A Tool for an Imperfect World: How Human Communication Is Limiting AI
- Published: 23/05/2025
- 10 minutes read
A Tool for an Imperfect World: How Human Communication Is Limiting AI
- Published: 23/05/2025
- 10 minutes read
As someone with a Master’s degree in English, I’ve spent years studying how people communicate. What fascinates me now is how that same complexity is the very thing that limits AI’s capabilities. The biggest bottleneck isn’t the machine; it’s us.
The problem of communication
When we talk, we rarely say exactly what we mean in exact terms. No language is made that way. All languages, in order to be defined as such, need to have six main elements. If a form of communication lacks any of the six elements, we can classify it only as a form of communication, not a language.
These six elements are:
Displacement – the ability to describe past, present, or future events
Productivity (or creativity) – forming new terms, sentences, and vocabulary
Arbitrariness – the lack of relation between an object and the term
Cultural transmission – the ability to teach, learn, and pass information through generations
Discreteness – having discrete units (sounds and symbols) which carry meaning and can be further combined to form a new meaning
Duality – also called double articulation – is a phenomenon where a language has meaningless sounds (phonemes) which combine into meaningful elements (morphemes)
On top of these elements, we have a complex set of rules that govern a language – grammar. It is broken into several categories: phonetics, morphology, and syntax, which govern how we structure language, and semantics and pragmatics, which are all about meaning and context. Many rules can be bent or broken arbitrarily, which gives us language variety in the form of dialects, sociolects, idiolects, jargons, slang, pidgin, and creole. There are also registers – the levels of expressions of formality – for which every language has rules. If exposed to any of the above varieties, the interesting fact is that people can fairly quickly pick it up and adapt to communicate via said variety.
The current models we use are Large Language Models (or LLMs) – it’s essentially a massive corpus of words, phrases, and sentences taken from all around the place. Books, the internet, or other media are used to ‘train’ the models, or to make them generate sentences based on their training data (hence these programs are called Generative AI).
While AI has made strides in parsing syntax and even generating fairly sophisticated text, semantics and pragmatics remain major hurdles. Why? Because these rely heavily on extralinguistic factors: tone, rhythm, cultural references, shared context, facial expressions, body language, or even silence. These inherent and often hidden carriers of meaning often get lost when AI produces a text.
Think about a simple phrase like “Nice job.” Depending on intonation and context, it can mean genuine praise, mild sarcasm, or biting criticism. Humans decode that easily. AI? Not so much.
This imperfect nature of our communication becomes a real issue when we try to use AI in contexts that demand nuance, like marketing.
The strain between imperfect language and AI in marketing
Marketing thrives on subtlety. It plays with emotion, identity, belonging, and desire, many of which are conveyed implicitly. So, when we use AI, a tool that thrives on structured input and predictable output, tension arises in this inherently chaotic space.
But that doesn’t mean AI is useless in marketing. In fact, there are several areas where it can shine, if we understand its strengths and limitations.
Where AI excels in marketing (and how to use it properly)
Data analysis at scale
One of AI’s superpowers is its ability to quickly process large volumes of text. For instance, if you’ve collected 50 customer interviews, manually analyzing them for recurring themes would take hours or even days.
AI can transcribe and summarize this data to highlight patterns, common objections, desires, and motivations, all of which are gold for building buyer personas. However, always review the AI’s output yourself. AI is still prone to “hallucinations” (i.e., confidently stating falsehoods), so use it as a tool to help you double-check your assumptions.
Testing creative variants
AI tools like ChatGPT, Copy.ai, or Jasper can generate dozens of ad headlines, scripts, or product descriptions in minutes. This is particularly useful during brainstorming or A/B testing phases. However, do not rely on AI as a final creative voice. Let copywriters adjust it with your brand in mind, and allow real-world customer reactions to shape what works.
Start with AI output
Test on a small audience segment
Iterate based on feedback.
Start with AI output
Test on a small audience segment
Iterate based on feedback.
Data Point: According to Salesforce’s State of Marketing report¹, high-performing marketers are 2.2x more likely to use AI for audience segmentation and dynamic testing. However, they also emphasize human review at every step of the creative process.
Mimicry of Style and Voice
One lesser-known use of AI is its ability to replicate tone and voice. If you’re a senior copywriter studying legends like David Ogilvy or Joseph Sugarman, you can ask AI to mimic their writing styles to explore how they might approach a campaign. This can allow copywriters to view the brand or the product from a different angle without spending too much time developing copy variants.
However, this tactic should never be used in final, customer-facing materials, especially if the AI mimics living public figures or proprietary styles. Use it as a learning tool only.
Where AI falls short (For Now)
Context retention and long-term learning are the current pitfalls of AI. When it comes to language, AIs cannot retain the register or jargon you might need to produce your marketing collateral, thereby risking diluting your brand and making it sound bland.
Or, if you’re interested, try playing “20 Questions” with an AI model, and you’ll see it often forgets earlier context or shifts interpretation halfway through. This is because most consumer-facing AIs don’t have persistent memory (unless specifically fine-tuned to do so), and they struggle with tracking long, dynamic conversations.
Compare this with humans, who build lifelong associative memories. We can recall things from childhood, hold grudges, or feel warm toward brands we interacted with years ago. AI can’t simulate that emotional or contextual memory yet because it struggles with pragmatics and semantics – something people innately utilize on a daily basis.
However, companies like OpenAI and Anthropic are testing long-context memory in enterprise settings. Still, even then, ethical risks around privacy, memory persistence, and hallucination remain significant barriers.
Another thing we as humans rely heavily on is understanding nuance and intent. Humans rely heavily on implied meaning. AI struggles here. For example, GPT models may not distinguish between a rhetorical question and a genuine inquiry, or might misinterpret a sarcastic tone as literal.
This can be disastrous in high-stakes marketing (e.g., political campaigns, healthcare products). Misunderstanding tone or intent could mean alienating customers, violating ethical norms, or even inviting legal risk. This is a modern example of what Grice called “conversational implicature”²; the idea that what’s meant in conversation often exceeds what’s said. AI still hasn’t caught up to the world between the words.
How to use AI for marketing
Given all this, marketers need to develop a structured, defensible way to use AI tools effectively without over-relying on them. Considering the fact that a lot of marketing strategies are now focused on the short term, AI can help with productivity. However, human behavior, and ultimately human purchasing decisions, are more often driven by long-term thinking.
A simple 4-step model might look like this:
Use AI to brainstorm ideas, summarize research, or draft early versions of content.
Review for hallucinations, off-brand tone, or cultural mismatches.
Test different AI-generated outputs on controlled audiences (ads, email CTAs, landing page variants).
Use feedback and performance data to refine final content before scaling.
What’s next for AI and communication?
Multimodal AI (systems combining text, image, and audio inputs) is already emerging. These models, such as Google Gemini’s latest Veo3 or OpenAI’s GPT-4 Turbo with vision, offer hope that AI will soon better understand situational context. Combined with future memory systems, this could help reduce some of the most obvious limitations in tone and intent recognition.
However, the real breakthrough may not come from better training alone, but from AI designed to ask better questions, not just provide answers. When an AI can say, “Can you clarify what you mean by that?” we’ll know it’s learning to use language the way we do.
Communication as a mirror
As someone deeply engaged in the complexities of human language, I find it both fascinating and humbling to observe how machines grapple with challenges similar to those I faced in semantics courses, struggling to understand the meaning behind context and concepts. (One famous thought experiment in semantics starts with a question: “Is a three-legged dog still a dog?”, and how many denominators you need to take away from said dog for it to stop being classified as such. The experiment aims to introduce students to the relationship between classifications and descriptors.)
AI is not fundamentally flawed; it is simply learning from us, and we are complex, emotional, and culturally nuanced beings who rarely articulate our true intentions clearly. This is not a deficiency in the tool; rather, it reflects its origins. By forming better questions, we might be able to extract better information from AI and utilize this tool to create better.
Key Learning Objectives
1. AI is powerful for data collation, ideation, and mimicry, but unsuitable for final creative decisions without human oversight.
2. Human language is imprecise by nature, and AI struggles with ambiguity, context, and subtlety.
3. A hybrid approach – human insight + machine speed – is the most effective path forward.
4. Marketers should treat AI like a tool, not a creative director.