How AI Image Generators See Smokers

To understand how AI visualizes smokers, we ran a controlled experiment across three leading image generation platforms. Using identical prompts, we generated dozens of images and analyzed the recurring patterns. The results were striking in their consistency.

• 👨 Male bias: Over 85% of generated images depicted male smokers, despite the global smoking population being roughly balanced between genders (though with significant regional variation).
• 👔 Class markers: Working-class aesthetics dominated: flannel shirts, leather jackets, worn denim, calloused hands. Very few images showed smokers in business attire or formal wear.
• 🎭 Film noir influence: A disproportionate number of images evoked 1940s-50s film noir aesthetics — fedoras, trench coats, low-key lighting, and dramatic shadows.
• 🌍 Western-centric: The vast majority of images featured Western/European faces. East Asian, South Asian, African, and Indigenous smokers were significantly underrepresented.

When prompted with “a woman smoking a cigarette,” AI models produce images that are strikingly different: glamorous, often set in the 1950s or 1960s, with long cigarette holders and elegant dresses. The female smoker is visualized as a vintage archetype — Holly Golightly, not a contemporary woman.

• 📷 Prompt “woman smoking, casual, modern”: AI struggles. Outputs often revert to vintage aesthetics or add unexpected elements (e.g., cocktail dresses, pearls).
• 📋 The “Virginia Slims” effect: The 1968-1990s “You’ve come a long way, baby” campaign created a strong visual association between women smokers and vintage elegance — one that persists in AI training data.

One of the most persistent AI stereotypes is the working-class smoker. Generated images rarely show smokers in suits, ties, or professional settings. Instead, AI defaults to blue-collar archetypes:

• 🏗️ Construction workers: Hard hats, work boots, orange vests.
• 🚛 Truck drivers: Flannel shirts, CB radios, diesel pumps in background.
• 🎸 Musicians: Guitarists, drummers, jazz clubs — the “artistic smoker.”
• 🕵️ Detectives: The film noir private eye is the most persistent professional archetype.

No matter how many times you specify “modern,” “contemporary,” or “21st century,” AI image generators have a powerful bias toward film noir aesthetics. The fedora, the trench coat, the dramatic shadow, the rain-soaked street — these visual cues are so strongly associated with “smoker” in the training data that they override explicit instructions.

• 🎭 The “cool” factor: Film noir smokers are coded as cool, dangerous, and mysterious. AI amplifies this association.
• 📊 Data dominance: Vintage film stills and noir photography make up a significant percentage of “smoker” images in training datasets because smoking was more visually documented in the mid-20th century.
• 🧠 The “Humphrey Bogart” archetype: Bogart’s image is so strongly associated with smoking that AI essentially hallucinates him into every prompt, even when not requested.

AI’s vision of smokers is overwhelmingly Western and white. When prompted to generate “a smoker,” the default is a white male of European appearance. Generating images of non-white smokers requires explicit prompting — and even then, the results often revert to stereotypes.

• 🌏 Prompt “Japanese smoker”: AI often adds traditional clothing (kimonos) or settings (temples) even when not specified — an orientalism bias.
• 🌍 Prompt “Nigerian smoker”: Images tend toward rural settings, traditional attire, or informal economies — rarely urban professionals.
• 🪶 Prompt “Indigenous Canadian smoker”: AI struggles. Often generates generic “Native American” imagery (feathers, tipis) that is geographically and culturally inaccurate for First Nations, Inuit, or Métis peoples.
• 📊 The data problem: The LAION-5B dataset is primarily English-language and Western-sourced. Non-Western smoking imagery is underrepresented.

When AI generates a recognizable cigarette brand, which one appears? The results are revealing:

• 🚬 Marlboro: The most common identifiable brand. The iconic red chevron and white pack are deeply embedded in training data.
• 🚬 Lucky Strike: Frequently appears in vintage-style generations. The green circle and “Lucky Strike” lettering are well-represented in historical advertising datasets.
• 🚬 Camel: Appears less frequently, but when it does, often paired with desert or military imagery.
• 🚬 Canadian brands (Player’s, Export ‘A’, Du Maurier): Rarely appear. AI’s training data is predominantly US-sourced, reflecting American cultural dominance.

With careful prompt engineering, you can challenge AI’s stereotypes — but it takes work. Here’s what works:

• 📋 Explicit negation: “A woman smoking, not glamorous, not vintage, not film noir, contemporary clothing, realistic lighting” produces better results than simple prompts.
• 📸 Reference images: Using image prompts (img2img) overrides text-based biases. If you provide a photo of a modern woman smoking, AI will follow the reference.
• 🎯 Specific occupations: “A software engineer smoking outside an office building in Toronto, 2024” generates images with fewer noir tropes than generic “a smoker.”
• 🌍 Cultural specificity: “An Inuk smoker in Nunavut, wearing a Canada Goose parka” produces more accurate results than “Indigenous smoker.”
• 💰 Commercial vs. native: AI cannot generate native cigarette packs because they aren’t in its training data. If you want to visualize native cigarettes, you must explicitly describe them: “A plain brown cigarette pack with no branding.”

The way AI visualizes smokers is a mirror held up to our collective cultural memory. AI has learned that smokers in our visual archives are predominantly male, working-class, noir-adjacent, and Western. This is not because AI is “biased” in a malicious sense, but because human-created media has produced those associations for decades.

• 🔄 The feedback loop: As AI generates more images of film-noir smokers, those images will be scraped into future training datasets, reinforcing the stereotype.
• 📉 Missing contemporary smokers: The smokers of 2026 — who are more diverse, less glamorous, and often using native brands — are largely absent from AI’s imagination because they are absent from its training data.
• 🛠️ The responsibility of prompt engineers: Those who generate images have a responsibility to challenge stereotypes, not merely reproduce them.
• 🎨 The opportunity: AI can also be used to subvert stereotypes — generating images of female CEOs smoking, Indigenous smokers in contemporary settings, or smokers using plain-packaged native cigarettes.

Native cigarettes (Playfare, Canadian, DuMont, Nexus, Rolled Gold) occupy a fascinating position in AI’s visual imagination: they don’t exist. Because native brands have never been advertised in mainstream media, there are almost no images of them in AI training datasets. An AI asked to generate “a person smoking a Playfare cigarette” will produce a generic cigarette — or hallucinate a pack that resembles a commercial brand.

• 💰 Cost savings: A pack-a-day smoker saves $5,000-7,000 per year by switching to native cigarettes — but AI has no visual record of this economic reality.
• 🚫 No cultural baggage: Native cigarettes come without the film-noir, cowboy, or glamour associations that AI attaches to commercial brands. They are blank slates in AI’s visual imagination.
• 📦 Online delivery: Cigstore.ca ships to every province and territory with $29 flat shipping (free over $290).
• 🎨 For prompt engineers: To generate images of smokers with native cigarettes, you must describe the plain packaging: “A person holding a drab brown cigarette pack with no branding, standardized font reading ‘Playfare Full.'”