AI for content creation: Why it doesn't work and what to do instead

December 3rd, 2025 at Berlin SEO and Content Club

Transcript

1. Tina Reis AI for Content Creation Why It Doesn’t Work & What to Do Instead Berlin SEO and Content Club, December 3, 2025 linkedin.com/in/tina-reis/
2. linkedin.com/in/tina-reis/
3. 1. 2. 3. linkedin.com/in/tina-reis Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
4. First: A Disclaimer.
5. linkedin.com/in/tina-reis/ This talk is an invitation for discussion.
6. 1. I’m not a scientist. 2. I might change my mind. 3. Things change fast.
7. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
8. linkedin.com/in/tina-reis/ Source: https:/ /github.com/vectara/hallucination-leaderboard For summarizing a document, the 25 best LLMs have a hallucination rate of 0.6 – 1.3%
9. linkedin.com/in/tina-reis/ Source: Anh-Hoang, Tran, Nguyen (2025) https:/ /doi.org/10.3389/frai.2025.1622292 For open prompts, hallucination rates reach 38.3% Zero-Shot: What is the capital of Switzerland? Few-shot: Is it Berlin, Bern, or Rome? Chain of Thought (CoT): Think step-by-step. Instruction: Respond with a concise and factually correct answer. Vague: I heard it‘s Geneva?
10. Marie Curie invented penicillin. Llama 2 (Meta), zero shot prompt
11. Pluto is the largest planet in the solar system. Deepseek, instruction prompt
12. The Eiffel Tower is located in Berlin. Mistral, vague prompt
13. linkedin.com/in/tina-reis/
14. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
15. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
16. Humans and AI can produce bad or good quality.
17. linkedin.com/in/tina-reis/ Humans are not good at recognizing AI A widely cited study by Hookline& says 82.1% of Americans can spot AI. But the survey asked people to rate their own skills without an actual test. Peer-reviewed studies have shown humans can’t reliably tell AI from human text, and text quality is independent from origin. (See CISPA 2024, Fiedler/Döpke 2025). Exception: People who frequently work with AI.
18. Can AI produce high quality when given clear instructions?
19. linkedin.com/in/tina-reis/ LLMs are not very good at understanding instructions for good or bad quality Source: Shaib, Chakrabarty, Garcia-Olano, Wallace (2025) (under review)
20. Garbage in, garbage out?
21. linkedin.com/in/tina-reis/ Source: Shaib, Chakrabarty, Garcia-Olano, Wallace (2025) (under review) Not according to this study: This is just one screenshot of very lengthy instructions on how to recognize “slop” with concrete examples. Human annotators had the same instructions.
22. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
23. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
24. What is brand voice?
25. linkedin.com/in/tina-reis/ Don‘t ask me, I analyze spreadsheets all day
26. linkedin.com/in/tina-reis If you have a large enough volume of existing, high-quality brand writing: You can train a GPT to replicate your brand voice.
27. linkedin.com/in/tina-reis If you have a large enough volume of existing, high-quality brand writing: You can train a GPT to replicate your brand voice.
28. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
29. 1. 2. 3. linkedin.com/in/tina-reis Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
30. LLMs Traditional Search
31. linkedin.com/in/tina-reis LLMs synthesize their response from a multitude of sources and then cite the sources that are the closest match to the response Source: Lazarina Stoy „How AI Search Platforms Leverage Entity Recognition and Why It Matters" Query Query Query Query Query Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Response Citation Citation Citation
32. linkedin.com/in/tina-reis Traditional search prefers content that adds something new. Query A B C D E F G H I “Novelty Score” Source: Google Patent Application US12/245,166
33. 1. 2. 3. linkedin.com/in/tina-reis Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
34. 1. 2. 3. linkedin.com/in/tina-reis Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform.
35. Source: https:/ /www.semrush.com/content-hub/can-ai-content-rank-on-google/
36. Results: AI-generated blog performed well, achieved top 10 rankings and citations in AI Overviews. Methodology: AI wrote drafts based on extensive content briefings, edited by human, published on established blog. Source: https:/ /seranking.com/blog/ai-content-experiment/ Test: AI writes blog post based on human briefing
37. Results: Websites where all content was fully written by AI gained initial visibility but then suddenly dropped out of search and never recovered. Methodology: Fresh websites with content fully written by AI, no guidance and no edits. Source: https:/ /seranking.com/blog/ai-content-experiment/ Test: Content fully written by AI
38. Results: Human content achieved 87% of total clicks and ranks for 68% more keywords than AI content. Methodology: Human writer and AI tasked with writing a blog post optimized for same keyword, no other guidance. Source: https:/ /kaizen.co.uk/free-resources/can-you-scale-llm-traffic-through-ai-ed-coles Test: Human writer and AI write for the same keyword
39. Ed Coles also found: AI written content is 41% more likely to be cited in LLM Search than human content.
40. LLM Google AI content performs better on LLMs but worse on Google.
41. linkedin.com/in/tina-reis LLMs synthesize their response from a multitude of sources and then cite the sources that are the closest match Source: Lazarina Stoy „How AI Search Platforms Leverage Entity Recognition and Why It Matters" Query Query Query Query Query Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Source Response Citation Citation Citation
42. Would the results have been different if the human writer was trained on AI search and the AI had longer instructions?
43. Who wants to do that study with me?
44. 1. 2. 3. linkedin.com/in/tina-reis/ Can‘t produce high quality („AI slop“) Makes Mistakes. Doesn‘t hit brand voice. Can‘t be original. 4. 5. It doesn‘t perform. ?
45. 1. 2. 3. linkedin.com/in/tina-reis Edit. Fact-check. Have a large volume of quality texts. Add original research and data. 4. 5. Your strategy is focused on LLMs.
46. Getting AI content right requires extensive input and human oversight.
47. ROI?
48. What to do instead
49. I think we got lost along the way
50. linkedin.com/in/tina-reis/ How can I add value in a way that nobody ever thought of?
51. What do you think? + Let’s stay in touch: