Gotheborg Forum AI Policy

Log Out | Topics | Search Moderators | Register | Edit Profile

Gotheborg Forum AI Policy

Why §6 has become necessary

This Forum is a long-term effort to gather, preserve, and share reliable, experience-based knowledge about Chinese and Japanese ceramics. Accuracy, clarity, and source integrity are critical to its value. With the rapid spread of Large Language Models (LLMs) such as ChatGPT, Gemini, Claude, and others, we face a new kind of challenge: text that appears confident and scholarly but may be partially or entirely fabricated references.

AI tools generate plausible-sounding responses, often filled with false names, dates, references, and conclusions. If such material becomes mixed-up with our archive, it risks contaminating our archive forever - that we have spent decades building - and may permanently mislead current or future collectors, scholars, and students.

For this reason, we have added a paragraph 6. - Use of Artificial Intelligence (AI) - to strictly prohibit the posting of AI-generated content on the Forum unless fully understood and fact-checked by the user.


How Large Language Models (LLMs) Operate

LLMs do not know anything. They generate text by calculating the statistical likelihood of words or phrases based on vast sets of human texts. They are trained to continue a pattern, not to verify a fact. Engineers think this is fun because they can run enormous data farms, but they don't deliver truths, only likelihoods.

This method results in what is commonly, politely called hallucinations - the confident generation of statements, sources, or facts that are completely fabricated.

Examples of how hallucinations emerge:

- A prompt asks for a scholarly reference; the AI fabricates a convincing title and author.

- A user requests historical background; the AI merges different time periods into a single narrative based on keywords.

- An identification of a porcelain mark is requested; the AI finds similar based on observed details like square and red, of which there are tens of thousands.

The model does not know it is wrong because it does not know anything. It only tries to deliver what satisfies the given parameters.


Examples of Fabricated References and Their Impact

Here are a few common forms of fabricated or misleading content generated by AI:

1. Nonexistent Books or Authors

"Ming Porcelain: Maritime Trade and Designs by Jonathan Keates (Oxford University Press, 1997)..."
- This book does not exist. Neither does the author (in this context). The publisher may be real, but the combination is invented.

2. Misdated Dynasties, places, relations or chronologies

"Famille Rose porcelain became popular in the Song dynasty."
- Famille Rose appeared in the early 18th century (Qing dynasty), not the Song. Such anachronisms are common in AI responses. It depends on what textual sources that has been included.

3. Merged or Corrupted Historical Narratives

- Combining trade events from different centuries, or mixing Japanese and Chinese terminology without regard to origin and that might appear as referring to the same object.

4. Fabricated Museum or Auction References

"This piece is identical to one sold by the Metropolitan Museum of Art in their 2015 'Jingdezhen Revival' catalogue."

- Neither the catalogue nor the sale may exist but the model can combine credible-sounding words to invent it.

5. Self-Referencing Data Loops

AI models trained on public web data may already now quote their own previous outputs or those of other bots, multiplying falsehoods. There are even example on that false data have been willfully manufactured and fed to the AIs to promote false narratives. This can today be produced by the millions without much cost or efforts. Same goes for text bots that interacts with humans fully automatic under false or nonexistent personas.

6. Visual mismatch of similar looking objects

It is very difficult even for a human being to be really certain about what a specific object is, but to an AI that has never handled anything IRL and lack tactile experience, it is almost impossible. Mistakes are impossible to avoid and more rule than exceptions.



Why It Matters for This Forum

The damage done by AI-generated content can rapidly wreck havoc with any carefully collected real data. One single erroneous name, location, attribution or date, can completely ruin the data about something by the bare mentioning of a "new discovery". The effect is similar to that of planting unrelated items to an archeological dig. The problem:

- Credibility erosion: Over time, false information as well as fake items embedded in archives becomes over time indistinguishable from genuine scholarship.

- Misleading new collectors: Beginners may base their understanding on fabricated material, assuming its accuracy, wrecking the entire scholarship and discrediting real scholars.

- Loss of trust: Members rely on the Gotheborg Forum for facts grounded in real-world experience, not synthetic approximations.

Unlike social media platforms where speed and engagement dominate, our ambition has always been to build something that lasts, an archive of human knowledge in a narrow and fairly specialized field.


Permitted Use of AI Tools

AI is not a personal assistance or a goto authority that knows everything, but an helpful tool, especially for:

- Spell-checking and proofreading

- Structuring texts

- Suggesting terminology

- Framing early-stage research questions

To protect what we have accomplished with so much efforts, our posts must always be based on our own knowledge, based on our own understanding. Any references must be real, verifiable, and checked, and if not, very clearly stated as such.


Suggested Reading and References

This is certainly going to change since AI is a rapidly developing field, but to better understand how LLMs work and why they can be dangerous when used carelessly we for the moment recommend the following articles and studies. This is certainly going to age bad.

- On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Emily M. Bender, Timnit Gebru, et al. (2021)

https://dl.acm.org/doi/10.1145/3442188.3445922

- Hallucination in Natural Language Generation, Cornell
https://arxiv.org/search/?query=hallucinations+in+ai+&searchtype=all&source=header

- Language Models are Few-Shot Learners Brown et al., OpenAI, 2020
https://arxiv.org/abs/2005.14165

- AI and the Future of Knowledge Production, Nature editorial, 2023
https://www.nature.com/search?q=AI+and+the+Future+of+Knowledge+Production

- OpenAI Help: ChatGPT - Release Notes, A changelog of the latest updates for ChatGPT
https://help.openai.com/en/articles/6825453




Checklist for Identifying Undisclosed AI-Generated Content

1. The use of AI tools is allowed - and even encouraged - under the conditions outlined above under "Permitted Use of AI Tools". This checklist is provided to assist moderators and members in identifying texts that may have been generated by an AI system and posted without clearly saying so. The reason this is brought up in such detail is that with the very fast acceptance of AI as part of the our daily life, it has become urgent that more than two decades of human efforts is not mixed up with computer generated misinformation. There is so far still real book available and peer-reviewed research to rely on, but for the vast majority of collectors, the dominating route to 'something', will be the fastest.


2. Posts that raise concern often exhibit one or several of the following characteristics:

a. Overly polished but vague language. The response sounds smooth and well-written but lacks precise, field-specific terminology or observable details.

b. Confident tone but with factual errors. Assertions are presented with authority, but key dates, names, dynastic periods, or stylistic terms are misused or conflated.

c. Fabricated sources. References may include non-existent books, incorrect author names, or imaginary institutions. These will often sound plausible but cannot be verified.

d. Lack of personal voice or perspective. Posts are sterile, offering no signs of personal experience, handling of objects, or lived knowledge. No first person "I" perspective or evidence of personal engagement with the material.

e. Inaccurate use of terms. Chinese or Japanese terminology is awkwardly or incorrectly used, or Romanizations are mixed and inconsistent.

f. Unnatural synonym use or padded phrasing. The text contains filler or unnatural word choices, seemingly chosen to sound formal or academic rather than to inform.

g. Adaptive but inconsistent logic. The answer may shift tone or stance to match the question without holding to a coherent internal line of reasoning.

h. Suspicious timing or frequency. A user (often new) may post unusually fast, frequent, or wide-ranging replies that exceed typical knowledge depth.


3. If a post appears to meet several of these criteria, moderators and senior members are encouraged to:

a. Cross-check references for verifiability.

b. Request clarification from the poster about source, method or own experience.

c. Just ask, if AI tools were used.

d. Off list, ask a moderator to take a look if concerns persist.


4. Members are reminded that while AI may be used - or even encouraged - for private research and editing, any post submitted to the Forum must reflect their own verified knowledge and understanding. Misuse may lead to moderation such as deletion of posts or in severe cases; loss of membership.



Final Notes

We as Discussion Board and Forum, will continues to support research, innovation and exploration, development of fields and integration of new findings, but not at the expense of truth and reality. Our AI policy is not anti-technology. It is pro-responsibility, trying to think a little bit ahead of the crowd.

AI is welcome as a private tool. It is not welcome as a public author of unverified claims under your name.

So, let us build a standard others can follow.

Jan-Erik Nilsson
Gotheborg Discussion Board Admin
Sept 2025