The proliferation of generative AI tools has been followed by the emergence of numerous AI-detection tools. Trained on both human and AI-generated content, these tools claim to be able to reliably distinguish between the two, focusing on concepts like burstiness and perplexity. However, AI detection tools have been plagued by frequent false positive and false negative results. This means that there is currently no AI detection tool with a reliable success rate.
Use the resources below to find additional information :
- “Can AI-generated text be reliably detected?” (February 2024)Article detailing an experiment to study AI detection tools and show that they “are not reliable in practical scenarios”.
- “Why AI detectors think the US Constitution was written by AI” (July 2023)Article by Benj Edwards in ArsTechnica that provides background into the fallability of AI detectors.
- OpenAI “AI Classifier” Tool Discontinuation Notice (July 2023)Blog post from OpenAI noting that they have shut down their AI detection tool “due to its low rate of accuracy”.
- “’Nobody wins in an academic-integrity arms race’: How artificial intelligence is changing the way colleagues think about cheating” (June 2023)Excerpts from an online forum hosted by Ian Wilhelm for The Chronicle of Higher Education discussing issues of plagiarism and generative AI.
Since the wide release of ChatGPT 3.5 in November 2022, generative AI tools have continued to emerge and evolve at a rapid pace. New tools, many of which are open-source and freely available, are made available seemingly every day. Popular tools improve their processing capabilities, offer free and paid tiers of access, and force large companies like Microsoft and Google to enter the generative AI field. As more users interact with generative AI tools, and as the amount of training data available to those tools grows, it can be difficult to both keep up with changes to existing tools and assess the quality of new tools.
Use the resources below to find additional information:
- “Artificial Intelligence, Explained” (July 2023)Article by Jennifer Monahan published through Carnegie Mellon University’s Heinz College that covers the basics of artificial intelligence, including a timeline that tracks its evolution from the 1950s to now.
- Generative AI – Statista Market InsightReport from Statista that provides data on market trends and the overall monetary value of generative AI products, as well as growth predictions in the coming years.
Open-Source AI Platforms
- The following websites offer access to new and open-source generative AI tools:
“Prompts” refer to a user’s initial request or input into a generative AI tool to receive a desired result (ex. “What are research topics related to...”). Currently, the semantic structure of those prompts is a major factor in the overall quality of the tool’s output. Below are articles that offer practical tips and advice on how to craft prompts when using generative AI tools for specific tasks. Additionally:
-
Many AI researchers predict that users will not need to focus as much on structuring their prompts in the future as tools continue evolving.
-
Using generative AI tools that have direct access to the live internet (ex. Bard, Bing Copilot) and/or more advanced tools (ex. ChatGPT 4 vs. ChatGPT 3.5) will also have an effect on output quality.
Use the following links for more information about prompt engineering, and guides for creating prompts for specific tasks:
In the context of generative AI tools, “hallucinations” refer to instances where tools will respond to users with seemingly accurate answers and information when in reality the outputs are false or nonexistent. For example, users have noticed that citations or bibliographic entries produced by LLMs often include made-up authors, articles, and journals. In part, hallucinations exist because of how generative AI tools function, predicting what the most likely word in a response should be instead of "thinking” through what the correct response actually is. While some tools are more prone to hallucinations than others, it is a problem that should make all users pause before trusting all generative AI outputs.
Use the resources below to find additional information:
- “When AI gets it wrong: Addressing AI hallucinations and bias”Article from MIT Sloan that provides background into AI hallucinations.
- UC Berkeley Library: Evaluating Resources GuideGuide from UC Berkeley that covers approaches to evaluate scholarly articles and resources found on the web for accuracy and bias. These skills are directly applicable to generative AI outputs to address hallucinations.
- “High rates of fabricated and inaccurate references in ChatGPT-generated medical content” (May 2023)Study published in Cureus on the accuracy rate of AI-generated citations in the medical field.
Most generative AI tools require users to provide personal information in order to use them, such as full names, phone numbers, and/or email addresses. Additionally, data, user prompts, and other content entered into generative AI tools can be added to their training datasets and used to further develop those tools. Consider the following when using generative AI tools:
-
Review all privacy policies – AI companies will post and update privacy and data retention polices on their tool websites. Look through those policies, including any changes from previous iterations, before inputting potentially sensitive data. For tools that lack a sound privacy policy (or any policy at all), consider using a different tool.
-
Be cautious with what you share – Know that any sensitive or personal information may be incorporated into generative AI datasets, so exercise caution when using proprietary data, controlled information, or student information (as this may constitute a FERPA violation).
Use the resources below to find additional information:
- “From ChatGPT to ThreatGPT: Impact of generative AI in cybersecurity and privacy” (August 2023)Paper published in IEEE Access detailing the various cybersecurity risks posed by ChatGPT and other generative AI tools.
- Generative Artificial Intelligence and Data Privacy: A Primer (May 2023)Report from the U.S. Congressional Research Service that provides background info into generative AI, associated privacy concerns, and the current privacy legislative context in the U.S.
- “Three ways AI chatbots are a security disaster” (April 2023)Article by Melissa Heikkilä in MIT Technology Review that discusses security issues of generative AI tools through the examples of jailbreaking, phishing scams, and data poisoning.
- “ChatGPT is a data privacy nightmare. If you’ve ever posted online, you ought to be concerned” (February 2023)Article by Uri Gal in The Conversation about ChatGPT and OpenAI’s issues with training data bias and privacy concerns.
Generative AI tools require a significant amount of computational processing power to function, which is provided by high-performance servers housed in physical data centers located across the country. These centers require massive amounts of electricity to keep tools operational, as well as water to keep the servers cool. Many AI companies have not revealed just how much electricity and water are used by their tools, or how much will be needed in the future. As such, there are significant unanswered questions about the environmental costs of keeping generative AI tools functional.
Use the resources below to find additional information
- “Generative AI’s environmental costs are soaring – and mostly secret” (Februray 2024)Article by Kate Crawford in Nature World View that uses U.S. Senator Ed Markey’s proposed bill, the "Artificial Intelligence Environmental Impacts Act", to frame a discussion around AI’s environmental costs.
- “Sam Altman seeks trillions of dollars to reshape business of chips and AI” (February 2024)Article on the future reshaping of global supply chains to support generative AI expansion, and the adverse environmental effects that could have.
- “Inside AI's giant land grab” (December 2023)Article by Issie Lapowsky in Business Insider that provides detail into the construction and spread of data centers across the U.S. in support of generative AI companies.
- “Risks and benefits of large language models for the environment” (Februray 2023)Article published in Environmental Science & Technology that covers the “direct and indirect environmental impacts and effects” of LLMs.
- “On the dangers of stochastic parrots: Can language models be too big?” (March 2021)Article presented at the 2021 ACM Conference on Fairness, Accountability, and Transparency that discusses the environmental costs associated with large-language models that require massive amounts of data storage and processing power.
Because many generative AI tools can produce human-like responses and complete various repetitive tasks, there are concerns about how AI tools will impact the human labor market. For example, some writers and language translators have been replaced by their employers and contractors and replaced by AI tools. Attention has also focused on the economic exploitation of some AI content moderators, or workers tasked with training and improving AI tools. While these are not the only labor-related concerns, continued attention is needed to ensure fair and equitable treatment of all workers potential impacted by AI.
Use the resources below to find additional information:
- Ethics of Artificial IntelligenceWebpage from UNESCO with inforamtion on their development of guiding ethical principles for AI use, including labor practices.
- “Hollywood actors strike ends with a deal that will impact AI and streaming for decades” (November 2023)Article by Angela Watercutter and Will Bedingfield in Wired that discusses the end of the 2023 strike by the Writers and Screen Actors Guilds. Specifically, the article discusses the agreements reached on anti-AI provisions.
- “AI is a lot of work” (June 2023)Article by Josh Dzieza in New York Magazine’s Intelligencer that provides an overview of AI content moderators and the issues they face.
- “We are all AI’s free data workers” (June 2023)Article by Melissa Heikkilä in MIT Technology Review with information on AI content moderators, and how data scraped from the Internet to create AI training datasets constitutes a form of “free labor”.
- “OpenAI used Kenyan workers on less than $2 per hour to make ChatGPT less toxic” (January 2023)Article by Billy Perrigo in Time that reports on the exploitation faced by Kenyan content moderators to support ChatGPT.
Content warning: this story contains descriptions of sexual abuse
To form their training data sets, many generative AI tools have scraped information from a wide range of sources on the Internet with no notice to the sources’ original owners/creators. Some of this information is protected under copyright, which may mean that generative AI companies face legal problems because of their use of such material. Two of the most pressing questions on copyright and AI are:
-
Does the use of copyrighted material by generative AI companies constitute copyright infringement, or is it a legal form of fair use?
-
Will human-prompted outputs from generative AI tools be afforded copyright protection?
As of March 2023, the U.S. Copyright Office has provided an answer to the second question: no. However, this may change as more generative AI companies are subjected to legal scrutiny.
Use the resources below to find additional information:
- “The Times sues OpenAI and Microsoft over A.I. use of copyrighted work” (December 2023)Article from the New York Times detailing the news organization’s ongoing lawsuit against OpenAI.
- “Generative AI meets copyright” (July 2023)Article by Pamela Samuelson published in Science that provides an overview of the various copyright questions connected to generative AI.
- “How we think about copyright and AI art” (April 2023)Article from Kit Walsh published through the Electronic Frontier Foundation on the different interpretations of applying copyright protection specifically to AI-generated art.
- Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence (March 2023)Document from the Federal Register on the U.S. Copyright Office’s most recent ruling that copyright protection only be afforded to works created purely by humans (that was affirmed following the August 2023 ruling in the court case Thaler v. Perlmutter).
Generative AI’s applications in research and writing have raised questions from scholars about how to cite such uses, and whether or not journal publishers will allow the inclusion of AI assistance in submitted papers. The major concerns include generative AI potentially providing incorrect information presented as fact (i.e. hallucinations), data privacy, and dataset bias. While there is no universal standard to citing and/or including generative AI use in published papers, many scholarly journals and publishers have created guidelines for scholars to follow when submitting works for publication.
Use the resources below to find additional information:
- Committee on Publication Ethics (COPE) - Authorship and AI ToolsStatement from COPE clarifying the organization’s position on incorporating AI use in published works.
- “ChatGPT listed as author on research papers: many scientists disapprove” (January 2023)Article by Chris Stokel-Walker in Nature that provides early examples of generative AI’s inclusion in published works, and an overview of the associated ethical/legal questions.
- Retraction Watch: "Papers and peer reviews with evidence of ChatGPT writing"List of published articles collected by Retraction Watch that show evidence "they were written by ChatGPT".
Generative AI and Publishing Guidelines
-
The following are examples of generative AI guidelines from select scholarly journals and publishers. These can (and will) evolve, so please check directly with these sources and with specific publishers for the most up-to-date guidelines: