that the plaintiffs will win all of the generative AI cases filed? Of course not. That is not how copyright works. Fair use requires an analysis of case-specific facts. Still, defendants now have a much higher burden to overcome and will need to find ways to distinguish their cases from this one."
https://scholarlykitchen.sspnet.org/2025/02/20/copyrights-big-win-in-the-first-decided-us-artificial-intelligence-case/?informz=1&nbd=&nbd_source=informz
Dear DR.Ljubomir Jacić,
thank you very much for the valuable information.
We Could Use a Model Licensing Framework for Scholarly Content Use in AI Tools
"As artificial intelligence (AI) systems become increasingly integrated into scholarly research and knowledge dissemination, publishers face a critical challenge: how to best integrate their content into this new ecosystem. Currently, an increasing number of publishers are collaborating with AI system suppliers to license their content for large language model training, small model development, and retrieval-augmented generation applications. Many of the largest publishers have announced licensing arrangements with the largest model developers and likely many, many more are under the radar at other publishers as well...
The scholarly publishing industry has already navigated the challenges of standardizing content licensing for online journal access. While the framework of model licenses hasn’t removed the need for negotiation, nor entirely replaced the need for legal review, it has simplified the process and saved everyone’s time. These past efforts provide a valuable blueprint for addressing the emerging complexities of AI content licensing. A community initiative to support the development of model licenses for the use of scholarly content by AI developers could streamline negotiations, reduce costs, and encourage more responsible AI content-usage practices. In turn, using license frameworks could ensure that AI-generated outputs respect the intellectual contributions of academic authors and uphold the credibility of published research. As the landscape is rapidly evolving, getting such an initiative moving quickly should be a priority."
https://scholarlykitchen.sspnet.org/2025/02/26/we-could-use-a-model-licensing-framework-for-ai-tools/?informz=1&nbd=&nbd_source=informz
Ensuring ethical, fair, and responsible AI development requires a multifaceted approach:
Addressing Bias and Inequality in AI
The use of artificial intelligence in scientific research with integrity and ethics
"This paper addresses the evolution of Artificial Intelligence (AI) in scientific research and the ethical and integrity challenges that arise with its integration. AI has become an indispensable tool for researchers, accelerating discoveries and optimizing processes. However, using these algorithms raises concerns about bias, transparency, and accountability. The ability of machines to learn and create knowledge challenges the paradigms of authorship and credibility, putting integrity and ethics under new scrutiny. The discussion emphasizes robust ethical governance, collaboration among stakeholders, ongoing education, and the creation of transparent and auditable algorithms. It further highlights the importance of maintaining ethics and integrity at the heart of AI research to ensure its advancement benefits humanity fairly and responsibly, emphasizing the need for a holistic approach involving education, transparency, accountability, and active participation of multiple stakeholders. Finally, it reiterates that as we embark on this new era of AI-driven discovery, we must embrace both the opportunities and the ethical challenges it presents, ensuring that the use of AI in scientific research continues to benefit humanity by promoting knowledge and well-being."
Article The use of artificial intelligence in scientific research wi...
What a case, the first case of student expelled from the University over the use of AI.
Ph.D. student sues UMN, files human rights complaint after AI plagiarism expulsion
"Doctorate student Haishan Yang denied accusations of plagiarism and filed a lawsuit against the University of Minnesota and an individual professor, citing defamation...
The University of Minnesota expelled a third-year Ph.D. student in January after he was accused of using generative artificial intelligence (AI) on an exam. He appears to be the first student expelled from the University over AI..."
https://mndaily.com/292797/campus-administration/ph-d-student-sues-umn-files-human-rights-complaint-after-ai-plagiarism-expulsion/
Will AI jeopardize science photography? There’s still time to create an ethical code of conduct
Generative artificial-intelligence illustrations can be helpful, but fall short as scientific records...
https://www.nature.com/articles/d41586-025-00532-2
Stepped Approach to Managing Responsible AI: An Interview with Aaron Wood on Employee Policies, Rights Reservation, and Research Integrity
"While some journal publishers have been addressing AI (or, at least text and data mining) for more than a decade, many organizations are at the earliest planning stages in terms of their AI strategy. However, the American Psychological Association (APA) has been consistently ahead of the curve in addressing the implications of AI as a publisher, a licensor, and a user...
I think the most important thing is to create the space in your organization for the conversation, because it can’t really be led from just one area or the other. You really need staff involved to have an open discussion about what’s going on. That’s the only way to get to the point where you’re looking at each application, each license, and each tool individually, and not thinking of AI in generalities."
https://scholarlykitchen.sspnet.org/2025/04/14/american-psychological-associations-stepped-approach-to-managing-responsible-ai-an-interview-with-aaron-wood-on-employee-policies-rights-reservation-and-research-integrity/?
The Open Access – AI Conundrum: Does Free to Read Mean Free to Train?
"When the Open Access (OA) movement gained momentum in the early 2000s, its proponents envisioned a world where research findings would be freely available to all readers — breaking down paywalls that limited knowledge dissemination and hindered scientific progress. By the 2010s, an increasing number of countries began to mandate open access publications, arguing that publicly funded research should be freely available to the public. The readers envisaged by proponents of OA were obviously human (academics as well as the wider public). While text mining had been considered as one potential application, they could not foresee the development of large language models (LLMs) which would begin to rapaciously ingest large amounts of text. OA literature has become particularly attractive for AI training precisely because it lacks the legal and technical barriers that might protect traditionally published content...
Barriers to using copyrighted content have not, however, provided effective protection (or remuneration) to published authors in other areas. As media reporting on ongoing lawsuits has made abundantly clear, the major AI models that have transformed the digital landscape since Chat-GPT’s launch in 2023 were trained on copyrighted material on a large scale...
OA licenses could mandate that AI tools trained on OA papers need to include citation capabilities in exchange for the free use of high-quality material, to ensure that the creators of academic are appropriately recognized. As OA publishing is rapidly becoming a significant share of academic output, OA policies around AI use could reshape the technologies engagement with research substantially. Otherwise, the underlying incentive structure of academic careers could be quickly eroded through widespread AI use, disproportionately affecting a new generation of researchers. Going forward, we may need to consider alternative impact indicators for AI reuse alongside traditional citation counts, if these become artificially depressed by extensive AI reuse without attribution. These are urgent issues for scholarly publishing, especially when it comes to OA, which will have an impact on future generations of research users worldwide."
https://scholarlykitchen.sspnet.org/2025/04/15/guest-post-the-open-access-ai-conundrum-does-free-to-read-mean-free-to-train/?informz=1&nbd=&nbd_source=informz
National research centre to strengthen responsible AI use
The Danish government has launched a major initiative, the National Centre for Artificial Intelligence in Society, led by the University of Copenhagen in collaboration with Aalborg University. It is aimed at strengthening Denmark’s leadership in responsible AI innovation and development...
https://www.universityworldnews.com/post.php?story=20250520183227783
Most AI researchers reject free use of public data to train AI models
A new survey finds only a quarter of AI researchers support unrestricted AI model training using publicly available data. As creatives protest unauthorised use of their work, tensions mount between government policy, industry ambitions and public expectations. Cian O’Donovan explores what researchers think, where they align with the public and where critical gaps remain...
"There are growing concerns among writers and academics about the use of their work in AI development without consent. The Atlantic magazine has covered that engineers at Meta had downloaded more than 7 million books and 80 million research papers from vast pirated libraries online. A recent petition by the Society of Authors, representing over 11,000 members, highlights the call for transparency and fair use practices in AI training.
New research from UCL reveals that many AI researchers support stronger ethical standards for training data. A global survey found that only one in four respondents believe AI companies should be allowed to train their models on any publicly available text or images. The majority favour more restrictive approaches, including the requirement for explicit permission from content creators..."
https://blogs.lse.ac.uk/impactofsocialsciences/2025/06/19/only-a-quarter-of-ai-researchers-support-a-free-for-all-on-training-data/
When AI Doesn’t Understand You: A New Form of Global Inequality
Many of the world’s languages are underrepresented in the data used to train generative AI chatbots and other AI-based tools, which means entire communities could be missing from the foundational models underlying artificial intelligence. The solution isn’t just getting better data but better governance...
https://undark.org/2025/08/07/opinion-ai-language-inequality/
A Smarter Way to License Research Articles for AI
"The recent leap in generative AI technology, led by companies like OpenAI, Anthropic, Google, and Microsoft, has sparked both awe and anxiety across industries. But while these models are impressive, they remain fundamentally limited by the quality and scope of the data they can access. In general, large language models (LLMs) cannot access research articles beyond abstracts or open-access articles. For the scientific and publishing communities, this presents both a challenge and an opportunity...
LLMs are not yet capable of meaningfully engaging with most peer-reviewed research articles. Why? Because most of the content is locked behind paywalls and isn’t accessible. If LLMs are the future of information discovery, valuable research disseminated by scholarly publishers risks being left behind — unless we build a bridge between authoritative research and intelligent retrieval...
Many publishers have been reluctant to engage with large AI companies for a variety of reasons. Some believe AI companies are infringing on their copyright, some believe this could significantly threaten their business, and others are waiting to see how their peers act. There is an arc of how publishers interact with AI and AI companies: at one end of the spectrum are groups that are blocking and fighting AI companies, while others announce large licensing deals and have formed direct partnerships ...
Crucially, our work to build the next generation of citations is created under indexing agreements directly with publishers. I think there is a function here for AI companies and publishers that can help bridge the gap in a way that protects researchers and publishers, yet connects the most powerful LLMs in the world to the most authoritative content.
Smart Citations (or their equivalent) could serve as a model for licensing research articles to LLMs for use in Retrieval-Augmented Generation (RAG). A pan-publisher unified dataset of annotated citations could be sublicensed, and usage could be tracked (even in a COUNTER-compatible way). I believe it is a new way to responsibly bridge the gap between AI / LLMs and scientific publishers.
This model delivers clear wins:
Attribution: If an LLM can’t interact with an article, it can’t cite it. By directly licensing content in a controlled way, content would be more likely to be surfaced and cited in chatbots.
Reliability: LLMs are hard to trust. By linking LLMs to authoritative content, their outputs become more trustworthy and comprehensive. Citation snippets could be surfaced for easy verification by the end user without taking away value from the version of record.
Discovery: Similar to abstracts helping content be more discoverable on the web while not giving up the entire article, citation snippets would preserve the version of record and drive researchers’ and publishers’ traffic.
Compensation: Publishers would be remunerated for the use of the version of record in a discovery environment. The more times an LLM interacts with the article, the more times the content is paid for, whether that is a “pay to crawl” type of licensing or as part of a subscription package demonstrating usage. Think of “AI reads” as the new reads and downloads.
This approach could be expanded, building the next generation of citations not just for the research space but building them for LLMs. The chatbots are being used by real people with real questions, sometimes directly impacting their lives in very serious ways. We as a community ought to think about how we can ensure society and the world have the best information possible. I think publishers are now more important than ever."
https://scholarlykitchen.sspnet.org/2025/08/07/guest-post-a-smarter-way-to-license-research-articles-for-ai/?informz=1&nbd=&nbd_source=informz
AI models display subliminal learning
When data created by one AI model are used to train another, they can subliminally pass on problematic tendencies. In one example, researchers fine-tuned a ‘teacher’ AI model to display preferences for owls, had it generate an owl-free dataset of meaningless numbers and then used those data to fine-tune a ‘student’ model. When asked for its favorite bird, the ‘student’ model was likely to still pick an owl, showing that models can transmit hidden traits through statistical patterns in their outputs. “We conclude that subliminal learning is a general phenomenon that presents an unexpected pitfall for AI development,” write researchers with Anthropic and AI safety group Truthful...
https://www.theverge.com/ai-artificial-intelligence/711975/a-new-study-just-upended-ai-safety
Scientists want to prevent AI from going rogue by teaching it to be bad first
Researchers are testing new ways to prevent and predict dangerous personality shifts in AI models before they occur in the wild...
https://www.nbcnews.com/tech/tech-news/ai-anthropic-researchers-predicting-dangerous-behavior-rcna223236
- Badges
- Science topic
- Similar topics
- Engineering
- Modeling