Will AI make copywriters obsolete?

Can AI replace jobs that require "creativity," the very quality where humans should demonstrate their greatest value?

Dentsu Inc.'s " AICO2 " is the world's only AI copywriter equipped with a "Creative Thinking Model" (*). This series explores the essence of creativity through the copy generated by AICO2, with humans debating from various perspectives.

For this commemorative first installment, our guest is Naoki Tanaka (Dentsu Lab Tokyo), an Executive Creative Director/Copywriter renowned for his expertise in the intersection of technology and creativity.

Joined by AICO2 Project Leader Hiroki Kita and creator Hirose Dai, who participated in its development, we discuss AICO2's current capabilities and the future of copywriting!

*Creative Thinking Model = While the original AICO learned copy written by copywriters, AICO2 additionally learns the creator's "intent" and "thought process," and Dentsu Inc. refers to this as the Creative Thinking Model.
 

Development began with a process of mutual exploration between creators and AI engineers

Hirose: For this session on the theme of Creative × Technology, we've invited Mr. Naoki Tanaka. But first, let's hear from Mr. Kita of Marketing Division 2, the project leader for AICO2, about how AICO2 came to be.

Tanaka: Ah, but first. We should note upfront that this conversation took place in early December 2024. This field is evolving at an incredible pace.

Kita: Absolutely. Things become outdated in the blink of an eye.

Hirose: Given how rapidly this field advances, let's hear from Kitano-san about the journey of developing AICO2. The story goes that he developed AICO2 because he found it tedious to constantly request copy development from a demanding copywriter during his daily work.

Kita: Well, that's beside the point... The original AICO debuted in 2017. Back then, I was genuinely shocked by how AICO could weave together words in ways humans couldn't even imagine. According to the original development team, it was based on relentlessly compiling massive amounts of copy into a database and learning patterns from that. Meanwhile, six years after AICO's birth in 2023, generative AI based on large language models (LLMs) began appearing one after another. The trigger for developing AICO2 was wondering what would happen if we updated AICO using an LLM as its foundation. In that sense, AICO2 has a fundamentally different underlying foundation compared to the original.

Related Series:The Serendipity AI Unleashes
 

Tanaka: So the internal model is different from the original AICO?

Kita: Yes. I was originally a dedicated user of the first AICO and had high hopes for AI copywriters. However, due to the independence and entrepreneurship of Hiroyuki Fukuda (currently CEO of Chanfuku Inc.), who was the leader of the AICO project, the project entered a temporary restructuring phase.

Meanwhile, independently of AICO, I was working with Dentsu Digital Inc.'s AI engineering team to develop a target user analysis tool. The engineers on that team then suggested, "We could reboot AICO using ChatGPT now." That's how I ended up leading the development of AICO2.

The original AICO trained on "various catchphrase patterns," but this time we focused on "the thought processes of Dentsu Inc. copywriters." The Dentsu Digital Inc. AI engineering team performed Fine-Tuning(※1) based on ChatGPT 3.5.

※1 Fine-Tuning = A method of adjusting an existing pre-trained model to specialize it for a specific task.
 

Hirose: But initially, there wasn't a single creative member on the development team, right?

Kita: Right, it was just me and the AI engineering team. While AICO2 took shape, from a creative quality standpoint, it wasn't something Dentsu Inc. could release externally. That's when an executive in charge of creative told us, "We can't release this as is, but we'll add some creators to the team. Rebuild it." That's how six creators, including Hirose-san, joined the development.

Tanaka: Top-tier copywriters. That's great.

Hirose: But even though we were called in to join, we were thinking, "No way we're gonna lose to some AI," while also thinking it'd be pretty cool if we could make something amazing.

Kita: The atmosphere was pretty tense at first, right?

Hirose: As I mentioned earlier, I heard that Kitano-san actually struggled with collaborating with creators, which is why he was trying to create an AI copywriter (laughs). Hearing that made me think the reason for developing AI felt very human-like and appealing.

Kita: At the first meeting, the AI engineer members kept asking the six creators, "What thought process do you use to create copy?"

Hirose: What struck me during that first meeting was how even Kitano-san and the AI engineers seemed to understand copywriting in theory, but fundamentally didn't grasp it. It's just like how we don't understand the thought processes behind how AI engineers build AI. I've been training new creative hires in copywriting for nearly ten years. I proposed to the team that the engineers attend the first and second sessions of that training. I thought the fastest way for them to grasp it would be to understand, "Oh, so this is the thought process behind copywriting," and get a more concrete image of it.

Tanaka: That's fascinating!

Kita: Hirose-san's copywriting training was a turning point in developing AICO2. It was truly eye-opening for me too. Up until then, the copywriters I'd worked with seemed to write copy based purely on intuition. But Hirose-san's approach was incredibly theoretical and systematic. I thought, "What is this? I can understand this too!"

Tanaka: I see. I especially feel this way, but copywriting is actually almost entirely logical. Though it does get that "I do it with intuition and genius!" image sometimes.

Kita: So, after completing two sessions of training, the engineering team wanted to create a dataset that would allow the AI to understand that "logic." To do that, they gave six creators some "homework." This homework is a trade secret, though.

Tanaka: I see. The engineers who grasped the structure in just two sessions and then assigned homework are impressive. What kind of homework was it?

Hirose: Well, it was... (whispering).

Tanaka: This is... a pain! (laughs) So you have to write it in that format.

Hirose: It's a pain, right? And you have to write over 100 of them.

Tanaka: That sounds like a huge amount of work—enough to cost a fortune.

Hirose: But the copy we submitted as homework for the AI to learn wasn't all copy I'd written myself. Dentsu Inc. has a large archive of "copy that, with the author's permission, can be used legally." We divided that among the six of us and submitted the "homework" to the engineering team.

Tanaka: Ah, I see. Even if it's not copy you wrote yourself, analyzing it in this format could make it easier to understand. This is tedious but interesting. It's basically re-experiencing it, right? I think it's a good way to break things down for someone learning copywriting. I give lectures at places like the Advertising Conference, and it's similar to what I have my students do there.

Kita: By creating datasets as homework, what we taught the AI was this: AICO2 isn't designed to learn copy as the goal itself, but to learn the "reasoning" behind it. For example, with Mr. Hirose, the idea is to mimic the thought processes inside his head.

Once development progressed to a certain point, the engineers tested it by generating about 300 pieces of copy. Six creators were immediately excited from the very first output, saying things like, "This is different from before!" and "This might be good!" As development continued, the output kept improving.

Copywriting is an "unnatural language model." Is it actually incompatible with natural language processing?

Tanaka: So you prepared your own dataset to train the copywriting process.

Hirose: But it seems it's not as simple as just increasing the amount of data to train on to improve quality.

Kita: Whether it's ChatGPT or Google Gemini, these so-called LLM (Large Language Models) are fundamentally based on natural language processing. They're models designed to generate clear, natural Japanese phrasing. But here's the problem: the catchphrases copywriters create don't actually use natural language, do they?

Tanaka: Exactly. It's more like "unnatural language." That was actually the point I was most concerned about before coming today. You use ChatGPT, right? I wondered if it might not be a good fit for copywriting. I mean, wouldn't it just spit out "ordinary words"? The interesting part of copy is when something essential is missing, or it feels rough around the edges, or unexpected words appear that the reader didn't anticipate—that's what really stirs people's hearts and minds. Trying to achieve that with AI feels like it would require some pretty specialized training.

Kita: You're right. That feeling where copy makes your heart skip a beat, or that part that catches you—it's not "natural"...

Kita: Actually, I originally developed this thinking, "What if we could generate copy without assigning a human copywriter?"

Hirose: Ah, you're being honest! (laughs)

Kita: So initially, I envisioned it being used by non-creatives.

Tanaka: Yeah, I thought that's where the current demand for AI copywriters lies. For example, salespeople, or general small and medium-sized businesses and shops that don't have in-house copywriters.

Kita: But once it was finished, I realized that choosing which copy written by AICO is good still requires professional skill. Maybe you could get used to it, but it takes time. When considering how AI copywriters are used, there are two approaches. One is where the AI copywriter acts as an assistant under a copywriter. The other is where the AI writes the copy like a creative director, and humans focus on "selecting" it.

Tanaka: I think it will be possible in the future, and I hope it becomes possible. The current AICO2 might need a bit more training (laughs). But even the current AICO2 is ready for immediate use. For example, there's definitely a huge demand for situations like wanting copy for this poster but lacking the budget or route to hire a copywriter. For that kind of use, it's already at a practical level.

Hirose: If you were to use AICO2 for work right now, how would you use it? Please give me your honest opinion.

Tanaka: Hmm, actually, when I tried it, I did discover new perspectives and angles. You can find things like, "Oh, I hadn't considered that viewpoint," or "Maybe this perspective is possible." But I want to increase the occurrence rate of these discoveries, like in a gacha game. If this "discovery" rate goes up, we might enter an era where AICO2 writes copy for us as an assistant.

Hirose: By the way, how do we actually increase this gacha drop rate? I've been wondering how we intentionally trigger these so-called creative leaps. I feel like that connects to this.

Tanaka: That's tough. The most important thing when writing copy is that "Ah, I hadn't thought of that perspective" moment of discovery, right? It's the "What to Say" part. Presenting a discovery to a product or its users – like "People thought this before, but actually, there's this other way to look at it" – is the first step to good copy. And deliberately aiming for that is really hard.

Kita: My AI engineer team and I originally researched analyzing marketing "target demographics." Maybe adding the "Who" perspective—who the copy targets? Who do we want to reach?—could actually boost that gacha's hit rate.

Hirose: I see, that's interesting. But I have the opposite opinion. While target audience theory exists, looking back at my own work, the times I wrote copy I thought was good, I wasn't really writing "what to say to the target audience." Actually, it happened more when I found "something I really wanted to say myself." Something that moved me the most, or made me think, "This is what I actually wanted to tell the world."

Tanaka: I get it. When I found the copy "Life needs restaurants," it wasn't about the target audience. It felt more like "This is what I really wanted to say" or "I wanted to express my gratitude to restaurants myself."

Hirose: Right? That's why I think unless the AI has that vague, passionate feeling inside—that "this is what I really wanted to say"—it's hard to produce copy at the level Tanaka-san demands.

Tanaka: What makes this discussion even trickier is that even highly experienced copywriters clash head-on over what's good or bad. It happens often in ad award judging too. Since every writer prioritizes different values and perspectives, development becomes even harder. In that sense, it's tough for AI to be everyone's partner; it would be better if it customized itself for each copywriter.

Hirose: Ah, I see. The more you maintain it, the more it truly becomes "your" partner. Your thought processes accumulate, and the AI develops its own personality.

Tanaka: Exactly. For example, Hirose would have "DAICO" as his partner, and I'd have "NAOCO" (laughs). If it evolves like that, it could become a service copywriters really value.

Kita: Hearing that just flipped a switch for me (laughs). I want to update it into something that professionals like Tanaka-san and Hirose-san would genuinely think, "This is actually useful." As Tanaka-san mentioned earlier, non-creators using it is exactly our intended use case, so hearing there's demand is a welcome compliment. But I really want to make the professionals go, "Whoa, this newcomer is amazing!"

Copywriters won't go extinct, but I still expect that level of evolution from AICO!

Tanaka: I want to get back to the topic of occurrence rates. There's something I've noticed while using AICO2. I'm someone who likes to categorize things logically, so when I classify copy, there's one zone that's like, "Everyone already knows that," or "It's so obvious it doesn't even feel like a proposal." Then there's the zone that's too far ahead, a bit hard to grasp – "That's just your own idea," "It's self-indulgent." And good copy sits right in the middle, right? Meaning, when you think "Now that you mention it, that's true," or "I hadn't noticed, but yeah, that makes sense," that's when the empathy switch clicks and your heart moves.

Hirose: We want to increase the occurrence rate of that zone even more, right?

Tanaka: Exactly. AI tends to use data from "a little while ago," so it often ends up in the "I've seen this before" or "Someone said that already" zone. What's even trickier is that something someone discovered and was praised as "new" quickly falls into the "I've seen this before" zone once it becomes established. If AICO2 could control that and start throwing out ideas like "This is a perspective surprisingly few people have considered," then the theory that copywriters are in serious trouble would really take hold, right?

Hirose: So the theme this time is "Will AI make copywriters extinct?" (laughs). When it comes to creating that "new thing," the biggest advantage humans have is having a physical body. No matter how far AI goes, it's still fed "past information" and builds from there. Humans, on the other hand, live in the moment, feeling and experiencing things through their bodies, so they have that much more input.

Tanaka: Yeah, that's true. Current AI converts input into something it can handle as data, but humans take in all kinds of information directly—like the smell of the place, the temperature that day, or even "awkwardness"—and write copy based on that.

Hirose: If that's the case, human copywriters won't go extinct after all. But we'd still want to evolve them to the point of extinction, wouldn't we? Might as well (laughs).

Tanaka: In terms of feasibility, I think it'll be possible someday. The key is probably the "unnatural language model."

Hirose: On the other hand, AI has its own strengths, right?

Kita: Having multiple thought processes and being able to output from each is unique to AI. Humans inevitably fall back on their own methods, so they can only handle two or three approaches. Other human weaknesses are fatigue and the time it takes. If the ideal is to generate countless outputs in a short time, leading to new discoveries, then increasing the rate of those discoveries is the current challenge, right?

Tanaka: What I find truly fascinating about creating AICO2 is that copywriters' thought processes are still largely a black box. Even in this very discussion today, some things became clearer, right? Somehow, developing AICO2 feels like an endeavor to analyze "copy" – or more precisely, the method of "moving people's hearts with words" – to the absolute limit. On a similar note, Professor Hiroshi Ishiguro of Osaka University says he builds androids "to understand humans."

Hirose: Professor Ishiguro of the " Matsukoroid," right?

Tanaka: Exactly. It's the same thing. The very act of trying to recreate a copywriter's thought process with AICO makes you appreciate copywriting more. And thinking about AICO makes copywriting itself more interesting. That's why I find this project truly fascinating.

Copywriters always have to chase new things. Even copy that's "unnatural and interesting" becomes "natural" after a few days. From the AI's perspective, it's like the correct answer keeps slipping away.

Kita: It's interesting how there's no single correct answer, even though it's logical. Having a Hirose-style AICO or a Tanaka-style AICO is totally valid. Do you have any requests for AICO2, like "I want this part changed"?

Tanaka: I'd love voice input support. When I'm brainstorming or writing copy, I don't want to touch anything but pencil and paper. Having to open my PC and type into AICO2 is stressful and breaks my flow. Heck, if it could learn from our copywriter meetings via voice recognition, we wouldn't have to manually input all our homework or notes.

Hirose: That would be fantastic! Learning and growing on its own by listening to human conversations.

Tanaka: Human copywriters still have so many aspects that aren't consciously recognized or data-driven.

Hirose: Doing this homework made me realize something: one of Dentsu Inc.'s underappreciated assets is the sheer volume and quality of creative ideas our people have conceived over the years. It's undeniable. If you want to build an AI copywriter, this might be the biggest barrier to entry.

Tanaka: Since no one else seems to be tackling this area, we should really push forward with it.

Kita: No other company in the advertising field currently handles projects at this scale with this many people. Only Dentsu Inc. can train AI on the thought processes of these professionals. AICO is still evolving, but today's discussion gave us many valuable hints. We'll keep refining it. Thank you all for today!

For more on AICO2's evolution and content, please see here
https://www.dentsu.co.jp/showcase/aico2.html