Guidelines for protecting  AI-generated content in Taiwan | Lee and Li

Newsrooms, OTT marketers and social media teams are embracing the use of generative tools to create headlines, summaries and marketing slogans in mass quantity. Taiwan’s existing copyright rules currently define functional limits to protection, aiming to protect human expression without covering text generated purely by AI.

Accordingly, media companies seeking to understand how much of their text is protected must determine which portion has been generated by human effort, and whether and how such input can be proven.

Defining protectable content

TsungYuan Shen
Associate Partner
Lee and Li
Hsinchu
Tel: +886 2 2763 8000 (ext.2539)
Email: tsungyuanshen@leeandli.com

Taiwan’s Copyright Act only protects expression, not ideas, methods, concepts or style. Edited collections formed by creative selection and arrangement qualify as compilation works. Default authorship and ownership rules still focus on work made by human creators, including works made in the course of employment or under commission.

Considered together, these provisions protect human contributions of media, even where AI provided the basic outline or structure of the text.

The Taiwan Intellectual Property Office (TIPO) has held that output produced entirely by AI without human creative input (for example, an image generated by a tool prompted to match the “style” of a sample image) is not protected under the Copyright Act.

In contrast, when AI is used as a tool – with humans contributing creative prompting, post-generation editing or creative selection and arrangement – protection may apply to human-authored portions of the finished product.

When AI-generated text includes previously existing work – such as quotes, background passages or internal archives – the four-factor test under article 65 of the Copyright Act should be applied to determine whether exploiting it complies with reasonable use:

1. 1. Purposes and nature of exploitation, including whether it is of a commercial nature or for non-profit educational purposes;
   2. Nature of the work;
   3. Amount and substantiality of the portion exploited in relation to the work as a whole; and
   4. Effect of the exploitation on the work’s current and potential market value.

Many studies examining article 65 and fair-use boundaries have been conducted. Media company legal teams may refer to such studies to formulate workable checklists.

Protectable v unprotectable

Protected Copy. In a typical copy workflow, text produced by the fol-
lowing steps may be protected because they reflect human expression or
creative arrangement:

1. 1. Rewriting AI drafts into publishable copy: tightening leads, recasting verbs, changing tone, and creating original headline-deck pairs;
   2. Restructuring paragraphs and narrative arc: selecting which facts to foreground, how to pace the piece, and where to insert context boxes; and
   3. Curating multiple generations into a final package: picking and sequencing slogans, assembling a “best of” synopsis, or building a side-by-side comparison. These are common selection/arrangement steps that fit the definition of compilation work.

Unprotected copy. The following scenarios typically do not create copyrightable content:

1. 1. Text generated via a “one-and-done” prompt with only superficial edits;
   2. Text generated by a prompt to follow a particular “style” (e.g. “make ABC text sound like X”) without distinct human input of wording or structure; and
   3. Fully automated translation, summarising or paraphrasing that lacks discernible human expression.

The above-mentioned scenarios generally fall outside the category of protected content because local practice protects concrete, human-authored expression, rather than abstract “style”.

In other words, style imitation or near-automatic outputs do not by themselves create protectable content. As the TIPO has stated, protection applies only when the final work bears a clear human imprint in wording, structure or editorial choices.

Grey areas. Short, functional phrases rarely qualify as protectable expressions, regardless of authorship. Likewise, AI-expanded lists of facts or programme synopses may be too thin to meet originality thresholds, unless the editor’s word choice and arrangement show individual expression.

Establishing human contribution

Rachel Chen
Senior Attorney
Lee and Li
Hsinchu
Tel: +886 3579 9911 (ext. 3206)
Email: rachelchen@leeandli.com

These suggested steps may increase the likelihood of protection of copy.

1. 1. Version trail and edit logs. The human contribution to copy text should be made visible and explainable. A record of prompts used, intermediate output generated, tracked edits and short notes capturing key editorial decisions should be retained. The goal is to create a record of the path of human input, including who rewrote which sentences, who changed what structure, and when such inputs were made. This supports authorship claims in the protectable portions of the copy and aligns with the TIPO’s philosophy of protecting human-generated content.
   2. Role separation and contract language. In documents intended to provide internal guidance, AI-drafted material should be distinguished from human-authored text, and rights should be asserted only in the latter.

Contracts signed with vendors and other service providers should include the following:

1. 1. 1. 1. source and rights warranties for any material used to condition or inform outputs;
         2. process transparency, with delivery of prompts and a brief description of human edits; and
         3. a book of clearly defined steps to be taken in case of dispute, including notification, removal of online publication, revision, and cost allocation.

1. 3. Distinguishing expression from style. Prior to publication, several questions should be considered: Does the piece show the team’s distinct wording and structure? Can specific human rewrites and restructuring be identified? If the “style veneer” is removed, are portions of text with human expression still clearly identifiable?

If answers to most of these questions are negative, the text should undergo substantive rewriting and narrative work. These tests should be integrated into the workflow to reduce post-publication disputes.

1. 4. Fair use worksheet for inputs. A short attribution is recommended when AI prompts include lengthy third-party text, or when comparison passages are published. A one-page form should be created for common work scenarios for editors including, for example, purpose and nature of the use, nature of the original work, amount and substantiality of the portion used, and effect on the potential market.

The goal is not to slow publication, but rather to document editorial purpose and restraint in use, leaving a paper trail that explains judgement calls.

1. 5. Labelling and archive discipline. Where practical, the phrase “AI-assisted draft; human-edited” should be noted in internal records. An index of final human-edited text and non-protectable AI raw outputs should be maintained. Such separation is not merely cosmetic; rather, it reflects the scope of rights the organisation can assert and license. Such separation also accelerates responses to takedown requests or client audits. In addition, retention periods should be set and access rules established. Details on major campaigns should be retained for three to five years. Periodic spot checks should be performed to confirm that labels, attachments and logs are complete.
   6. Training and escalation. Categories of issues that frequently escalate should be identified and highlighted, such as thin expression cases that may not be protectable, fair use close calls for substantial third-party excerpts, and sources whose contracts or terms of service restrict reuse. All escalations should go through a ticket or dedicated channel with timestamps, link to drafts and attachments (including prompts, excerpts and screenshots) to enable reviewers to decide quickly.

In addition, regular scenario drills should use sample drafts to practise worksheet, publish/hold decisions and revert cycles.

Looking ahead

The general objectives of Taiwan’s regulators are likely to remain focused on protection of human expression, without protection for style or fully automated output. What will evolve is the operational layer around that core.

More practical guidance can be expected from the authorities, as well as increasing practitioner literature on newsroom and platform workflows, and more internal guidance across the TMT sector on documentation of human authorial control.

Courts and regulations may clarify policy in grey areas, but the recommended approach will likely remain unchanged. Workflows should therefore be designed to ensure that human input is both substantive and visible.

Teams that build well-formulated proof-of-authorship habits will move faster through review, face fewer take-down fights, and approach licensing with a clearer defensible scope of rights.

LEE AND LI, ATTORNEYS-AT-LAW (HSINCHU OFFICE)
Address: 5F, Science Park Life Hub, No.1,
Industry E. 2nd Rd., Hsinchu Science Park,
Hsinchu 30075, Taiwan, R.O.C.
Tel: +886 3579 9911
Email: attorneys@leeandli.com
www.leeandli.com