Branding in the era of generative AI: risks and opportunities
What happens when all brands use the same neural networks
The problem is not hypothetical. Midjourney, DALL·E and their analogs are trained on similar data sets, so the 'default' output is similar across all: identical perspectives, identical color palettes, identical compositions. When dozens of competitors in one category use one tool without customization, they get visually similar results — and the brand loses what it pays for most: recognizability.
There is also a more subtle effect that is rarely discussed. Researchers describe it as a feedback loop: results of neural network work are fed into new training datasets, and the algorithm begins to consider 'good design' as exactly what it generated more often before. This results in copy of copy — the visual landscape narrows by reproducing itself. This is a systemic risk for the entire category, not an error of a single brand.
Knauf and Alfa-Bank confirmed this in practice: standard models yielded 80–90% defects relative to the brand style, and companies had to train their own models on 1500+ images with descriptions in order to raise the quality above 70%. These are serious investments — but precisely these are the difference between a brand with a face and a brand that has dissolved in the noise.
Where AI really strengthens the brand and where it destroys it
The main mistake is to treat generative AI as a universal tool. It has a zone where it saves weeks of work, and a zone where it is dangerous to let it in.
| Task | You can trust AI | Why |
|---|---|---|
| Draft texts, brainstorming, headline variations | Yes | Speed is more important than accuracy, errors are not visible to the consumer |
| Mass personalization of advertising creatives | Yes, with control | Only works with user transparency and control over data (Nielsen Norman Group) |
| Routine adaptations of layouts for different formats and platforms | Yes | Mechanical work, the brand book sets the boundaries |
| Final visual representation of the brand face, key campaign | No without human review | Requires more than 5000 requests for 12 final images (case Verkhovye x Okkam) — without expert selection, an average result will emerge |
| Positioning strategy, brand DNA | No | AI cannot hold a position — it averages, rather than defends |
| Factual content (news, figures, expert statements) | Only with verification | Approximately 20% of neural network responses contain inaccurate or outdated data (data from the European Broadcasting Union and BBC) |
How to build a workflow of 'AI + human' without losing identity
Here, few companies provide a specific algorithm — usually it all boils down to 'use with caution'. The working sequence confirmed by the practices of the Evelan agency consists of five steps.
- Check the brand foundation — if positioning and visual code are blurred, AI will blur them even more.
- Establish a content management system — who sets the prompts, who approves the results, where the version history is stored.
- Choose a pilot area — do not launch AI across all channels at once, but test on a low-risk segment (for example, social media, not outdoor advertising).
- Implement cycles of human review — every AI output goes through an editor or brand strategist before publication.
- To capture visible results — compare metrics before and after, so that the decision to scale is based on data, not on the feeling that 'this is trendy'.
Separately, the flexibility of the brand book is worth noting. This is perceived as the main challenge of the era: to transition from rigid rules to flexible systems where AI adapts the message to context, but cannot go beyond the brand DNA. This is not about abandoning the brand book — it is about describing not only 'what is not allowed', but also the range of acceptable variations that are understandable to an algorithm.
Why AI-generated content causes distrust — and what to do about it
The numbers here are uncomfortable. According to the Edelman Trust Barometer 2024, global trust in AI companies has dropped from 61% to 53% over five years, and in the US — from 50% to 35%. This is not abstract anxiety, but a direct impact on conversion.
Studies show that adding the label 'AI-generated' reduces engagement and conversion regardless of the actual quality of the content. Moreover, the effect is stronger for premium products — the higher the price, the more sharply the audience reacts to suspicions of 'cheap' production. This is a key finding that changes strategy: mechanically revealing the use of AI without explaining the context works against the brand.
The solution is not to hide the use of AI (which is legally and reputationally risky), but to communicate it meaningfully: explaining what exactly AI does (for example, personalizes the selection), while maintaining the user's sense of control. Nielsen Norman Group notes this pattern: personalization is perceived positively only when transparent and when the user understands they can influence the data processed by the algorithm. An 'AI' label without explanation is not transparency — it's a warning flag.
Legal risks: what companies actually face
Three zones of risk are rarely examined together, but they should be.
- Copyright. In Russian law, only a person can be an author — Article 1228 of the Civil Code of the Russian Federation explicitly states this. Therefore, an image generated by a neural network without substantial creative human involvement may legally not receive protection as a copyright object. Practical conclusion: document prompts, versions, and refinements — without this, it is almost impossible to prove authorship in court.
- Trademarks. Generative models may accidentally reproduce elements of another party's registered trademark or distinctive trade style — this is a direct risk of a claim from a competitor.
- Deepfakes and licensing of individuals. Using a person's image (even a stylized one) without consent is a separate legal risk that is amplified by the increasing realism of video generation.
Risk minimization is based on three practices: documenting the entire generation process, using models with transparent licensing of training data (or training your own, as Knauf and Alfa-Bank did), and always running final materials through legal review before publishing in a major campaign.
How to tell if AI branding is working, rather than just cutting costs
Saving on production is not a success indicator if the audience stops trusting the brand. You need to look at the set of metrics, not just one number.
| Metric | What it shows | Benchmark from practice |
|---|---|---|
| CTR / VTR | Does the content attract attention and hold it | Case Verkhov'e x Okkam: CTR 1.4%, VTR 80% |
| Index of trust in the material | How much the audience believes the content is authentic | Verkhov'e x Okkam: 78/100 |
| Ratio of defects to finished content | Real efficiency of investments in training the model | Standard models — 80-90% defects before retraining |
| Engagement dynamics when the AI fact is revealed | Risk of losing audience trust | Decline regardless of quality (data from Freak) |
If CTR is growing, but the trust index is falling — this is not success, it is a delayed problem that will surface in loyalty and repeat purchases later.
The new role of the person: not an executor, but a filter
The role of designers and copywriters shifts from production to quality control and strategic editing. Practically, this looks like three new functions.
- Brand DNA Strategy — sets the boundaries within which AI can vary the message, and ensures that flexibility does not turn into identity blurring.
- Editor-Filter — checks each AI output for factual accuracy (remembering 20% hallucinations) and brand tone consistency.
- Legal Cleanliness Controller — maintains prompt and version documentation, mitigates copyright and trademark risks prior to publication.
Companies that benefit from AI are not those that replaced people with algorithms, but those that restructured the process: humans set the boundaries and make the final decision, AI accelerates the draft stage. It's a synergy, not a replacement.
What happens when all brands use the same neural networks
The problem is not hypothetical. Midjourney, DALL·E and their analogs are trained on similar data sets, so the 'default' output is similar across all: identical perspectives, identical color palettes, identical compositions. When dozens of competitors in one category use one tool without customization, they get visually similar results — and the brand loses what it pays for most: recognizability.
There is also a more subtle effect that is rarely discussed. Researchers describe it as a feedback loop: results of neural network work are fed into new training datasets, and the algorithm begins to consider 'good design' as exactly what it generated more often before. This results in copy of copy — the visual landscape narrows by reproducing itself. This is a systemic risk for the entire category, not an error of a single brand.
Knauf and Alfa-Bank confirmed this in practice: standard models yielded 80–90% defects relative to the brand style, and companies had to train their own models on 1500+ images with descriptions in order to raise the quality above 70%. These are serious investments — but precisely these are the difference between a brand with a face and a brand that has dissolved in the noise.
Where AI really strengthens the brand and where it destroys it
The main mistake is to treat generative AI as a universal tool. It has a zone where it saves weeks of work, and a zone where it is dangerous to let it in.
| Task | You can trust AI | Why |
|---|---|---|
| Draft texts, brainstorming, headline variations | Yes | Speed is more important than accuracy, errors are not visible to the consumer |
| Mass personalization of advertising creatives | Yes, with control | Only works with user transparency and control over data (Nielsen Norman Group) |
| Routine adaptations of layouts for different formats and platforms | Yes | Mechanical work, the brand book sets the boundaries |
| Final visual representation of the brand face, key campaign | No without human review | Requires more than 5000 requests for 12 final images (case Verkhovye x Okkam) — without expert selection, an average result will emerge |
| Positioning strategy, brand DNA | No | AI cannot hold a position — it averages, rather than defends |
| Factual content (news, figures, expert statements) | Only with verification | Approximately 20% of neural network responses contain inaccurate or outdated data (data from the European Broadcasting Union and BBC) |
How to build a workflow of 'AI + human' without losing identity
Here, few companies provide a specific algorithm — usually it all boils down to 'use with caution'. The working sequence confirmed by the practices of the Evelan agency consists of five steps.
- Check the brand foundation — if positioning and visual code are blurred, AI will blur them even more.
- Establish a content management system — who sets the prompts, who approves the results, where the version history is stored.
- Choose a pilot area — do not launch AI across all channels at once, but test on a low-risk segment (for example, social media, not outdoor advertising).
- Implement cycles of human review — every AI output goes through an editor or brand strategist before publication.
- To capture visible results — compare metrics before and after, so that the decision to scale is based on data, not on the feeling that 'this is trendy'.
Separately, the flexibility of the brand book is worth noting. This is perceived as the main challenge of the era: to transition from rigid rules to flexible systems where AI adapts the message to context, but cannot go beyond the brand DNA. This is not about abandoning the brand book — it is about describing not only 'what is not allowed', but also the range of acceptable variations that are understandable to an algorithm.
Why AI-generated content causes distrust — and what to do about it
The numbers here are uncomfortable. According to the Edelman Trust Barometer 2024, global trust in AI companies has dropped from 61% to 53% over five years, and in the US — from 50% to 35%. This is not abstract anxiety, but a direct impact on conversion.
Studies show that adding the label 'AI-generated' reduces engagement and conversion regardless of the actual quality of the content. Moreover, the effect is stronger for premium products — the higher the price, the more sharply the audience reacts to suspicions of 'cheap' production. This is a key finding that changes strategy: mechanically revealing the use of AI without explaining the context works against the brand.
The solution is not to hide the use of AI (which is legally and reputationally risky), but to communicate it meaningfully: explaining what exactly AI does (for example, personalizes the selection), while maintaining the user's sense of control. Nielsen Norman Group notes this pattern: personalization is perceived positively only when transparent and when the user understands they can influence the data processed by the algorithm. An 'AI' label without explanation is not transparency — it's a warning flag.
Legal risks: what companies actually face
Three zones of risk are rarely examined together, but they should be.
- Copyright. In Russian law, only a person can be an author — Article 1228 of the Civil Code of the Russian Federation explicitly states this. Therefore, an image generated by a neural network without substantial creative human involvement may legally not receive protection as a copyright object. Practical conclusion: document prompts, versions, and refinements — without this, it is almost impossible to prove authorship in court.
- Trademarks. Generative models may accidentally reproduce elements of another party's registered trademark or distinctive trade style — this is a direct risk of a claim from a competitor.
- Deepfakes and licensing of individuals. Using a person's image (even a stylized one) without consent is a separate legal risk that is amplified by the increasing realism of video generation.
Risk minimization is based on three practices: documenting the entire generation process, using models with transparent licensing of training data (or training your own, as Knauf and Alfa-Bank did), and always running final materials through legal review before publishing in a major campaign.
How to tell if AI branding is working, rather than just cutting costs
Saving on production is not a success indicator if the audience stops trusting the brand. You need to look at the set of metrics, not just one number.
| Metric | What it shows | Benchmark from practice |
|---|---|---|
| CTR / VTR | Does the content attract attention and hold it | Case Verkhov'e x Okkam: CTR 1.4%, VTR 80% |
| Index of trust in the material | How much the audience believes the content is authentic | Verkhov'e x Okkam: 78/100 |
| Ratio of defects to finished content | Real efficiency of investments in training the model | Standard models — 80-90% defects before retraining |
| Engagement dynamics when the AI fact is revealed | Risk of losing audience trust | Decline regardless of quality (data from Freak) |
If CTR is growing, but the trust index is falling — this is not success, it is a delayed problem that will surface in loyalty and repeat purchases later.
The new role of the person: not an executor, but a filter
The role of designers and copywriters shifts from production to quality control and strategic editing. Practically, this looks like three new functions.
- Brand DNA Strategy — sets the boundaries within which AI can vary the message, and ensures that flexibility does not turn into identity blurring.
- Editor-Filter — checks each AI output for factual accuracy (remembering 20% hallucinations) and brand tone consistency.
- Legal Cleanliness Controller — maintains prompt and version documentation, mitigates copyright and trademark risks prior to publication.
Companies that benefit from AI are not those that replaced people with algorithms, but those that restructured the process: humans set the boundaries and make the final decision, AI accelerates the draft stage. It's a synergy, not a replacement.
Recent articles
MegaFon, BMW, VTB and more 240+ companies. See how it looks in the showcase reel:
Our branding agency is approached to order:
Branding agency services
