AI-Assisted Tagging in OpenStreetMap: A Case for Responsible Innovation and Copyright Compliance

Introduction

Hi everyone,

I’m David Osipov, an active OSM mapper based in Georgia. I’m passionate about improving the map, particularly enriching data for my home country, which boasts a rich cultural heritage and a burgeoning tourism industry.

Recently, I’ve been exploring the use of AI, specifically large language models (LLMs) like Gemini AI, to help me create more informative and multilingual name and description tags for points of interest (POIs). This approach, while innovative, has sparked debate within the OSM community, with some expressing concerns about the copyright implications of using AI-generated content. I understand and appreciate these concerns, and this post aims to address them directly, fostering a transparent and informed dialogue about responsible AI integration within OSM.

Why AI? Addressing OSM’s Challenges and Expanding its Impact

The OSM project thrives on the collaborative spirit of a diverse global community, united by a shared commitment to creating a free and open map for everyone. AI tools, when used responsibly and ethically, have the potential to significantly enhance this collaborative effort, accelerating data creation and expanding the map’s reach and impact in several ways:

  1. Bridging Language Barriers: Multilingualism is essential for making OSM truly accessible to a global audience. AI translation capabilities can efficiently generate high-quality translations of tags, enabling users who speak different languages to engage with and contribute to the map. This aligns with the broader goals of international organizations like WIPO, which advocate for inclusive and equitable access to knowledge and information. (World Intellectual Property Organization [WIPO], 2024).
  2. Enhancing Data Completeness and Accuracy: Many POIs in OSM lack descriptions or have incomplete information, limiting the map’s usefulness for users seeking detailed information. AI can assist in filling these gaps by extracting and summarizing relevant details from publicly available sources, improving the map’s overall accuracy and comprehensiveness.
  3. Unlocking the Power of Big Data: AI excels at processing and analyzing large volumes of data, a task that would be prohibitively time-consuming for humans, especially when dealing with multilingual sources. By leveraging AI, we can tap into the vast amount of information available online, enriching OSM with insights that would otherwise remain hidden or inaccessible due to language barriers.
  4. Empowering Mappers: AI can act as a powerful tool for mappers, freeing them from tedious and repetitive tasks like manual translation and data entry. For instance, manually translating a single description tag into 20 languages could easily take several hours. While I am proficient in Georgian, English, and Russian, I simply don’t have the language skills or the time to handle 20 languages. Hiring professional translators for each tag would be far too expensive for a volunteer project like OSM. Consider this: just reading through the 10 or more sources I typically gather for a single POI, which often totals over 20 pages of text, can take well over half an hour! AI assistance, however, allows me to overcome these limitations. It enables me to efficiently generate high-quality translations, saving me countless hours and making it feasible to contribute data in multiple languages. With AI assistance, I can achieve the same result in a matter of 30-40 minutes, allowing me to contribute significantly more data to OSM. This allows mappers to focus on more complex and creative contributions, such as adding new POIs, verifying data, and improving map features, ultimately leading to a richer and more dynamic map. It’s important to emphasize that AI is not meant to replace human mappers; it’s a tool to enhance our capabilities and make our contributions more efficient and impactful.

My AI-Assisted Tagging Process: An example and a breakdown of the AI-Human Partnership

To provide a transparent and concrete illustration of my AI-assisted tagging process, let’s delve into a specific example: the Rezo Gabriadze Marionette Theater and its whimsical Clock Tower in Tbilisi. These iconic landmarks, imbued with the artistic spirit of their creator, Rezo Gabriadze, deserve rich and detailed representation on the OSM platform.

Theater: https://www.openstreetmap.org/node/1567308849

Clock Tower: https://www.openstreetmap.org/way/1062718406

Before my edits, these POIs had a rather sparse presence on OSM. The theater, while marked on the map, possessed only rudimentary information:

  • Name: Initially only in English. Later, Georgian and Russian translations were added.
  • Address: Basic street address.
  • Website: A link to the official website.

The Clock Tower, a captivating structure that enchants visitors with its hourly angel performance, had even less data:

  • Name: Limited to Georgian and English translations, with Italian and Polish added later.
  • Tourism Tag: Marked as a tourist attraction.

This lack of detail and multilingual representation significantly limited the map’s usefulness for individuals seeking a deeper understanding of these unique landmarks.

My AI-assisted process, however, enabled me to dramatically enhance these entries, breathing life into their digital representations on OSM. Here’s a breakdown of the transformation:

Theater:

  • Name Tags: Expanded from 3 languages (English, Georgian, Russian) to a comprehensive 21, reflecting the global reach of this renowned theater.
  • Description Tag: A rich and evocative description, meticulously crafted and translated into 19 languages (Arabic, Azerbaijani, German, Spanish, Persian, French, Hebrew, Hindi, Armenian, Italian, Japanese, Georgian, Korean, Dutch, Polish, Portuguese, Russian, Turkish, Ukrainian, Chinese (Simplified), Chinese (Traditional)).
  • Additional Information: Added details about the architect, the theater’s capacity (80 seats), the year it was founded (1981), and a link to the online ticket booking platform, providing practical information for potential visitors.

Clock Tower:

  • Name Tags: Increased from 4 languages (Georgian, English, Italian, Polish) to a remarkable 23 languages.
  • Description Tag: A captivating narrative, translated into 20 languages, describing the tower’s unique features, the hourly angel performance, the architectural style, and the charming local name, “The Tower with the Angel.”
  • Architect and Construction Year: Added information about the architect (Rezo Gabriadze) and the year of construction (2010), providing historical context.

This significant enrichment of data demonstrates the power of AI as a collaborative tool for mappers. It allows for the rapid and efficient addition of detailed and multilingual information, making OSM more comprehensive and accessible to a global audience.

Now, let’s take a closer look at the step-by-step process that enabled this transformation:

Stage 1: Laying the Foundation - Source Selection and Contextualization

  1. Gathering Multilingual Insights: I embarked on a quest for knowledge, meticulously gathering information from a diverse range of publicly available sources:
  • Official Website: The theater’s official website https://gabriadze.com provided a wealth of information about the theater’s history, performances, and artistic philosophy.
  • Wikipedia: English, Russian and Georgian Wikipedia articles https://en.wikipedia.org/wiki/Rezo_Gabriadze_Marionette_Theater, https://ru.wikipedia.org/wiki/Тбилисский_государственный_театр_марионеток_имени_Резо_Габриадзе, https://ka.wikipedia.org/wiki/რევაზ_გაბრიაძე offered comprehensive overviews, historical context, and details about Rezo Gabriadze’s life and work.
  • Tourism Websites: Reputable tourism websites like https://www.georgianjournal.ge/arts-culture/34663-rezo-gabriadze-puppet-theatre-tbilisi.html and https://tbilisi-life.info/place/puppet_theatre/ provided valuable insights from a visitor’s perspective, highlighting the theater’s unique charm and appeal.
  • Blogs and Travelogues: Personal accounts and blog posts offered firsthand experiences and unique perspectives, enriching the narrative with personal anecdotes and observations.

Importantly, I ensured that all sources were freely accessible, avoiding materials behind paywalls or with restrictive usage rights.

  1. Guiding the AI with OSM Wisdom: To ensure my tags aligned seamlessly with OSM’s established conventions and tagging best practices, I provided the AI with relevant guidance from the OSM wiki. This included articles on:
  • Amenity Tagging: https://wiki.openstreetmap.org/wiki/Key:amenity
  • Tourism Tagging: https://wiki.openstreetmap.org/wiki/Key:tourism

This step ensured that the AI understood the specific language and structure of OSM tags, making the output more compatible with the platform’s requirements.

  1. Infusing Personal Knowledge: My deep familiarity with Tbilisi, gained from living here and exploring its hidden gems, allowed me to contribute valuable insights that might not be found in standard sources. I added my own knowledge about:
  • Historical Context: Its significance in the context of Georgian culture.
  • Architectural Nuances: The tower’s unique design, the probable materials used, based on my visit of these POIs.
  • Performance Details: The types of puppet shows, the target audience, and the schedule of performances.
  • Mapillary Insights: I analyzed Mapillary images to capture recent changes or details not mentioned in written sources, such as new signage, accessibility features.

This step ensured that the tags reflected a nuanced and firsthand understanding of the POIs, enriching them with details that go beyond standard descriptions.

  1. Extracting the Essence - Structured Summary Request: With the sources gathered and contextualized, I turned to the AI’s analytical prowess. I prompted Gemini AI to create a structured summary of the information, focusing on these key aspects:
  • History: Key dates, events, and individuals involved in the creation and development of the theater and clock tower.
  • Architecture: Architectural style, materials used, notable features, and any unique design elements.
  • Performances: Types of puppet shows, target audience, schedule, and any special events or performances.
  • Unique Features: Any distinctive characteristics that set these POIs apart, such as the hourly angel performance on the clock tower.
  • Local name: As a local, I know that this tower is called “The tower with an angel” by Georgians.

This structured summary served as a concise and organized foundation for the subsequent tag creation, ensuring that the AI focused on the most relevant information and avoided irrelevant tangents or extraneous details.

Stage 2: Crafting Precise Names - Name Tag Generation

  1. Derivation and Refinement: I prompted the AI to derive the POI’s real name, considering variations and naming conventions across the sources. The AI then generated name tags in multiple languages, adapting them to fit OSM’s character limitations.

My role was then to meticulously review and refine these tags, ensuring:

  1. Accuracy: Each name tag accurately reflected the POI’s official name in the respective language.
  2. Consistency: The naming convention was consistent across all languages, avoiding unnecessary variations or inconsistencies.
  3. Natural Language Flow: The tags read naturally and idiomatically in each language, avoiding awkward or literal translations.

This step involved cross-referencing the AI’s output with the original sources, consulting language dictionaries and resources, and, in some cases, seeking feedback from native speakers to ensure the highest level of accuracy and fluency.

Stage 3: Painting a Vivid Picture - Description Tag Creation and Translation

  1. Iterative Refinement: With the name tags finalized, I turned to the task of crafting a compelling and informative description tag. Gemini AI generated an initial English description based on the structured summary and the unique features I had identified.

I then embarked on a process of iterative refinement, carefully scrutinizing the AI’s output and engaging in a dialogue with the AI to improve its content and phrasing. This involved:

  1. Adding Context: Providing the AI with additional prompts to elaborate on specific aspects of the POIs, such as their historical significance or architectural details.
  2. Clarifying Ambiguities: Rephrasing sentences or adding clarifying details to ensure the description was clear and unambiguous. I’ve lowered the temperature to 0.6 to help AI be more accurate
  3. Enhancing Engagement: Using more evocative language and incorporating descriptive details to make the description more engaging and captivating for readers.

This iterative process, a true partnership between AI and human creativity, resulted in a description tag that was both informative and engaging, capturing the essence of the theater and clock tower.

  1. Expanding the Linguistic Palette - Multilingual Expansion: With the English description finalized, I instructed the AI to translate it into other languages, ensuring that the translated tags were not only accurate but also flowed naturally in each target language. This involved:
  • Translation Verification: I carefully reviewed the translations I understood (English, Russian, Georgian, and to some extent French) for accuracy and clarity, making adjustments.
  • Language Adaptation: I further adapted the translations for natural language flow in each target language, ensuring they sound natural to native speakers. I couldn’t consult with native speakers of other languages than English, Russian, and Georgian.

The culmination of this meticulous process is a set of comprehensive and multilingual tags that paint an engaging picture of the Rezo Gabriadze Marionette Theater and Clock Tower in OSM.

Copyright Analysis: Addressing Legal Concerns and Building a Strong Case for Fair Use

The use of AI in creative endeavors, particularly when utilizing copyrighted materials, raises novel legal questions. My AI-assisted tagging process, however, is carefully designed to comply with copyright law, particularly the principles of fair use enshrined in the U.S. Copyright Act (17 U.S.C. § 107) (U.S. Copyright Office, 1976), as well as similar legal doctrines recognized in other jurisdictions.

The following in-depth analysis demonstrates how my approach aligns with each of the four fair use factors, providing a robust legal foundation for my methods:

1. Purpose and Character of the Use:

  • Highly Transformative: At the heart of fair use lies the concept of transformation. My process transcends mere replication. The AI, under my guidance and informed by my knowledge, assists me in transforming factual information gleaned from various sources into succinct, informative OSM tags. This transformation serves a distinct purpose – furnishing location-based data for OSM users – and caters to a different audience than the original sources (World Intellectual Property Organization [WIPO], 2024). This transformative use finds support in landmark cases like Campbell v. Acuff-Rose Music, Inc. (510 U.S. 569 (1994)), which recognized parody as a transformative fair use.

Beyond simply summarizing, the AI aids me in re-contextualizing and repurposing factual data, creating a new type of content tailored specifically for OSM’s unique requirements. This process aligns with the spirit of fair use, which encourages the creation of new works that build upon existing knowledge without stifling innovation (U.S. Copyright Office, 2023a).

  • Public Benefit: My contributions directly benefit the public by enhancing OSM, a free and universally accessible map (OpenStreetMap Foundation, 2024). The multilingual tags, in particular, foster cross-cultural understanding and global collaboration, echoing the broader aims of open knowledge and universal accessibility championed by international organizations like WIPO (WIPO, 2024).

2. Nature of the Copyrighted Works:

  • Predominantly Factual: The sources I utilize are primarily factual, concentrating on historical and architectural facets of Georgian POIs. Copyright protection for factual works is inherently less robust than that for creative works, as facts themselves are not copyrightable (Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991)). This principle was solidified in the landmark case of Feist Publications, which held that compilations of facts are only eligible for copyright protection if they demonstrate a minimal degree of creativity in their selection and arrangement.

My process focuses on extracting these factual elements from diverse sources, further reinforcing the applicability of fair use.

3. Amount and Substantiality of the Portion Used:

  • Limited by Design: The inherent 255-character constraint on OSM tags intrinsically limits the amount of content I can extract from any individual source, ensuring that I am not utilizing substantial portions of the original works.

This limitation, coupled with my multi-step process of summarization, refinement, and translation, guarantees that the essence of the original works is distilled without appropriating substantial expressive content.

  • Summarization and Refinement: My meticulous multi-step process of summarizing, refining, and translating further ensures that I am not utilizing substantial portions of the original works. My focus is on distilling key facts and unique details, not on verbatim copying. This meticulous approach minimizes the amount of copyrighted material incorporated into the final tags, aligning with the fair use principle of using only as much as necessary to achieve the transformative purpose.

4. Effect on the Market:

  • No Harm, Potential Enhancement: My non-commercial intent, the public availability of the sources I use, and the non-competitive nature of OSM tags with the source materials all strongly suggest minimal, if any, market harm.

My tags are succinct, factual descriptions that do not serve as substitutes for the original works. They fulfill a distinct purpose and cater to a different audience. In fact, my work could potentially bolster the market for these sources by prompting OSM users to seek out more comprehensive information (U.S. Copyright Office, 2024b).

Addressing Additional Concerns: Transparency, Community, and AI Ethics

Beyond the four fair use factors, I embrace additional principles to ensure ethical and responsible AI use within the OSM community:

1. OSM License (ODbL): My contributions are released under the Open Data Commons Open Database License (ODbL), which permits the unrestricted use, adaptation, and dissemination of data (OpenStreetMap Foundation, 2024). This aligns with OSM’s collaborative, open-source ethos, guaranteeing that my contributions are freely accessible to all.

2. OSM Community Guidelines: My process adheres to OSM’s Automated Edits Code of Conduct, which stresses caution, community engagement, and adherence to tagging conventions (OpenStreetMap Wiki, 2024). I actively seek feedback from the community and believe that open dialogue is vital for establishing best practices for AI utilization within OSM.

3. Transparency and Attribution: I consistently identify myself as the contributor and attribute the AI tools employed, ensuring transparency and avoiding any misleading assertions of original authorship. This approach reinforces the human element in my process and acknowledges the collaborative nature of AI-assisted creation.

4. Independent Verification: I meticulously verify the accuracy of the tags in multiple languages, demonstrating a commitment to data integrity and minimizing the risk of perpetuating any potential biases inherent in AI models. This meticulous verification process ensures that the tags are reliable and trustworthy, enhancing the overall quality of OSM data.

5. Necessity of AI: The scale and multilingual nature of my project necessitate the use of AI. Manually achieving comparable results would be prohibitively time-consuming and resource-intensive, impeding the timely enrichment of OSM data. AI, in this context, is not a shortcut but an essential tool for enabling a level of contribution that would be impossible for a single individual to achieve manually.

6. AI Ethics and Bias Mitigation: I recognize the importance of addressing potential biases in AI systems, as highlighted by the U.S. Copyright Office in its recent Notice of Inquiry (U.S. Copyright Office, 2023b). My focus on factual information, multi-source verification, and human review helps mitigate this risk. Furthermore, OSM’s collaborative nature allows any contributor to edit and improve the tags, fostering a collective effort to guarantee accuracy and fairness.

Risk Analysis: Potential Challenges and Mitigation Strategies

While AI-assisted tagging offers significant benefits, it’s important to acknowledge potential risks and develop strategies for mitigation:

  1. Evolving Copyright Landscape: Copyright law surrounding AI-generated content is evolving, and future legal interpretations could impact the fair use analysis. Staying informed about legal developments and adapting my practices as needed is crucial. I will continue to monitor legal developments and engage with legal experts to ensure my approach remains compliant.
  2. AI Bias and Misinformation: AI models can perpetuate biases present in their training data, potentially leading to inaccurate or misleading information in the tags. My multi-source verification and human review processes help mitigate this risk. Additionally, engaging with the OSM community to develop guidelines for identifying and addressing bias in AI-generated content is crucial. I am committed to working with the community to develop best practices for ethical AI use in OSM.
  3. Over-Reliance on AI: While AI is a powerful tool, it’s essential to avoid over-reliance and maintain human oversight. My process emphasizes human judgment and critical evaluation at each stage, ensuring that the AI is used as a partner, not a replacement for human expertise.
  4. Data Privacy: AI training often involves vast datasets, raising concerns about data privacy. It’s crucial for AI providers to adhere to ethical and legal standards for data collection and usage. As an OSM contributor, I rely on the AI providers’ compliance with these standards, highlighting the need for greater transparency from AI companies regarding their data practices. I will continue to advocate for greater transparency from AI providers and support efforts to develop ethical guidelines for AI data collection.

Addressing Specific Concerns from the OSM Community

I want to directly address some of the specific concerns raised in the OSM community discussion:

  • Discouraging Human Contributions: My goal is not to replace human contributions but to enhance them. By using AI to handle tedious tasks, I free up time for myself and other mappers to focus on more complex and creative contributions. My tags are intended to complement, not replace, human-created descriptions.
  • Transparency of AI Training Data: I acknowledge the importance of transparency regarding AI training data. While I do not have specific knowledge of the exact datasets used to train the AI models I employ, I trust that the AI providers are adhering to legal and ethical standards in their data collection and training practices, as outlined in their terms of service. I support efforts within the AI and OSM communities to promote greater transparency from AI companies regarding their data practices.
  • Risk of AI Bias: I am committed to ensuring that my AI-generated tags are not biased or discriminatory. My focus on factual information, multi-source verification, and human review processes help mitigate this risk. I am also open to collaborating with the OSM community to develop guidelines for identifying and addressing potential biases in AI-generated content.

Conclusion

AI offers a transformative opportunity to enrich OSM, creating a more informative and accessible map for a global audience. My approach demonstrates that this can be done responsibly, with respect for copyright, transparency, a commitment to community collaboration, and a focus on mitigating potential risks. I welcome your feedback and look forward to continued dialogue as we navigate this evolving landscape together.

License notice:

AI-Assisted Tagging in OpenStreetMap: A Case for Responsible Innovation and Copyright Compliance © 2024 by David Osipov assisted my Gemini AI is licensed under Creative Commons Attribution 4.0 International

Frequently Asked Questions (FAQs) About AI-Assisted Tagging in OSM

Q1: What about copyright in the AI-generated text itself?

A1: The copyrightability of AI-generated content is a complex and evolving legal question, with different countries taking different approaches. In the United States, current copyright law does not recognize AI systems as authors (U.S. Copyright Office, 2023). My process involves substantial human input, including source selection, prompting, iterative refinement, and verification, which could potentially satisfy the “minimal degree of creativity” standard for human authorship in certain jurisdictions. However, to address this uncertainty and to reinforce my commitment to open knowledge, I explicitly dedicate any copyright I may have in the tags to the public domain under the Open Data Commons Open Database License (ODbL).

Q2: Does the UK’s copyright law for computer-generated works apply to your tags?

A2: The UK Copyright, Designs and Patents Act 1988 (CDPA) grants copyright protection to “computer-generated works” even without a human author. However, the UK law focuses on works created autonomously by AI, where human involvement is limited to setting initial parameters (UK Intellectual Property Office, 2024). My active involvement in refining and directing the AI output, incorporating my knowledge and judgment, may distinguish my process from the autonomous creation envisaged by the CDPA.

Q3: What about potential bias in the AI-generated tags?

A3: AI bias is a valid concern. I take steps to mitigate it by focusing on factual information, using multi-source verification, and conducting human review in the languages I understand (Georgian, English, and Russian). Additionally, OSM’s collaborative nature allows any contributor to edit and improve the tags, fostering a collective effort to address bias. I encourage other OSM contributors to review my tags and provide feedback.

Q4: Are there any specific legal challenges to using copyrighted materials to train AI models?

A4: The legality of using copyrighted materials to train AI models is an area of ongoing debate and litigation. Some argue that such use constitutes copyright infringement, while others contend that it falls under fair use or other exceptions. As a user of publicly available AI tools, I rely on the AI providers to ensure they are adhering to legal and ethical standards in their data collection and training practices, as outlined in their terms of service. I support efforts within the AI and OSM communities to promote greater transparency from AI companies regarding their data practices.

Q5: How do you ensure your AI-generated tags comply with OSM tagging conventions and best practices?

A5: I meticulously review the AI-generated tags to ensure they adhere to OSM’s established tagging conventions and best practices, as documented in the OSM wiki. I also cross-reference the tags with existing OSM data to ensure consistency and avoid duplication.

Q6: What steps do you take to ensure that your AI-generated tags are not offensive or inappropriate?

A6: I review all AI-generated tags for potentially offensive or inappropriate language and make adjustments as needed to ensure the tags are neutral, objective, and respectful. I welcome feedback from other OSM contributors on this issue.

Q7: How can other OSM contributors provide feedback on your AI-generated tags?

A7: You can provide feedback on my AI-generated tags through various channels:
The OSM forum: this thread
By contacting me directly: via Telegram (Telegram: Contact @david_osipov) or email (david.z.osipov@gmail.com)

I am always open to feedback and suggestions for improvement.

Q8: What specific AI tools do you use in your process?

A8: I utilize several general-purpose AI models, primarily Google’s Gemini AI and Meta’s Llama AI. I occasionally use Anthropic’s Claude and OpenAI’s ChatGPT as well. These tools are not specifically designed for OSM tag creation but are powerful LLMs with a wide range of applications, including text generation, translation, and summarization.

Q9: Why do you use multiple AI tools instead of just one?

A9: Each AI model has its strengths and weaknesses. By using multiple tools, I can leverage their different capabilities to achieve the best possible results. For example, Gemini AI excels at generating creative text formats, while Llama AI is particularly strong in translation. Using a combination of tools allows me to cross-check information and ensure a higher level of accuracy and quality in the final tags.

Q10: How do you decide which languages to translate the tags into?

A10: I prioritize languages based on the POI’s location and its potential relevance to different linguistic communities. For POIs in Georgia, I translate into Georgian, English, and Russian as these are the most commonly used languages by locals and visitors. I also consider translating into other languages spoken by significant tourist groups or diaspora communities.

Q11: Do you ever use AI to generate tags for POIs that already have descriptions?

A11: I primarily focus on adding tags to POIs that lack descriptions or have minimal information. If a POI already has a detailed description, I may use AI to translate it into other languages or to add specific details that are missing. However, I always respect the work of other OSM contributors and avoid overwriting or duplicating existing information.

Q12: What are the limitations of AI-generated tags?

A12: AI-generated content, while helpful, has limitations. AI models can sometimes produce inaccurate or nonsensical information, especially when dealing with complex or nuanced topics. My process involves careful review and verification to minimize these errors. Additionally, AI-generated tags, due to their concise nature, may not capture the full richness and complexity of a POI. They are intended to provide a brief overview and encourage users to seek out more detailed information from other sources.

Q13: Are you concerned about the potential for your AI-assisted tagging to be used for malicious purposes?

A13: Like any tool, AI can be misused. However, my contributions to OSM are transparent and publicly accessible. Any malicious edits would be quickly identified and reverted by the OSM community. The open and collaborative nature of OSM serves as a safeguard against such misuse.

Q14: What are your thoughts on the future of AI in OSM?

A15: I believe AI has the potential to revolutionize OSM, making it even more comprehensive, accurate, and accessible. However, it’s crucial to use AI responsibly and ethically, with a focus on transparency, accountability, and community collaboration.

References:

Legislation:

• Directive (EU) 2019/790 of the European Parliament and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC, 2019 O.J. (L 130).
• REGULATION (EU) 2024/… OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of … laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act). (2024, April 19). European Parliament. Retrieved September 30, 2024, from https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CONSIL%3APE_24_2024_INIT

Court Decisions and Administrative Rulings:

• U.S. Copyright Office Review Board. (2022, February 14). Decision Affirming Refusal of Registration of A Recent Entrance to Paradise [Memorandum Opinion]. https://www.copyright.gov/rulings-filings/review-board/docs/a-recent-entrance-to-paradise.pdf
• U.S. Copyright Office. (2023, February 21). Cancellation Decision re: Zarya of the Dawn (VAu001480196). https://www.copyright.gov/docs/zarya-of-the-dawn.pdf
• Wilson, S. V. (2023, April 20). Guidance Letter to The MLC on AI-Created Works [Letter]. U.S. Copyright Office. Retrieved September 30, 2024, from https://www.copyright.gov/ai/USCO-Guidance-Letter-to-The-MLC-Letter-on-AI-Created-Works.pdf

Scholarly Works:

• Guadamuz, A. (2023). The monkey selfie revisited: Copyright in the age of artificial intelligence. Journal of Intellectual Property Law & Practice, 18(11), 947–955. https://doi.org/10.1093/jiplp/jpad121
• Sobel, B. L. W. (2024). Elements of Style: Copyright, Similarity, and Generative AI. Harvard Journal of Law & Technology, 38(1). https://www.bensobel.org/files/articles/Sobel_Elements-of-Style_Public-Draft-May-18-2024.pdf

Reports and Toolkits:

• 2023 State AI Legislation Summary. (2023). BSA | THE SOFTWARE ALLIANCE. Retrieved September 30, 2024, from https://www.bsa.org/files/policy-filings/09222023statelegai.pdf
• United States Copyright Office. (2024, July). Copyright and Artificial Intelligence. Part 1: Digital Replicas. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-1-Digital-Replicas-Report.pdf
• World Intellectual Property Organization (WIPO). (2024). Getting the Innovation Ecosystem Ready for AI: An IP policy toolkit. Geneva: WIPO. DOI: 10.34667/tind.48978

Websites:

• ChatGPT General FAQ. (n.d.). OpenAI. Retrieved September 30, 2024, from https://help.openai.com/en/articles/6783457-chatgpt-general-faq
• Commons:AI-generated media. (n.d.). Wikimedia Commons. Retrieved September 30, 2024, from https://commons.wikimedia.org/wiki/Commons:AI-generated_media
• Copyright Office Launches New Artificial Intelligence Initiative. (2023, March 16). U.S. Copyright Office. Retrieved September 30, 2024, from https://www.copyright.gov/newsnet/2023/1004.html
• Is AI-Generated Output Protected by Copyright? (n.d.). Copyright Alliance. Retrieved September 30, 2024, from https://copyrightalliance.org/faqs/is-ai-generated-protected-by-copyright/
• Copyright and License (n.d.). In OpenStreetMap website. Retrieved September 30, 2024, from https://www.openstreetmap.org/copyright
• Open Data Commons Open Database License (ODbL) (n.d.). Retrieved September 30, 2024, from https://opendatacommons.org/licenses/odbl/1-0/
• Lawton, G. (2024, June 5). Is AI-generated content copyrighted? TechTarget. Retrieved September 30, 2024, from https://www.techtarget.com/searchcontentmanagement/answer/Is-AI-generated-content-copyrighted

Blog Posts and Online Articles:

• Copyrightability of AI-Generated Works. (n.d.). Copyright Alliance. Retrieved September 30, 2024, from https://copyrightalliance.org/copyright-ai-generated-works/
• International Copyright Issues and Artificial Intelligence. (2023, July 26). U.S. Copyright Office. Retrieved September 30, 2024, from https://www.copyright.gov/events/international-ai-copyright-webinar/International-Copyright-Issues-and-Artificial-Intelligence.pdf
• Islam, A. (2023, March 21). A History of Generative AI: From GAN to GPT-4. Marktechpost. Retrieved September 30, 2024, from https://www.marktechpost.com/2023/03/21/a-history-of-generative-ai-from-gan-to-gpt-4/
• Martineau, K. (2023, April 20). What is generative AI? IBM Research Blog. Retrieved September 30, 2024, from https://research.ibm.com/blog/what-is-generative-AI
• Spring 2023 AI Listening Sessions. (n.d.). U.S. Copyright Office. Retrieved September 30, 2024, from https://www.copyright.gov/ai/listening-sessions.html
• Three Takeaways When Registering Your Copyright in an AI-Assisted Work. (n.d.). Copyright Alliance. Retrieved September 30, 2024, from https://copyrightalliance.org/ai-assisted-work-copyright-registration/
• Who Owns the Copyright to AI-Generated Works? (n.d.). Copyright Alliance. Retrieved September 30, 2024, from https://copyrightalliance.org/faqs/artificial-intelligence-copyright-ownership

Wikipedia:

• Artificial intelligence. (2024, September 26). In Wikipedia. Retrieved September 30, 2024, from https://en.wikipedia.org/wiki/Artificial_intelligence

OSM Wiki:

• Amenity Tagging. (n.d.). In OpenStreetMap Wiki. Retrieved September 30, 2024, from https://wiki.openstreetmap.org/wiki/Key:amenity
• Automated Edits code of conduct. (n.d.). In OpenStreetMap Wiki. Retrieved September 30, 2024, from https://wiki.openstreetmap.org/wiki/Automated_Edits_code_of_conduct
• Facebook AI-Assisted Road Tracing. (n.d.). In OpenStreetMap Wiki. Retrieved September 30, 2024, from https://wiki.openstreetmap.org/wiki/Facebook_AI-Assisted_Road_Tracing
• Machine learning. (n.d.). In OpenStreetMap Wiki. Retrieved September 30, 2024, from https://wiki.openstreetmap.org/wiki/Machine_learning
• Tourism Tagging. (n.d.). In OpenStreetMap Wiki. Retrieved September 30, 2024, from https://wiki.openstreetmap.org/wiki/Key:tourism
• WikiProject Georgia Abkhazia South-Ossetia. (n.d.). In OpenStreetMap Wiki. Retrieved September 30, 2024, from https://wiki.openstreetmap.org/wiki/WikiProject_Georgia_Abkhazia_South-Ossetia]

Please no. The internet is slowly filling up with derivative LLM-generated crap. Projects like OpenStreetMap are mostly free from this, with mappers contributing their knowledge and being generally very precise with names and content.

Now you are suggesting having LLM’s generate:

Description Tag: A captivating narrative, translated into 20 languages […]

How does a mapper know these descriptions in twenty languages are accurate? This approach will just lead to loads of data without any clear provenance and of dubious accuracy, and if I then ask the mapper contributing it where they got, for example, the Dutch description, I get “I had ChatGPT (or whatever) generate it!”.

Besides, users don’t need ‘a captivating narrative’. A link to the theatre’s website will do, thank you.

Besides, LLMs need data sources untouched by LLMs to train on. This includes sites like Wikipedia and OpenStreetMap. Adding AI into the mix for textual descriptions and names completely breaks that too, and eventually LLMs won’t even be able to work on OSM without regurgitating false or outdated data.

And that is another fundamental problem with LLMs. Getting names right is hard (exonyms in particular) and using LLMs will inevitably lead to outdated or incorrect names being used. OpenStreetMap often is the one place where the names are correct and up-to-date (because names do change); I would like to keep it that way.

By the way, this is not a desirable approach. You can’t just invent a translation for a name yourself (or have an LLM do that for you). OpenStreetMap is a map, not a translator. It is perfectly normal for a theatre in Tbilisi not to have a Dutch or Japanese or Frisian name. Don’t go making these up based on LLM output. Exonyms (foreign names) have to actually be in use before you can put them on OSM.

32 Likes

I didn’t read all that, but my concern isn’t much about copyright, but other things, for example: Node History: ‪თაბორის რეაბილიტაციის ცენტრი‬ (‪12185682258‬) | OpenStreetMap

Here you added an italian name to a place that doesn’t have an italian name.

In other cases you add descriptions in languages you don’t know (I guess, since they go from italian to hebrew to korean to arab). How can you know those translations are accurate? “Other mappers will fix it” is not a reply I appreciate much.

Here you added the architect name in 11 languages:


Wouldn’t be better to add 1 architect:wikidata=* value?

In general, this approach seems to clutter elements with lot of tags that seem to mimic what is already present in Wikipedia/Wikidata ecc. informations that can already be retrieved via wikipedia=* and wikidata=* tags.

17 Likes

I supply the AI with up-to-date information about the POI, it doesn’t use its own knowledge about the POI. AIs don’t have knowledge about Georgian POIs in general.

OSM Georgian community established several rules on name creation, which I also supply to the AI in terms of doubt. Still, I see having at least smt is better than nothing.

That’s LLMs creators responsibility, not ours.

The narrative in sense of telling a story about the POI in just 255 symbols.

Pretty much 40-60% of POI in the capital of Georgia Tbilisi have outdated POIs info, including names and other tags.

You are correct that this specific POI might not have a widely established Italian name, but we can’t be sure whether it 100% doesn’t have an Italian name. My intention was to provide a translated name for users who might be searching for the POI in Italian. However, I acknowledge that this approach might not be appropriate for all cases, and I am open to refining my process based on community feedback.

I understand your concern about the accuracy of translations in languages I am not proficient in. While I cannot guarantee 100% accuracy in every translation, I take several steps to minimize errors:

  1. Source Selection: I rely on reputable and reliable sources for information, including official websites, Wikipedia articles in multiple languages, and established tourism platforms.

  2. AI as a Tool: I use AI as a tool to assist with translation, not as a replacement for human judgment. I carefully review the AI’s output, comparing it with the source material and making adjustments as needed.

  3. Community Verification: I believe in the power of community collaboration. By making my edits public, I allow other mappers who are fluent in the respective languages to review and correct any errors. This collaborative approach helps ensure the accuracy and quality of the data.

Thank you for stressing it out. That was my mistakes and I’ve addressed it several minutes ago.

Don’t make up names in other languages!

There is a wine bar in Tbilisi called ზეღვინო (so name=ზეღვინო), which also styles itself as ‘ZeGvino’ on their own website (so alt_name=ZeGvino and/or name:ka-Latn=ZeGvino).

Now what are you going to call this in Dutch? If I read their about there is some explanation about the naming, indicating that it means something like ‘super wine’. So name:nl=Superwijn and name:it=Ottimo vino? Of course not. If some Dutch blogger was writing about this bar, they’d recommend visiting (or avoiding) ‘ZeGvino’, not some made up literal translation.

It is really valuable for data consumers and mappers to know for sure that some entity does not have a name in Italian (or whatever language), and users don’t want made up translations if something does not have a name in their language. You’re just polluting the database this way.

(Edit: although name:de=Überwein does sound really cool…)

11 Likes

Lol! :laughing: So you’re telling me that among the residents of Tbilisi, Georgia there are significant native speaking populations of the following 20+ languages, and that they all have different local names for this clock tower?

loc_name	ანგელოზის კოშკი
loc_name:ar	البرج ذو الملاك
loc_name:az	Mələkli Qüllə
loc_name:de	Der Turm mit dem Engel
loc_name:en	The Tower with the Angel
loc_name:es	La Torre del Ángel
loc_name:fa	برج با فرشته
loc_name:fr	La Tour de l'Ange
loc_name:he	המגדל עם המלאך
loc_name:hi	फ़रिश्ता वाला टावर
loc_name:hy	Հրեշտակով աշտարակ
loc_name:it	La Torre con l'Angelo
loc_name:ja	天使の塔
loc_name:ka	ანგელოზის კოშკი
loc_name:ko	천사의 탑
loc_name:nl	De Toren met de Engel
loc_name:pl	Wieża z Aniołem
loc_name:pt	A Torre com o Anjo
loc_name:ru	Башня с ангелом
loc_name:tr	Melek Kulesi
loc_name:uk	Вежа з ангелом
loc_name:zh-CN	带天使的塔
loc_name:zh-TW	帶天使的塔

Yes lets do our best to keep OSM from being overrun by AI slop like the rest of the internet.

14 Likes

I think it would be more helpful to provide a transliteration from the Georgian script to the Latin alphabet by using int_name[:<xx>], since that is undoubtably useful to anyone who doesn’t read Georgian.

I think you should just link wikipedia pages instead of using AI for descriptions. Organic Maps will show the start of the article for any POIs with a link to wikipedia, saving you both the pain of creating and the pain of translating descriptions.

If you like doing that, I think your time would be better spent making wikipedia/wikidata (maybe even wikivoyage) articles for POIs that are missing them in the languages you do know. You can then link to those articles from OSM.

5 Likes

I hope it helps:
name=ზეღვინო
name:en=ZeGvino
name:ka=ზეღვინო
name:ru=Зе Вино
name:fr=ZeGvino
name:de=ZeGvino
name:es=ZeGvino
name:it=ZeGvino
name:pt=ZeGvino
name:zh-CN=ZeGvino
name:zh-TW=ZeGvino
name:ja=ZeGvino
name:ko=ZeGvino
name:ar=زي غيفينو
name:hi=ज़ेग्विनो
name:tr=ZeGvino
name:hy=ԶեԳվինո
name:az=ZeGvino
name:fa=زه گویینو
name:uk=ЗеГвіно
name:pl=ZeGvino
name:nl=ZeGvino
name:he=זגווינו

Last time someone did something similar (automated translations and transliterations), they got blocked, a DWG member wrote:

However, generally speaking, transliteration and translation should be avoided. If there really is a sign in a certain language, or some other licence-compatible way of verifying that name, then it makes sense for a place to have multiple languages. If there isn’t, it doesn’t.

I don’t know if something changed since then, if not, it seems to apply here, since is just transliterations? I tag @SomeoneElse since he wrote the quoted message and could go in further detail as he handled the avinet_ua situation if I remember correctly.

  1. Why are you LoLing? I guess we should leave emotions and conduct a civilized discussion with arguments and counterarguments.
  2. My goal in adding multiple loc_name tags is to enhance the accessibility and discoverability of POIs for users who speak different languages. While it’s true that Tbilisi might not have significant native speaking populations for all 20+ languages I included, these languages represent a considerable portion of tourists and international visitors.

Providing translations of the local name (“ანგელოზის კოშკი” - The Tower with the Angel) allows users searching in their native language to find the POI more easily. This approach aligns with OSM’s goal of creating a globally accessible map.

That list now states that there is a Dutch name for that bar. This does not seem true, so your data is lying.

By the way, name:zh-CN, zh-TW and ja wouldn’t be written using Latin characters. Its Japanese name would be something like ザ・ガビノ, but again, that would only be the case if this bar was moderately famous in Japan, and people wrote about it.

Don’t make up translations.

9 Likes

You (and the LLM) do not know if the Dutch would even call it ‘De Toren met de Engel’. We’d probably call it ‘Engelentoren’ instead. But as it stands, we don’t really call it anything except in a descriptive sense (e.g., “The Georgians also call this ‘the tower with the angel’.”). That is just not enough for a loc_name:nl.

Now if someone blew up that tower and sparked a war (or a ‘special operation’) by doing so, and suddenly media around the world where writing about this, then it is quite possible that it would acquire a Dutch (nick)name. That might be ‘Engelentoren’, but it could also be ‘Angelozis koshki’ borrowing from Georgian. It really depends, and is by its very nature hard to predict.

Fortunately, that tower seems not to have been hit by any stray Shaded-drone yet, so it really doesn’t have a loc_name:nl or alt_name:nl.

8 Likes

Do not add translations - machine-generated or otherwise - of names. Never do that. The “Pont Neuf” in Paris is not called “Neue Brücke” in German! Only add names in other languages where the place actually uses that name for themselves (or is known under that name). Translations that are not tied to local knowledge add zero value - anyone who wishes a name (or a thousand) translated using a generic translation method can do that in one click in their browser.

Never add a “captivating narrative” to anything, it is not verifiable and will be deleted.

Also, if you try to keep your forum postings to a length that makes them readable in under five minutes you will find that more people read them.

26 Likes

Because the idea that a local name would exist in 20+ languages is ridiculous and funny! loc_name description from the wiki:

the name of a feature as it is known locally, but only where this is deemed to be too much of a slang name or otherwise unofficial-sounding.

So by putting 20+ loc_name:lang tags on a feature you are saying that there are 20+ native speaking language communities in the area with a specific slang/unnofficial term for the feature.

If the name is something tourists commonly use, then it’s not a loc_name. The whole point of that tag is for names that tourists would be less likely to encounter unless speaking to a local.

Anyway, this is the sort of mistake that can easily happen when human thought is replaced with the non-thinking, statistical analysis of an AI/LLM.

2 Likes

I appreciate the feedback and concerns raised regarding the use of AI-assisted tagging in OSM. I understand the importance of maintaining data quality and ensuring that all additions to the database are verified and reflect real-world usage.

I acknowledge the points made regarding the use of name and loc_name tags, and I am committed to refining my approach to align with community guidelines and best practices. Moving forward, I will prioritize accuracy and ensure that translated names are only added when they are verifiably used by the relevant communities.

However, I also want to reiterate the significant benefits that AI-assisted translation can offer for tourists visiting Georgia and other regions with limited multilingual support in OSM:

Multilingualism is Essential for Tourist Accessibility:

Providing information in multiple languages is crucial for making OSM accessible to a global audience. Tourists who don’t speak Georgian, English or other local languages often face difficulties navigating and understanding the map.

By utilizing AI to translate descriptions and potentially names (with careful consideration and verification), we can significantly improve the experience for these users. This allows them to:

  • Discover relevant POIs: Tourists can search for places in their native language, even if they don’t know the local names. This enhances discoverability and helps them find places that might be of interest to them.

  • Understand and engage with the map: Translated descriptions allow tourists to learn more about the places they are visiting, gaining a deeper understanding of the cultural and historical significance of Georgian POIs. This enriches their travel experience and fosters cross-cultural understanding.

  • Navigate with greater confidence: Having access to information in their native language can help tourists feel more comfortable and confident exploring a new area, leading to a more enjoyable and fulfilling travel experience.

For Georgia really all that is needed is information in English and Russian. Most tourists visiting without a guide will have a decent proficiency of English (or Russian in this case obviously). Any effort to aid multilingualism for Georgia should focus on making sure Georgian names have a Latin transliteration as well (e.g., by providing name:ka-Latn). That probably does not require any LLM though.

Tourists who don’t speak English or Russian tend to visit with a tour operator who will handle most of this for them. Think busloads of Chinese or an organized cultural vacation for a group of Germans. These groups will just follow their guide, possibly augmented by a guidebook in their own language.

And of course, anyone can use automated translation tools if they want to. There is no need to provide automated translations on OpenStreetMap.

If you want to aid (independent) tourists in Georgia, make sure that points of interest have their websites linked and are up-to-date.

That is a large assumption. Most tourists will either get a guidebook in their own language, go with a guided group, or use English. It doesn’t make much sense to add all sorts of translated descriptions (and certainly not names). Besides, for any somewhat famous point of interest, Wikipedia will have an article, and if that article was translated, it will be available to the user as long as there is a wikipedia-tag.

6 Likes

Hi,

This was too long, so I didn’t read it.

In the future, when using AI to generate posts for the OSM forum, please ask the chatbot to be more succinct so that people will actually read it.

I am opposed to whatever is being proposed here because it’s too long to comprehend.

29 Likes